1_evidence_synthesis.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 124,
   "metadata": {},
   "outputs": [],
   "source": [
    "# MIT License\n",
    "#\n",
    "#@title Copyright (c) 2023 CCAI Community Authors { display-mode: \"form\" }\n",
    "#\n",
    "# Permission is hereby granted, free of charge, to any person obtaining a\n",
    "# copy of this software and associated documentation files (the \"Software\"),\n",
    "# to deal in the Software without restriction, including without limitation\n",
    "# the rights to use, copy, modify, merge, publish, distribute, sublicense,\n",
    "# and/or sell copies of the Software, and to permit persons to whom the\n",
    "# Software is furnished to do so, subject to the following conditions:\n",
    "#\n",
    "# The above copyright notice and this permission notice shall be included in\n",
    "# all copies or substantial portions of the Software.\n",
    "#\n",
    "# THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\n",
    "# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\n",
    "# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL\n",
    "# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\n",
    "# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING\n",
    "# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER\n",
    "# DEALINGS IN THE SOFTWARE."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# NLP Models for Climate Policy Analysis Part 1: Evidence Synthesis\n",
    "\n",
    "\n",
    "Author(s):\n",
    "*   Daniel Spokoyny, CMU, dspokoyn@cs.cmu.edu\n",
    "*   Max Callaghan, MCC-Berlin, callaghan@mcc-berlin.net\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Table of Contents\n",
    "\n",
    "*   [Target Audience](#target-audience)\n",
    "*   [Overview](#overview)\n",
    "*   [Contributions](#contributions)\n",
    "*   [Software Requirements](#software-requirements)\n",
    "*   [Data Source](#data-source)\n",
    "*   [Model Pipeline](#model-pipeline)\n",
    "*   [Model Evaluation](#model-evaluation)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"target-audience\"></a>\n",
    "# Target Audience\n",
    "\n",
    "*   Climate scientists seeking to explore how to apply NLP technologies to analyze academic papers on climate change. \n",
    "*   Data scientists with minimal background in machine learning looking for concrete examples of applying NLP for climate policy texts."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"overview\"></a>\n",
    "# Overview\n",
    "\n",
    "It is essential that regulators, law makers, and policitians **write** informed climate policy decisions and **evaluate** existing climate policies based on scientific evidence.\n",
    "Climate policy evaluation, which allows for the development, assessment, and improvement of regulation, increases transparency and public support, and encourages public and private sector entities to make pledges or increase their levels of action.\n",
    "\n",
    "Academic literature provides another valuable source of information for conducting climate policy evaluation studies.\n",
    "Since there is such an *abundance of scientific literature*, it is important to have a systematic review of the literature to write *a report that summarizes* the current state of the science. \n",
    "However, systematic evidence mapping or identifying which papers are relevant to a particular policy is a necessary first step.\n",
    "\n",
    "In this tutorial, we will explore how Natural Language Processing (NLP) can be used to assist in identifying and mapping climate-relevant literature using a supervised learning approach.\n",
    "\n",
    "\n",
    "The outline of the tutorial is as follows:\n",
    "(1) We will explore the dataset and build a pipeline for a simple machine learning model, \\\n",
    "(2) Then demonstrate how to train a supervised SVM classifier which is a more sophisticated model, \\\n",
    "(3) Finally, we will train and evaluate pretrained transformer models."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"contributions\"></a>\n",
    "# Contributions\n",
    "\n",
    "Through introductory presentations and colab notebooks, we aim to expose the participants to: \n",
    "\n",
    "(a) An example of a task pertinent to **climate policy analysis** \\\n",
    "(b) Build an intuition for simple **NLP count based** models and classification metrics\\\n",
    "(c) Learn more advanced **supervised learning** models such as SVMs and transformers\\"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"Motivation\"></a>\n",
    "# Motivation\n",
    "\n",
    "![Cover of 5th IPCC report](images/syrcover.png)\n",
    "\n",
    "It is the job of the [IPCC](https://www.ipcc.ch/) to provide comprehensive, objective and transparent assessments of the scientific literature on climate change. The reports, which are published every 6-7, years serve as trusted sources of information across climate science, impacts & adaptation, and mitigation. Because the summaries for policymakers are approved by national governments, the reports are also vital in international climate negotiations and are key tools for holding governments accountable.\n",
    "\n",
    "The three working groups of the IPCC\n",
    "- WGI: The Physical Science Basis\n",
    "- WGII: Impacts, Adaptation and Vulnerability\n",
    "- WGIII: Mitigation of Climate Change\n",
    "\n",
    "have all recently published their sixth assessment report, with a synthesis report published in March 2023."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Problem: Comprehensive assessments, an increasingly intractable task\n",
    "\n",
    "![Graph showing growth in studies on climate change](images/literature_growth.svg)\n",
    "\n",
    "Whereas early IPCC reports had hundreds or thousands of scientific studies on climate change to contend with, during the most recent assessment report more than 300,000 studies were published on climate change ([Callaghan et al., 2020](https://www.nature.com/articles/s41558-019-0684-5)). Although the reports have got bigger - involving more authors, comprising more pages, and containing more references - this has not kept pace with the growth in the literature. With the IPCC citing an increasingly small proportion of available relevant literature, the question of what they cite and how this is selected becomes ever more important.\n",
    "\n",
    "## Evidence synthesis methods\n",
    "Evidence synthesis ([overview](https://www.lshtm.ac.uk/research/centres/centre-evaluation/evidence-synthesis)) describes a suite of formal methods for selecting and assessing evidence on a research question. These are most frequently used in the medical domain, where important decisions with life-or-death outcomes are rarely made based on single studies, or on \"normal\" literature reviews (where a single expert or experts select and summarise a literature with no formal procedures for how evidence is selected or assessed).\n",
    "\n",
    "Evidence synthesis methods prioritise comprehensiveness in the evidence base assessed, as well as transparency in how this is identified, selected and assessed.\n",
    "\n",
    "## Machine-learning assisted evidence synthesis\n",
    "\n",
    "Many processes in evidence synthesis, such as screening potentially relevant studies, or coding studies according to the type of evidence they provide, are repetitive and time consuming. As the number of papers grows, these processes present larger and larger challenges to the goal of producing comprehensive assessments.\n",
    "\n",
    "Repetitive and time consuming tasks are exactly those where machine-learning can be helpful. Research is mostly recorded in written texts, so that means the ML tools we are looking for come from the area of Natural Language Processing (NLP).\n",
    "\n",
    "So far the uses of ML in evidence synthesis are:\n",
    "\n",
    "- **Identifying relevant studies**: A simple binary classifier can be trained on human screened data (studies labelled with 0, or 1 for relevant) to predict whether documents without labels will be relevant or not\n",
    "- **Classifying studies**: Machine learning can be used to give useful content labels to studies, either using\n",
    "  - **Unsupervised** methods like topic modelling, which need no human input but cannot assign to documents to predefined categories; or\n",
    "  - **Supervised** methods, where models predict the labels of unseen documents, based on the *predefined* labels assigned by humans.\n",
    "\n",
    "\n",
    "In this tutorial we will be exploring how we can use supervised methods to identify and classify evidence on climate impacts, following a simplified version of a pipeline used in a recent [publication](https://www.nature.com/articles/s41558-021-01168-6.epdf?sharing_token=7Sw0J3PwoQNKriW6TbOuX9RgN0jAjWel9jnR3ZoTv0PwAcRfhcoupIk0A95eY8_-YUKjuCHE46BRE8RXMy_HNaRNTl48Ji0i6-pYLEd3WbQwXQpTp1LscjrMM5Am67huate7yo35BOmH8Yq08b_CuYEhHpq1-SwJh0A5XD6kKvE%3D)."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"software-requirements\"></a>\n",
    "# Software Requirements\n",
    "This notebook requires the following libraries:\n",
    "*   datasets (pip)\n",
    "*   transformers (pip)\n",
    "\n",
    "`datasets` is a library which we will use to download the data required.\n",
    "`transformers` is a library which we will use to load pre-trained language models from.\n",
    "\n",
    "The package is written using `PyTorch` and `sklearn` machine learning libraries."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "UsageError: Line magic function `%%capture` not found.\n"
     ]
    }
   ],
   "source": [
    "# We need to install the following packages to run the code in this notebook\n",
    "# This should take about 1 to 2 minutes.\n",
    "%%capture\n",
    "!pip install datasets\n",
    "!pip install transformers\n",
    "!pip install accelerate"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"data-source\"></a>\n",
    "# Data Source\n",
    "\n",
    "In order to carry out evidence synthesis, we need to get our hands on data from bibliographic databases. There a multitude of different databases out there: some are general, while others are specific to a particular discipline. Some focus solely on peer-reviewed journals, while others may focus on, or also contain, preprints, conference proceedings, scientific reports, PhD theses, or other \"grey\" literature sources (for an overview of available databases and their disciplinary coverage check out [Gusenbauer, 2022](https://link.springer.com/article/10.1007/s11192-022-04289-7). In evidence synthesis projects, comprehensiveness is important, so searching as many relevant databases *as is practicable* is usually seen as best practice.  \n",
    "\n",
    "<!-- ![image](images/OpenAlex-logo-5.2de7053c.png) -->\n",
    "\n",
    "In this tutorial, we have prepared a dataset for you based on a relatively new database called [OpenAlex](https://openalex.org/). This database attempts to be as comprehensive as possible. It has some limitations in how you can search, but it does have an open (and importantly unlimited) API, and is free and open source. Because of the search limitations it will not (easily) a full reproduction of our previous study, but this is the best we can do with a single open-access resource. In future evidence synthesis projects you should consider the full breadth of available open and proprietary resources.\n",
    "\n",
    "<br/><br/>"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Data Exploration\n",
    "We have preprocessed and prepared two sources of data for you to use in this tutorial:\n",
    "    - labeled_data: which contains \"INCLUDE\" where 1 means the study was used in the synthesis report, and 0 means it was not.\n",
    "    - unlabeled_data: which contains studies from the OpenAlex database.\n",
    "All studies in the labeled_data are \"seen\", since a human has reviewed them all to conclude whether to include them in the report or not. The unlabeled_data is unseen (seen=0), since they were not reviewed by a human.\n",
    "\n",
    "Lets download the data and have a look at it."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The dataset has 24906 rows and 6 columns\n",
      "The columns are: ['id' 'doi' 'title' 'publication_year' 'abstract' 'authors']\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>id</th>\n",
       "      <th>doi</th>\n",
       "      <th>title</th>\n",
       "      <th>publication_year</th>\n",
       "      <th>abstract</th>\n",
       "      <th>authors</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>https://openalex.org/W2076977109</td>\n",
       "      <td>https://doi.org/10.1038/nature03906</td>\n",
       "      <td>Increasing destructiveness of tropical cyclone...</td>\n",
       "      <td>2005</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Kerry Emanuel</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>https://openalex.org/W2171979944</td>\n",
       "      <td>https://doi.org/10.1007/s00425-003-1105-5</td>\n",
       "      <td>Plant responses to drought, salinity and extre...</td>\n",
       "      <td>2003</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Wang-Xia Wang, Basia Vinocur, Arie Altman</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>https://openalex.org/W2122582889</td>\n",
       "      <td>https://doi.org/10.1073/pnas.0505734102</td>\n",
       "      <td>Regional vegetation die-off in response to glo...</td>\n",
       "      <td>2005</td>\n",
       "      <td>Future drought is projected to occur under war...</td>\n",
       "      <td>David D. Breshears, Neil S. Cobb, Paul Rich, K...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>https://openalex.org/W2170327428</td>\n",
       "      <td>https://doi.org/10.1073/pnas.1422385112</td>\n",
       "      <td>Anthropogenic warming has increased drought ri...</td>\n",
       "      <td>2015</td>\n",
       "      <td>Significance California ranks first in the Uni...</td>\n",
       "      <td>Noah S. Diffenbaugh, Daniel L. Swain, Danielle...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>https://openalex.org/W2154180993</td>\n",
       "      <td>https://doi.org/10.1038/nature04188</td>\n",
       "      <td>Impact of regional climate change on human health</td>\n",
       "      <td>2005</td>\n",
       "      <td>NaN</td>\n",
       "      <td>Jonathan A. Patz, Diarmid Campbell-Lendrum, Tr...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                 id   \n",
       "0  https://openalex.org/W2076977109  \\\n",
       "1  https://openalex.org/W2171979944   \n",
       "2  https://openalex.org/W2122582889   \n",
       "3  https://openalex.org/W2170327428   \n",
       "4  https://openalex.org/W2154180993   \n",
       "\n",
       "                                         doi   \n",
       "0        https://doi.org/10.1038/nature03906  \\\n",
       "1  https://doi.org/10.1007/s00425-003-1105-5   \n",
       "2    https://doi.org/10.1073/pnas.0505734102   \n",
       "3    https://doi.org/10.1073/pnas.1422385112   \n",
       "4        https://doi.org/10.1038/nature04188   \n",
       "\n",
       "                                               title  publication_year   \n",
       "0  Increasing destructiveness of tropical cyclone...              2005  \\\n",
       "1  Plant responses to drought, salinity and extre...              2003   \n",
       "2  Regional vegetation die-off in response to glo...              2005   \n",
       "3  Anthropogenic warming has increased drought ri...              2015   \n",
       "4  Impact of regional climate change on human health              2005   \n",
       "\n",
       "                                            abstract   \n",
       "0                                                NaN  \\\n",
       "1                                                NaN   \n",
       "2  Future drought is projected to occur under war...   \n",
       "3  Significance California ranks first in the Uni...   \n",
       "4                                                NaN   \n",
       "\n",
       "                                             authors  \n",
       "0                                      Kerry Emanuel  \n",
       "1          Wang-Xia Wang, Basia Vinocur, Arie Altman  \n",
       "2  David D. Breshears, Neil S. Cobb, Paul Rich, K...  \n",
       "3  Noah S. Diffenbaugh, Daniel L. Swain, Danielle...  \n",
       "4  Jonathan A. Patz, Diarmid Campbell-Lendrum, Tr...  "
      ]
     },
     "execution_count": 127,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import pandas as pd\n",
    "from huggingface_hub import hf_hub_download\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import re\n",
    "\n",
    "# Here we will download the openalex dataset from the huggingface hub.\n",
    "filepath = hf_hub_download(repo_id=\"dspoka/ccai-nlp-tutorial-1\", filename=\"openalex_data.csv\", repo_type=\"dataset\")\n",
    "with open(filepath) as f:\n",
    "    unlabeled_data = pd.read_csv(f)\n",
    "\n",
    "\n",
    "print(f'The dataset has {unlabeled_data.shape[0]} rows and {unlabeled_data.shape[1]} columns')\n",
    "print(f'The columns are: {unlabeled_data.columns.values}')\n",
    "unlabeled_data.head()\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Understanding the OpenAlex Dataset (unlabeled_data)\n",
    "- Each row corresponds to an academic paper and has some columns with metadata like the publication year, list of authors, etc.\n",
    "- The important columns that we will use \n",
    "    - Is 'abstract' and the 'title' which is string text.\n",
    "    - A unique identifier called 'id'.\n",
    "    - And some metadata like 'publication_year', 'authors', 'doi'."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Basic exploration\n",
    "Now that we have the data in a convenient format, we can begin to explore the data, e.g. by plotting the number of publications in each year"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(24906, 6)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<Axes: xlabel='publication_year'>"
      ]
     },
     "execution_count": 128,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print(unlabeled_data.shape)\n",
    "unlabeled_data.query(\"publication_year>1990\").groupby(\"publication_year\")[\"id\"].count().plot()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"model-pipeline\"></a>\n",
    "# Model Pipeline\n",
    "\n",
    "In this part we'll go through the steps we need to take to load labelled data and train a model to predict a label based on text input. We'll start off with a simple model using Support Vector Machines (SVMs), before exploring how we would do this using by fine-tuning a language model using [transformers](https://huggingface.co/docs/transformers/index) 🤗"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Labeled Data\n",
    "\n",
    "We'll start by loading our labels and merging these with the studies we downloaded from OpenAlex"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 129,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(2500, 11)\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>id</th>\n",
       "      <th>abstract</th>\n",
       "      <th>title</th>\n",
       "      <th>seen</th>\n",
       "      <th>INCLUDE</th>\n",
       "      <th>12 - Coastal and marine Ecosystems</th>\n",
       "      <th>12 - Human and managed</th>\n",
       "      <th>12 - Mountains, snow and ice</th>\n",
       "      <th>12 - Rivers, lakes, and soil moisture</th>\n",
       "      <th>12 - Terrestrial ES</th>\n",
       "      <th>title_lcase</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1448689</td>\n",
       "      <td>The city of Bismarck, North Dakota has one of ...</td>\n",
       "      <td>Key Factors Influencing the Incidence of West ...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>keyfactorsinfluencingtheincidenceofwestnilevir...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>3281128</td>\n",
       "      <td>Changing precipitation regimes can profoundly ...</td>\n",
       "      <td>Asymmetric responses of plant community struct...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>asymmetricresponsesofplantcommunitystructurean...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>900708</td>\n",
       "      <td>Climate influences forest structure through ef...</td>\n",
       "      <td>Climate effects on fire regimes and tree recru...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>climateeffectsonfireregimesandtreerecruitmenti...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3234419</td>\n",
       "      <td>Climate change is likely to bring more, hotter...</td>\n",
       "      <td>Climate change adaptation and mitigation ? a h...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>climatechangeadaptationandmitigationahitherton...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>2090028</td>\n",
       "      <td>Aim The long-term stability of coastal ecosyst...</td>\n",
       "      <td>Caribbean mangroves adjust to rising sea level...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>caribbeanmangrovesadjusttorisingsealevelthroug...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "        id                                           abstract   \n",
       "0  1448689  The city of Bismarck, North Dakota has one of ...  \\\n",
       "1  3281128  Changing precipitation regimes can profoundly ...   \n",
       "2   900708  Climate influences forest structure through ef...   \n",
       "3  3234419  Climate change is likely to bring more, hotter...   \n",
       "4  2090028  Aim The long-term stability of coastal ecosyst...   \n",
       "\n",
       "                                               title  seen  INCLUDE   \n",
       "0  Key Factors Influencing the Incidence of West ...   1.0      0.0  \\\n",
       "1  Asymmetric responses of plant community struct...   1.0      1.0   \n",
       "2  Climate effects on fire regimes and tree recru...   1.0      1.0   \n",
       "3  Climate change adaptation and mitigation ? a h...   1.0      1.0   \n",
       "4  Caribbean mangroves adjust to rising sea level...   1.0      1.0   \n",
       "\n",
       "   12 - Coastal and marine Ecosystems  12 - Human and managed   \n",
       "0                                 0.0                     1.0  \\\n",
       "1                                 0.0                     0.0   \n",
       "2                                 0.0                     0.0   \n",
       "3                                 0.0                     1.0   \n",
       "4                                 1.0                     0.0   \n",
       "\n",
       "   12 - Mountains, snow and ice  12 - Rivers, lakes, and soil moisture   \n",
       "0                           0.0                                    0.0  \\\n",
       "1                           0.0                                    0.0   \n",
       "2                           0.0                                    0.0   \n",
       "3                           0.0                                    0.0   \n",
       "4                           0.0                                    0.0   \n",
       "\n",
       "   12 - Terrestrial ES                                        title_lcase  \n",
       "0                  0.0  keyfactorsinfluencingtheincidenceofwestnilevir...  \n",
       "1                  1.0  asymmetricresponsesofplantcommunitystructurean...  \n",
       "2                  1.0  climateeffectsonfireregimesandtreerecruitmenti...  \n",
       "3                  0.0  climatechangeadaptationandmitigationahitherton...  \n",
       "4                  0.0  caribbeanmangrovesadjusttorisingsealevelthroug...  "
      ]
     },
     "execution_count": 129,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "filepath = hf_hub_download(repo_id=\"dspoka/ccai-nlp-tutorial-1\", filename=\"labelled_data.csv\", repo_type=\"dataset\")\n",
    "with open(filepath) as f:\n",
    "    labeled_data = pd.read_csv(f)\n",
    "\n",
    "# We create a lower case version of the title without spaces and punctuation, to allow for merging\n",
    "labeled_data[\"title_lcase\"] = labeled_data[\"title\"].apply(lambda x: re.sub(\"\\W\", \"\", x).lower() if isinstance(x, str) else x)\n",
    "\n",
    "# We also want to get rid of documents without abstracts, as we can't use these for learning\n",
    "labeled_data = labeled_data.dropna(subset=[\"abstract\"])\n",
    "print(labeled_data.shape)\n",
    "labeled_data.head()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 130,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "seen\n",
       "1.0    2500\n",
       "Name: count, dtype: int64"
      ]
     },
     "execution_count": 130,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "labeled_data['seen'].value_counts()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 131,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The dataset has 18556 rows and 8 columns\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>OA_id</th>\n",
       "      <th>doi</th>\n",
       "      <th>title</th>\n",
       "      <th>publication_year</th>\n",
       "      <th>abstract</th>\n",
       "      <th>authors</th>\n",
       "      <th>title_lcase</th>\n",
       "      <th>seen</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>https://openalex.org/W2122582889</td>\n",
       "      <td>https://doi.org/10.1073/pnas.0505734102</td>\n",
       "      <td>Regional vegetation die-off in response to glo...</td>\n",
       "      <td>2005</td>\n",
       "      <td>Future drought is projected to occur under war...</td>\n",
       "      <td>David D. Breshears, Neil S. Cobb, Paul Rich, K...</td>\n",
       "      <td>regionalvegetationdieoffinresponsetoglobalchan...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>https://openalex.org/W2170327428</td>\n",
       "      <td>https://doi.org/10.1073/pnas.1422385112</td>\n",
       "      <td>Anthropogenic warming has increased drought ri...</td>\n",
       "      <td>2015</td>\n",
       "      <td>Significance California ranks first in the Uni...</td>\n",
       "      <td>Noah S. Diffenbaugh, Daniel L. Swain, Danielle...</td>\n",
       "      <td>anthropogenicwarminghasincreaseddroughtriskinc...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>https://openalex.org/W2145638163</td>\n",
       "      <td>https://doi.org/10.1098/rstb.2010.0021</td>\n",
       "      <td>Community and ecosystem responses to recent cl...</td>\n",
       "      <td>2010</td>\n",
       "      <td>There is ample evidence for ecological respons...</td>\n",
       "      <td>Gian-Reto Walther</td>\n",
       "      <td>communityandecosystemresponsestorecentclimatec...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>https://openalex.org/W2120998920</td>\n",
       "      <td>https://doi.org/10.1016/j.tree.2004.07.021</td>\n",
       "      <td>Emerging infectious diseases of plants: pathog...</td>\n",
       "      <td>2004</td>\n",
       "      <td>Emerging infectious diseases (EIDs) pose threa...</td>\n",
       "      <td>Pamela Sue Anderson, Andrew A. Cunningham, Nik...</td>\n",
       "      <td>emerginginfectiousdiseasesofplantspathogenpoll...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>https://openalex.org/W2085588880</td>\n",
       "      <td>https://doi.org/10.1111/j.1365-2486.2007.01404.x</td>\n",
       "      <td>Influences of species, latitudes and methodolo...</td>\n",
       "      <td>2007</td>\n",
       "      <td>New analyses are presented addressing the glob...</td>\n",
       "      <td>Camille Parmesan</td>\n",
       "      <td>influencesofspecieslatitudesandmethodologieson...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                              OA_id   \n",
       "2  https://openalex.org/W2122582889  \\\n",
       "3  https://openalex.org/W2170327428   \n",
       "6  https://openalex.org/W2145638163   \n",
       "7  https://openalex.org/W2120998920   \n",
       "8  https://openalex.org/W2085588880   \n",
       "\n",
       "                                                doi   \n",
       "2           https://doi.org/10.1073/pnas.0505734102  \\\n",
       "3           https://doi.org/10.1073/pnas.1422385112   \n",
       "6            https://doi.org/10.1098/rstb.2010.0021   \n",
       "7        https://doi.org/10.1016/j.tree.2004.07.021   \n",
       "8  https://doi.org/10.1111/j.1365-2486.2007.01404.x   \n",
       "\n",
       "                                               title  publication_year   \n",
       "2  Regional vegetation die-off in response to glo...              2005  \\\n",
       "3  Anthropogenic warming has increased drought ri...              2015   \n",
       "6  Community and ecosystem responses to recent cl...              2010   \n",
       "7  Emerging infectious diseases of plants: pathog...              2004   \n",
       "8  Influences of species, latitudes and methodolo...              2007   \n",
       "\n",
       "                                            abstract   \n",
       "2  Future drought is projected to occur under war...  \\\n",
       "3  Significance California ranks first in the Uni...   \n",
       "6  There is ample evidence for ecological respons...   \n",
       "7  Emerging infectious diseases (EIDs) pose threa...   \n",
       "8  New analyses are presented addressing the glob...   \n",
       "\n",
       "                                             authors   \n",
       "2  David D. Breshears, Neil S. Cobb, Paul Rich, K...  \\\n",
       "3  Noah S. Diffenbaugh, Daniel L. Swain, Danielle...   \n",
       "6                                  Gian-Reto Walther   \n",
       "7  Pamela Sue Anderson, Andrew A. Cunningham, Nik...   \n",
       "8                                   Camille Parmesan   \n",
       "\n",
       "                                         title_lcase  seen  \n",
       "2  regionalvegetationdieoffinresponsetoglobalchan...     0  \n",
       "3  anthropogenicwarminghasincreaseddroughtriskinc...     0  \n",
       "6  communityandecosystemresponsestorecentclimatec...     0  \n",
       "7  emerginginfectiousdiseasesofplantspathogenpoll...     0  \n",
       "8  influencesofspecieslatitudesandmethodologieson...     0  "
      ]
     },
     "execution_count": 131,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Now we load the Open alex data, and create the same title variable for merging\n",
    "\n",
    "unlabeled_data = unlabeled_data.rename(columns={\"id\": \"OA_id\"})\n",
    "unlabeled_data[\"title_lcase\"] = unlabeled_data[\"title\"].apply(lambda x: re.sub(\"\\W\", \"\", x).lower() if isinstance(x, str) else x)\n",
    "unlabeled_data = unlabeled_data.dropna(subset=[\"abstract\"])\n",
    "unlabeled_data.loc[:, \"seen\"] = 0\n",
    "print(f'The dataset has {unlabeled_data.shape[0]} rows and {unlabeled_data.shape[1]} columns')\n",
    "unlabeled_data.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 132,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(20771, 15)\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>id</th>\n",
       "      <th>abstract</th>\n",
       "      <th>title</th>\n",
       "      <th>seen</th>\n",
       "      <th>INCLUDE</th>\n",
       "      <th>12 - Coastal and marine Ecosystems</th>\n",
       "      <th>12 - Human and managed</th>\n",
       "      <th>12 - Mountains, snow and ice</th>\n",
       "      <th>12 - Rivers, lakes, and soil moisture</th>\n",
       "      <th>12 - Terrestrial ES</th>\n",
       "      <th>title_lcase</th>\n",
       "      <th>OA_id</th>\n",
       "      <th>doi</th>\n",
       "      <th>publication_year</th>\n",
       "      <th>authors</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1448689.0</td>\n",
       "      <td>The city of Bismarck, North Dakota has one of ...</td>\n",
       "      <td>Key Factors Influencing the Incidence of West ...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>keyfactorsinfluencingtheincidenceofwestnilevir...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>3281128.0</td>\n",
       "      <td>Changing precipitation regimes can profoundly ...</td>\n",
       "      <td>Asymmetric responses of plant community struct...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>asymmetricresponsesofplantcommunitystructurean...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>900708.0</td>\n",
       "      <td>Climate influences forest structure through ef...</td>\n",
       "      <td>Climate effects on fire regimes and tree recru...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>climateeffectsonfireregimesandtreerecruitmenti...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3234419.0</td>\n",
       "      <td>Climate change is likely to bring more, hotter...</td>\n",
       "      <td>Climate change adaptation and mitigation ? a h...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>climatechangeadaptationandmitigationahitherton...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>2090028.0</td>\n",
       "      <td>Aim The long-term stability of coastal ecosyst...</td>\n",
       "      <td>Caribbean mangroves adjust to rising sea level...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>caribbeanmangrovesadjusttorisingsealevelthroug...</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "          id                                           abstract   \n",
       "0  1448689.0  The city of Bismarck, North Dakota has one of ...  \\\n",
       "1  3281128.0  Changing precipitation regimes can profoundly ...   \n",
       "2   900708.0  Climate influences forest structure through ef...   \n",
       "3  3234419.0  Climate change is likely to bring more, hotter...   \n",
       "4  2090028.0  Aim The long-term stability of coastal ecosyst...   \n",
       "\n",
       "                                               title  seen  INCLUDE   \n",
       "0  Key Factors Influencing the Incidence of West ...   1.0      0.0  \\\n",
       "1  Asymmetric responses of plant community struct...   1.0      1.0   \n",
       "2  Climate effects on fire regimes and tree recru...   1.0      1.0   \n",
       "3  Climate change adaptation and mitigation ? a h...   1.0      1.0   \n",
       "4  Caribbean mangroves adjust to rising sea level...   1.0      1.0   \n",
       "\n",
       "   12 - Coastal and marine Ecosystems  12 - Human and managed   \n",
       "0                                 0.0                     1.0  \\\n",
       "1                                 0.0                     0.0   \n",
       "2                                 0.0                     0.0   \n",
       "3                                 0.0                     1.0   \n",
       "4                                 1.0                     0.0   \n",
       "\n",
       "   12 - Mountains, snow and ice  12 - Rivers, lakes, and soil moisture   \n",
       "0                           0.0                                    0.0  \\\n",
       "1                           0.0                                    0.0   \n",
       "2                           0.0                                    0.0   \n",
       "3                           0.0                                    0.0   \n",
       "4                           0.0                                    0.0   \n",
       "\n",
       "   12 - Terrestrial ES                                        title_lcase   \n",
       "0                  0.0  keyfactorsinfluencingtheincidenceofwestnilevir...  \\\n",
       "1                  1.0  asymmetricresponsesofplantcommunitystructurean...   \n",
       "2                  1.0  climateeffectsonfireregimesandtreerecruitmenti...   \n",
       "3                  0.0  climatechangeadaptationandmitigationahitherton...   \n",
       "4                  0.0  caribbeanmangrovesadjusttorisingsealevelthroug...   \n",
       "\n",
       "  OA_id  doi  publication_year authors  \n",
       "0   NaN  NaN               NaN     NaN  \n",
       "1   NaN  NaN               NaN     NaN  \n",
       "2   NaN  NaN               NaN     NaN  \n",
       "3   NaN  NaN               NaN     NaN  \n",
       "4   NaN  NaN               NaN     NaN  "
      ]
     },
     "execution_count": 132,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Now we add the OpenAlex rows which are not in the labelled data to the labelled_data\n",
    "# print(unlabeled_data[~unlabeled_data[\"title_lcase\"].isin(labeled_data[\"title_lcase\"])].shape)\n",
    "\n",
    "df = pd.concat(\n",
    "    [labeled_data, unlabeled_data[~unlabeled_data[\"title_lcase\"].isin(labeled_data[\"title_lcase\"])]]\n",
    ")\n",
    "print(df.shape)\n",
    "df.head()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Inspecting our labels\n",
    "\n",
    "We can take a quick look at how many of which type of labels we have"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 133,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of seen vs. unseen documents\n",
      "seen\n",
      "0.0    18262\n",
      "1.0     2500\n",
      "Name: title, dtype: int64\n",
      "--------------------------------------------------------------------------------\n",
      "Number of included vs. excluded documents\n",
      "INCLUDE\n",
      "0.0    1194\n",
      "1.0    1306\n",
      "Name: title, dtype: int64\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def inspect_labels(df):\n",
    "    # We can quickly plot how many documents have been labelled, and how many not\n",
    "    # df.groupby(\"seen\")[\"title\"].count().plot.bar();\n",
    "    print('Number of seen vs. unseen documents')\n",
    "    print(df.groupby(\"seen\")[\"title\"].count())\n",
    "    print('-'*80)\n",
    "    # and also plot how many of the labelled docs were included\n",
    "    # df.groupby(\"INCLUDE\")[\"title\"].count().plot.bar()\n",
    "\n",
    "    print('Number of included vs. excluded documents')\n",
    "    print(df.groupby(\"INCLUDE\")[\"title\"].count())\n",
    "    # Finally we can plot how many examples of each impact type we have\n",
    "    df[impacts].sum().plot.barh();\n",
    "impacts = [x for x in df.columns if re.match(\"12 -\", x)]\n",
    "inspect_labels(df)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 134,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "def train_test_split_df(df):\n",
    "    ''' Gets train/test splits for INCLUDE prediction task.\n",
    "    '''\n",
    "    # select only the seen documents\n",
    "    all_seen = df[df[\"seen\"] == 1].sample(2000).index\n",
    "    all_x = df[df[\"seen\"] == 1][\"abstract\"].values\n",
    "    all_y = df[df[\"seen\"] == 1][\"INCLUDE\"].values.astype(int)\n",
    "    # percent of data to use for train, val, test\n",
    "    train_size = 0.8\n",
    "    # train_size = 0.01\n",
    "    # get the indices for each split\n",
    "    x_train, x_test, y_train, y_test = train_test_split(all_x, all_y, train_size=train_size, random_state=42)\n",
    "    assert len(x_train) == len(y_train)\n",
    "    assert len(x_test) == len(y_test)\n",
    "    print (f\"Train size: {len(x_train)}\")\n",
    "    print (f\"Test size: {len(x_test)}\")\n",
    "    return x_train, y_train, x_test, y_test\n",
    "\n",
    "    "
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## A pipeline for a simple machine learning model\n",
    "\n",
    "Now we want to put together a pipeline for a simple machine learning model. We'll start off by trying to predict the binary INCLUSION label, then move onto the impact type, which is a **multilabel** output. This means each document can be labelled with one or more of the impact types. \n",
    "\n",
    "In each case, though, the first thing we need to do is to encode our text input, so it can be used in our models. To do this, we'll take a look at approaches from the feature extraction module from [scikitlearn](https://scikit-learn.org/stable/modules/classes.html#module-sklearn.feature_extraction.text): `CountVectorizer` and `TfidfVectorizer`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 135,
   "metadata": {},
   "outputs": [],
   "source": [
    "def visualize_doc_vec(X, features=None):\n",
    "    '''Simple function to visualise a document-term matrix'''\n",
    "    print('Document-term matrix:\\n', X.todense())\n",
    "    # And we could visualise this simply like this\n",
    "    fig, ax = plt.subplots()\n",
    "    doc_map = ax.imshow(X.todense(), cmap=\"Blues\")\n",
    "    ax.set_xticks(np.arange(len(features)))\n",
    "    ax.set_xticklabels(features, rotation=45, ha=\"right\")\n",
    "    cbar = fig.colorbar(doc_map)\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 136,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Document-term matrix:\n",
      " [[0 0 1 ... 0 0 1]\n",
      " [1 1 1 ... 0 0 0]\n",
      " [0 0 2 ... 1 1 0]]\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from nltk.stem.porter import PorterStemmer\n",
    "from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer\n",
    "\n",
    "\n",
    "def show_raw_counts(texts):\n",
    "    '''\n",
    "    Function to show the raw counts of words in a corpus and visualize them.\n",
    "    '''\n",
    "    # We'll instantiate a vectorizer class\n",
    "    vectorizer = CountVectorizer()\n",
    "\n",
    "    # And apply the fit_transform method to our texts\n",
    "    X = vectorizer.fit_transform(texts)\n",
    "\n",
    "    # This gives us a matrix of shape n_texts, vocab_size\n",
    "    X.todense()\n",
    "\n",
    "    # The features each column represents can be accessed here\n",
    "    features = vectorizer.get_feature_names_out()\n",
    "    visualize_doc_vec(X, features)\n",
    "\n",
    "# Let's take for now a couple of example texts\n",
    "texts = [\n",
    "    \"Climate change is impacting human systems\",\n",
    "    \"Climate change is caused by fossil fuels\",\n",
    "    \"System change not climate change\",\n",
    "]\n",
    "# And now lets visualize the raw word counts in our 3 sentences.\n",
    "show_raw_counts(texts)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "As we can see, each document is a row, each column is a feature. The cells contain the number of times each word occurs in each document. \n",
    "\n",
    "However, by adding a small amount of complexity we can generate a more useful way of representing these texts. \n",
    "\n",
    "Term frequency inverse document frequency ([TF-IDF](https://en.wikipedia.org/wiki/Tf%E2%80%93idf)) gives greater weight to those terms which appear infrequently in the corpus, meaning that very frequent but uninformative terms like \"the\" are given less importance."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 137,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Document-term matrix:\n",
      " [[0.         0.         0.28561676 ... 0.         0.         0.48359121]\n",
      " [0.43535684 0.43535684 0.25712876 ... 0.         0.         0.        ]\n",
      " [0.         0.         0.61046311 ... 0.51680194 0.51680194 0.        ]]\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAf4AAAGOCAYAAAB2TWHkAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAABWfklEQVR4nO3deVxN+f8H8Pe57UZlSTUSMZYQorTYmYjGOoMsIxNiLDNM9jVjmRikQdmN7zCUQXZhwhgjUjEaWQdJFFkqoei+f3/0u2e6yozcc7vlvJ7zOI/HOJ17zucu57zP5/1ZjsDMTAAAACALCl0XAAAAAEoOAj8AAICMIPADAADICAI/AACAjCDwAwAAyAgCPwAAgIwg8AMAAMgIAj8AAICM6Ou6AAAAALr04sULys3NlWRfhoaGZGxsLMm+tAWBHwAAZOvFixdkYlqZ6NUzSfZnbW1NN2/eLNXBH4EfAABkKzc3l+jVMzJq6EukZ6jZzvJyKfXij5Sbm1uqAz/a+AEAAPQMSdBweZcbh5CQELKzsyNjY2NydXWlmJiYf93+yZMnNHr0aPrwww/JyMiI6tatSwcOHCjWMVHjBwAAEIhIEDTfRzGEh4eTv78/rVq1ilxdXSk4OJg8PT3pypUrZGlpWWj73Nxc6tixI1laWtL27dvJxsaGkpKSqEKFCsUrJp7OBwAAcpWZmUnm5uZk5DiSBD0jjfbFeTmUc34lZWRkkJmZ2X9u7+rqSs2bN6cVK1YQEZFSqSRbW1v66quvaMqUKYW2X7VqFS1atIguX75MBgYG71xOpPoBAAAklJmZqbbk5OQU2iY3N5fi4uLIw8NDXKdQKMjDw4Oio6OL3O+ePXvI3d2dRo8eTVZWVuTg4EDfffcd5eXlFat8CPwAAACCIM1CRLa2tmRubi4ugYGBhQ6Xnp5OeXl5ZGVlpbbeysqKUlNTiyzijRs3aPv27ZSXl0cHDhygmTNn0pIlS2jevHnFeqto4wcAABAU+Yum+yCi5ORktVS/kZFmTQgqSqWSLC0tac2aNaSnp0dOTk6UkpJCixYtooCAgLfeDwI/AACAhMzMzP6zjd/CwoL09PQoLS1NbX1aWhpZW1sX+ZoPP/yQDAwMSE9PT1xXv359Sk1NpdzcXDI0fLtRBUj1AwAASJjqfxuGhobk5OREUVFR4jqlUklRUVHk7u5e5GtatmxJ169fJ6VSKa67evUqffjhh28d9IkQ+AEAAIhI8U+6/12XYoZUf39/Wrt2Lf3vf/+jS5cu0ciRIyk7O5t8fX2JiMjHx4emTp0qbj9y5Eh69OgRjR07lq5evUr79++n7777jkaPHl2s4yLVDwAAUMwa+xv3UQze3t704MEDmjVrFqWmppKjoyNFRkaKHf5u375NCsU/NxO2trZ06NAh+uabb6hx48ZkY2NDY8eOpcmTJxevmBjHDwAAciWO43ceR4K+huP4X+VQTmzwW4/j1xXU+AEAACTs1V/aIfADAADoINWvK2Xj9gQAAAAkgRo/AAAAUv0AAAAyglQ/AAAAvI9Q4wcAAECqHwAAQEYEQYLAj1Q/AAAAlDKo8QMAACiE/EXTfZQBCPwAAABo4wcAAJARDOcDAACA9xFq/AAAAEj1AwAAyAhS/QAAAPA+Qo0fAAAAqX4AAAAZQaofAAAA3keo8QMAACDVDwAAICNI9QMAAMD7CDV+AAAAkiDVX0bq0gj8AAAAMkr1I/ADAAAIggSd+8pG4C8beQkAAACQBGr8AAAAGM4HAAAgIzJq4y8btycAAAAgCdT4AQAAkOoHAACQEaT6AQAA4H2EGj8AAABS/QAAADKCVD8AAAC8j1DjBwAA2RMEgQSZ1PgR+AEAQPYQ+AEAAORE+P9F032UAWjjBwAAkBHU+AEAQPaQ6gcAAJAROQV+pPoBAABkBDV+AACQPTnV+BH4AQBA9uQU+JHqBwAAkBHU+AEAAGQ0jh+BHwAAZA+pfgAAAHgvocYPAACyl/9UXk1r/NKURdtQ4wcAANkTSBDT/e+8vEPkDwkJITs7OzI2NiZXV1eKiYl547YbN24sdExjY+NiHxOBHwAAZE/joP8OfQTCw8PJ39+fAgICKD4+npo0aUKenp50//79N77GzMyM7t27Jy5JSUnFfq8I/AAAADoQFBREfn5+5OvrSw0aNKBVq1ZRuXLlaMOGDW98jSAIZG1tLS5WVlbFPi4CPwAAgCDR8pZyc3MpLi6OPDw8xHUKhYI8PDwoOjr6ja97+vQp1ahRg2xtbalHjx508eLFYrzJ/z9OsV8BAADwvpEizf//qf7MzEy1JScnp9Dh0tPTKS8vr1CN3crKilJTU4ssYr169WjDhg20e/du2rx5MymVSmrRogXduXOnWG8VgR8AAEBCtra2ZG5uLi6BgYGS7Nfd3Z18fHzI0dGR2rZtSzt37qQqVarQ6tWri7UfDOcDAADZk2ICH9Xrk5OTyczMTFxvZGRUaFsLCwvS09OjtLQ0tfVpaWlkbW39VsczMDCgpk2b0vXr14tVTtT4AQBA9qTs1W9mZqa2FBX4DQ0NycnJiaKiosR1SqWSoqKiyN3d/a3KnJeXRwkJCfThhx8W672ixg8AAKAD/v7+NHjwYHJ2diYXFxcKDg6m7Oxs8vX1JSIiHx8fsrGxEZsK5syZQ25ublS7dm168uQJLVq0iJKSkmjYsGHFOi4CPwAAgA4e0uPt7U0PHjygWbNmUWpqKjk6OlJkZKTY4e/27dukUPyTmH/8+DH5+flRamoqVaxYkZycnOjUqVPUoEGD4hWTmbl4RQUAAHg/ZGZmkrm5OVkM2kgKw3Ia7UuZ+4zSN31BGRkZam38pQ3a+AEAAGQEqX4AAJA9KXv1l3YI/AAAIHsI/AAAADIip8CPNn4AAAAZQY0fAABAB8P5dAWBHwAAZA+pfgAAAHgvocYPAACyJ6caPwI/AADInpwCP1L9AAAAMoIaPwAAAHr1AwAAyAdS/QAAAPBeQo0fAABkT041fgR+AACQPYEkCPxlpJEfgR8AAGRPTjV+tPEDAADICGr8AAAAGM4HAAAgH0j1AwAAwHsJNX4AAJA9OdX4EfgBAED2BCF/0XQfZQFS/QAAADKCGj8AAMhefo1f01S/RIXRMgR+AAAACVL9ZWU4H1L9AAAAMoIaPwAAyB569QMAAMiInHr1I/ADAIDsKRQCKRSaRW7W8PUlBW38AAAAMoIaPwAAyB5S/QAAADIip859SPUDAADICGr8AAAge0j1AwAAyAhS/QAAAPBeQo0fAABkT041fgR+AACQPTm18SPVDwAAICOo8QMAgOwJJEGqv4w8lxeBHwAAZE9OqX4EfgAAkD05de5DGz8AAICMoMYPAACyh1Q/AACAjCDVDwAAAO8l1PgBAED2kOoHAACQEaT6AQAA4L2EwA8AACD8k+5/1+VdJu4LCQkhOzs7MjY2JldXV4qJiXmr14WFhZEgCNSzZ89iHxOBHwAAZE+V6td0KY7w8HDy9/engIAAio+PpyZNmpCnpyfdv3//X19369YtmjBhArVu3fqd3isCPwAAgA4EBQWRn58f+fr6UoMGDWjVqlVUrlw52rBhwxtfk5eXRwMHDqRvv/2WatWq9U7HReAHAADZ0zTNX3BUQGZmptqSk5NT6Hi5ubkUFxdHHh4e4jqFQkEeHh4UHR39xnLOmTOHLC0taejQoe/8XhH4AQBA9qRM9dva2pK5ubm4BAYGFjpeeno65eXlkZWVldp6KysrSk1NLbKMJ0+epPXr19PatWs1eq8YzgcAALIn5Tj+5ORkMjMzE9cbGRlptmMiysrKokGDBtHatWvJwsJCo30h8AMAAEjIzMxMLfAXxcLCgvT09CgtLU1tfVpaGllbWxfa/u+//6Zbt25Rt27dxHVKpZKIiPT19enKlSv00UcfvVX5kOoHAADZK+le/YaGhuTk5ERRUVHiOqVSSVFRUeTu7l5oe3t7e0pISKDz58+LS/fu3al9+/Z0/vx5srW1fetjo8YPAACyp4uZ+/z9/Wnw4MHk7OxMLi4uFBwcTNnZ2eTr60tERD4+PmRjY0OBgYFkbGxMDg4Oaq+vUKECEVGh9f8FgR8AAEAHvL296cGDBzRr1ixKTU0lR0dHioyMFDv83b59mxQK6RPzAjOz5HsFAAAoAzIzM8nc3JxafHeI9I0/0Ghfr15k06lpnpSRkfGfbfy6VOI1fqVSSXfv3iVTU9My80ADAAAoecxMWVlZVLVqVa3UfAuS00N6Sjzw3717t1idEAAAQN6Sk5OpWrVqui7Ge6PEA7+pqSkREV2/mUympTgVUhIWHL2m6yKUClM61NF1EUqFRuN367oIpUbjRja6LgL9/IWzrotQKujyd6nMfU7pW0aIcUObpBzHX9qVeOBXpUJM32Kc4/vOqFx5XRehVJD770BFYVhO10UoNfRNNGtrlQJ+l/lKw++yJFLockr1Yxw/AACAjGA4HwAAyJ5AEqT6JSmJ9iHwAwCA7CkEgRQaRn5NX19SEPgBAED25NS5D238AAAAMoIaPwAAyJ6cevUj8AMAgOwphPxF032UBUj1AwAAyAhq/AAAAIIEqfr3ucYfEhJCdnZ2ZGxsTK6urhQTEyN1uQAAAEqMqle/pktZUOzAHx4eTv7+/hQQEEDx8fHUpEkT8vT0pPv372ujfAAAACChYgf+oKAg8vPzI19fX2rQoAGtWrWKypUrRxs2bNBG+QAAALROkOi/sqBYgT83N5fi4uLIw8Pjnx0oFOTh4UHR0dGSFw4AAKAkqHr1a7qUBcXq3Jeenk55eXlkZWWltt7KyoouX75c5GtycnIoJydH/HdmZuY7FBMAAEB75DSOX+vD+QIDA8nc3FxcbG1ttX1IAAAAeINiBX4LCwvS09OjtLQ0tfVpaWlkbW1d5GumTp1KGRkZ4pKcnPzupQUAANAC9Op/A0NDQ3JycqKoqChxnVKppKioKHJ3dy/yNUZGRmRmZqa2AAAAlCaqp/NpupQFxZ7Ax9/fnwYPHkzOzs7k4uJCwcHBlJ2dTb6+vtooHwAAAEio2IHf29ubHjx4QLNmzaLU1FRydHSkyMjIQh3+AAAAygo5PZb3nabsHTNmDI0ZM0bqsgAAAOgEevUDAADAewkP6QEAANlDqh8AAEBGpOiVX1Z69SPVDwAAICOo8QMAgOwJ/79ouo+yAIEfAABkT069+hH4AQBA9qR4ul5ZeTof2vgBAABkBDV+AACQPaT6S0Cb76JIz+gDXR2+VPDtVFvXRSgVnAIO67oIpcLNkM90XYRSY3j4n7ouAvw/Xf4uMzMzyWpjyR2vjMRtjSHVDwAAICNI9QMAgOwh1Q8AACAj6NUPAAAA7yXU+AEAQPaQ6gcAAJAROU3Zi1Q/AACAjKDGDwAAsienx/Ii8AMAgOwJguYT+JSRuI/ADwAAIKfOfWjjBwAAkBHU+AEAQPbklOovdo3/xIkT1K1bN6patSoJgkC7du3SQrEAAABKjqpzn6ZLWVDswJ+dnU1NmjShkJAQbZQHAAAAtKjYqf4uXbpQly5dtFEWAAAAnUCqHwAAQEZUvfo1XYorJCSE7OzsyNjYmFxdXSkmJuaN2+7cuZOcnZ2pQoUK9MEHH5CjoyNt2rSp2MfUeue+nJwcysnJEf+dmZmp7UMCAACUeuHh4eTv70+rVq0iV1dXCg4OJk9PT7py5QpZWloW2r5SpUo0ffp0sre3J0NDQ9q3bx/5+vqSpaUleXp6vvVxtV7jDwwMJHNzc3GxtbXV9iEBAACKRSHRUhxBQUHk5+dHvr6+1KBBA1q1ahWVK1eONmzYUOT27dq1o169elH9+vXpo48+orFjx1Ljxo3p5MmTxX6vWjV16lTKyMgQl+TkZG0fEgAAoFhKOtWfm5tLcXFx5OHhIa5TKBTk4eFB0dHR//l6ZqaoqCi6cuUKtWnTpljvVeupfiMjIzIyMtL2YQAAAEqF15u0i4qD6enplJeXR1ZWVmrrrays6PLly2/cd0ZGBtnY2FBOTg7p6elRaGgodezYsVjlK3bgf/r0KV2/fl38982bN+n8+fNUqVIlql69enF3BwAAoHOCQKSQqFf/603aAQEBNHv2bM12/v9MTU3p/Pnz9PTpU4qKiiJ/f3+qVasWtWvX7q33UezAHxsbS+3btxf/7e/vT0REgwcPpo0bNxZ3dwAAADqnkCDwq16fnJxMZmZm4vqist4WFhakp6dHaWlpauvT0tLI2tr6zcdQKKh27dpEROTo6EiXLl2iwMBA7Qb+du3aETMX92UAAACllpQP6TEzM1ML/EUxNDQkJycnioqKop49exIRkVKppKioKBozZsxbH1OpVKqNnHsbmKsfAABAB/z9/Wnw4MHk7OxMLi4uFBwcTNnZ2eTr60tERD4+PmRjY0OBgYFElD9KztnZmT766CPKycmhAwcO0KZNm2jlypXFOi4CPwAAyJ6Uqf635e3tTQ8ePKBZs2ZRamoqOTo6UmRkpNjh7/bt26RQ/DP4Ljs7m0aNGkV37twhExMTsre3p82bN5O3t3exjovADwAAsqerKXvHjBnzxtT+8ePH1f49b948mjdv3juUTB2m7AUAAJAR1PgBAED2pHisbll5LC8CPwAAyN67TLlb1D7KgrJSTgAAAJAAavwAACB7uurcpwsI/AAAIHsKkqCNn8pG5EeqHwAAQEZ0VuM/Me3j/5zS8H1XsfnbT8uoLUNmjtJ1ESju2066LgKUMmu8m+i6CGQ56CddF6FUuL/JR9dFKBFI9QMAAMiILmbu0xUEfgAAkL38x/Jq+pAeiQqjZWjjBwAAkBHU+AEAQPbQxg8AACAjcmrjR6ofAABARlDjBwAA2RP+/z9N91EWIPADAIDsIdUPAAAA7yXU+AEAQPbkVONH4AcAANkTBIEEjSfwKRuRv1ip/sDAQGrevDmZmpqSpaUl9ezZk65cuaKtsgEAAIDEihX4f/vtNxo9ejSdPn2ajhw5Qi9fvqROnTpRdna2tsoHAACgdapUv6ZLWVCsVH9kZKTavzdu3EiWlpYUFxdHbdq0kbRgAAAAJQUz972ljIwMIiKqVKmSJIUBAADQBYUgaPyQHk1fX1LeOfArlUoaN24ctWzZkhwcHN64XU5ODuXk5Ij/zszMfNdDAgAAgIbeeRz/6NGj6a+//qKwsLB/3S4wMJDMzc3FxdbW9l0PCQAAoBVyauN/p8A/ZswY2rdvHx07doyqVav2r9tOnTqVMjIyxCU5OfmdCgoAAKA1wj/t/O+6lJEZe4uX6mdm+uqrrygiIoKOHz9ONWvW/M/XGBkZkZGR0TsXEAAAAKRTrMA/evRo2rJlC+3evZtMTU0pNTWViIjMzc3JxMREKwUEAADQNgUJpNCwyq7p60tKsVL9K1eupIyMDGrXrh19+OGH4hIeHq6t8gEAAGidpml+KYYDlpRip/oBAACg7MJc/QAAIHt4SA8AAICMyGkCn3cexw8AAABlD2r8AAAge5irHwAAQEYUJEGqv4wM50PgBwAA2ZNTjR9t/AAAADKCGj8AAMiegjSvCZeVmjQCPwAAyJ4gCCRomKvX9PUlpcQDv2r2v6zMzJI+dKnDebm6LgLlPnuq6yJQJn4LUArxy+e6LkKpoMvzUxUnMGustEo88GdlZRERUe2atiV9aCjCukFrdV0EWqfrAgDAG1nt+FLXRaCsrCwyNzfX6jGkeKpu2ajv6yDwV61alZKTk8nU1PSd0iKZmZlka2tLycnJZGZmpoUSogwoQ9krQ2kpB8qAMkhZBmamrKwsqlq1qhZKp05OM/eVeOBXKBRUrVo1jfdjZmam04ssyoAylMYylJZyoAwog1Rl0HZNX47QuQ8AAIDKTqpeUwj8AAAge5jApxQzMjKigIAAMjIyQhlQBpShlJUDZUAZSlsZoDCBMU4CAABkKjMzk8zNzWndiUtUrrypRvt69jSLhrWpTxkZGTrvV/FvkOoHAADZw8x9AAAAMiKnmfvKyg0KAAAASACBHwAAZE+QaCmukJAQsrOzI2NjY3J1daWYmJg3brt27Vpq3bo1VaxYkSpWrEgeHh7/uv2bIPADAIDsqVL9mi7FER4eTv7+/hQQEEDx8fHUpEkT8vT0pPv37xe5/fHjx6l///507Ngxio6OJltbW+rUqROlpKQU772iV3++xYsXU+3atalnz566LorsBQYGkiAINGXKFF0XBQBew8xlpi37bah69f/v5BVJevUPblXvrXv1u7q6UvPmzWnFihVERKRUKsnW1pa++uqrt7r+5eXlUcWKFWnFihXk4+Pz1uUsFTV+pVKp6yJQbGws9evXjyIjI3VWhqLuweR4X/bq1SuaNm0aLV++XGdlUH3u9+7do1evXumsHKVBaTg/S4uXL18SUf4Ft6SVhu9BqVSKQT8pKUnHpZGWQqKFKP9mouCSk5NT6Hi5ubkUFxdHHh4e/5RBoSAPDw+Kjo5+qzI/e/aMXr58SZUqVSr2e9U5hSK/GJs2baIjR44QUckHvLCwMBo8eDD17t2bDh48WKLHJlI/oe7evUtJSUk6ubNWfe5JSUkUGxtL165do8ePH6v9TVvu3r1LREQzZ86kpUuX0rhx42jZsmVaPWZRVJ/73r17ycvLiw4cOFDkiauLcumC6vxMTU3VyfGJCr/3kg6Cd+7coUePHpGBgQHt27ePtmzZUqI3hEqlUvweYmJi6PLly3Tnzp0SO/7rZZgzZw55e3tTXFxciZZBm6RM9dva2pK5ubm4BAYGFjpeeno65eXlkZWVldp6Kyurtz7XJk+eTFWrVlW7eXgbpSLwMzNlZGTQjBkz6OjRo0RUcsMimFm8iCxatIi6du1KQ4cOFW9ASqoMqhPq22+/pU8++YTat29PjRo1op9//lkMvCVRDkEQaOfOneTh4UGff/459e7dm3x8fCguLk6r38mcOXNo4MCB9OLFCyIiGjt2LH3//ff0zTfflFjwV/0OBEGgXbt2Uf/+/al///5kb29faOYxbQdh1f6vXLlCsbGx9Pfff5f4TWDB4BoREUF16tSh2NjYEi0D0T+/y2PHjtH8+fOJ6J+bkZKQmZlJfn5+5O3tTT/++CN1796dTExMSF+/5EZDq97v5MmTqUePHtSmTRsaMWIE7d69u8TLMGXKFFq1ahWNHz+eLCwsSuz4ZUlycjJlZGSIy9SpUyU/xoIFCygsLIwiIiLI2Ni4eC/mUkCpVDIz88aNG9nW1pbPnz9f4sfevn07t2zZknv06MEKhYIrVqzIBw4cKLFyMDPPnTuXLS0tedeuXZydnc1ubm5cu3ZtvnLlSomV4cSJE2xqasrLly9nZubVq1ezIAi8bNkyrR73yZMn/NdffzEzc2Zmpvi9LF68mBUKBf/www9aO/aFCxfE/1cqlZyamsqNGjXixYsXMzPzy5cv+cWLF3z48OES/S527NjBFSpU4Fq1arGxsTGvXr2ac3JySuTYeXl54v9v2rSJQ0JCWBAEdnBw4LNnz5ZIGZjVz08LCwsePXo0//nnn4X+rk2vXr3iiIgIrlu3LhsYGHBISAgz5/8utK3g+ztz5gw7ODjw6dOnecuWLTx48GBu0KABh4eHa70cKqdOnWI7Ozv+/fffmZk5JyeH09LS+MiRI3z37t0SK4eUMjIymIj45z+ucsSf9zRafv7jKhMRZ2Rk/Odxc3JyWE9PjyMiItTW+/j4cPfu3f/1tYsWLWJzc/N3Phd1EvhfP1lVF5nLly+zu7s7r169mpnzT7iScObMGTY2Nua1a9fyrVu3+Ny5c9yvXz8uX758iQX/R48ecZs2bXjr1q3MzBwZGclmZma8atUqZla/EGuDav+zZs1iX19fZma+c+cO16hRg0eNGiVu9/jxY8mPXfACeuzYMXZ2dub4+HhxnTaDf2RkJFepUoXXr18vrrt+/Trb29vzH3/8wWlpaTxv3jxu27Yt6+vrc+vWrXn37t2Sl0NFdW7cuXOH69evz2vWrOG4uDhesGABC4LAgYGB/OzZM60d/3VTp05la2trXrlyJU+bNo1dXFzYxsZGq8E/LCyML126JP771KlTbGZmxmvXrlXbriSCvuoYV69e5WrVqrGdnR336NGD09PTmVm716jXz/mTJ0/yyJEjxX+fP3+ehw0bxvb29iUW/Pfu3cs1atRgZubY2FieMmUK161blw0NDdnLy4tv3LhRIuWQkirwbzl1lXdduKfRsuXU2wd+ZmYXFxceM2aM+O+8vDy2sbHhwMDAN75m4cKFbGZmxtHR0e/8nnVa4//xxx95//79auvGjh3LNWrU4BcvXpRYOcLCwtjR0ZEzMzPFdS9evGBvb2+uVKkS//rrr1ovQ0pKCteuXZsfP37MR44c4fLly/PKlSuZmTk7O5uXL1/Oqampkhyr4AVFFXRVF7jx48fz7Nmz+d69e2xjY8PDhw8X/7Zv3z7esGGDVmudjx494kqVKnHLli3Vanaq4K/KREglISGBv/zyS27YsCH/+OOP4noHBweuW7cuW1pacq9evTgoKIivXbvG9erV42+//VbSMrzu8OHDHBwczKNGjVILLKGhoSUa/G/cuMHVq1fnbdu2ievu37/PnTp1YltbW60E/+TkZG7VqhXfvn1bXBcUFMQ9evRg5vzfx549e7hPnz7s7u7OO3bskLwMRXn48CFfvHiRt2/fzu7u7uzl5VUo+GvrvAgMDOTu3btzt27deMCAAWp/+/PPP9nPz48bNmzIGzdulPS4RVU2Hjx4wBUrVuQmTZpwpUqV2M/Pj8PCwvjChQtsYGDAe/bskbQMJUGXgT8sLIyNjIx448aNnJiYyMOHD+cKFSqI1/pBgwbxlClTxO0XLFjAhoaGvH37dr537564ZGVlFes9l2jgL/hDSk5O5u7du7MgCDxw4EBesmSJuL59+/Zi0CuJu/oNGzawsbExP3/+nJn/CYZ//PEHC4LA+vr6fOTIEcmOV/A9FbzBadeuHXfu3JlNTU153bp14vpbt25xq1ateOfOnZKVITk5WSzH3r17+bvvvmPm/OaGmjVrsq2trVpNPzc3l7/44gseP368pDdlqjJcv36d//77b2bOzyrUrFmTXVxc1IJ/UFAQC4IgZkGkcunSJR49ejTXq1dPrFW+ePGCv/vuO169ejU/evRI/E307t2b58yZo1Z2qU2ZMoUFQeCGDRvyo0eP1P4WGhrKhoaGPHPmTPH3qi2JiYlsbm7OZ86cYeZ/zt9bt26xra0tN27cWMzMSJmRUt3UXLhwgf/++2/etm0bC4LAmzZtYk9PT/by8uIBAwZw37592dzcnNPS0iQ7torqu01KSuJbt27x9evXmTn/fYaFhbGbmxt37dqVHz58yMzMy5cv582bN0vymyj4WQYGBnLlypV5yJAh3Lp1a/FzKOjChQvcu3fvQjcFUpXhzJkzfOnSJU5KSmLm/OzHjBkzeM+ePfzkyRNmzr/pcXd3L/HmUSmoAn/YqWu850KqRkvYqWvFCvzM+b+d6tWrs6GhIbu4uPDp06fFv7Vt25YHDx4s/rtGjRpMRIWWgICAYr3nEgv8BWsuN27c4PT0dH7y5AknJibymDFj2NbWll1dXfm7775jd3d39vPz00o5ijoxMzMzuXHjxvzFF1+o3TldvnyZ+/fvzyNHjlRLPWqi4Ak1f/58njlzpnh3t27dOq5evTp/8skn4jbZ2dns5eXFH3/8sWRpxWfPnrGDgwO3a9eOw8PDWRAEtVRhx44d2czMjG/fvi22b0+dOpWrVq3Kly9flqQMzP98FxEREdy0aVNesmQJ379/n5nzg7+dnV2h4L98+XJOTEyUrAwqFy9e5FGjRnG9evXUbrpUsrOzedq0aVy5cuUSaedXpfYLNkGoLFmyhCtVqiTWOKXwpoDVqFEj/uKLL9TWZWdns4eHB1etWpVtbW3Fi5yUN0IZGRncuHFjHjhwIEdFRfG0adPY2tqafX19+cSJE8ycn31o3LixZOemiup97Nixg+vWrcs1a9Zkc3NzHjlypBj8wsLCuFWrVtygQQMeMWIEC4LACQkJkpYjISGBly5dylFRUczMfPPmTR43bhybmpryzz//rLbt9evXtdIcOGnSJLa2tuYqVapwly5deNeuXWp/f/HiBT948IC9vLzY2dm5xJpnpaQK/OHR13hvQqpGS3h08QO/Lmg98IeGhqq1106aNInr1KnDFhYW3Lp1a96yZQsz53/4I0eO5EGDBrEgCCwIguRpPNUJffr0af7hhx94wYIFYnvtunXr2N3dnQcNGsSPHj3i1NRUnj59Ordv316ymlXBE/P27dvcr18/NjQ05MWLF/OLFy/4yZMnPH78eK5duza3adOGfXx8uGXLlty4cWPOzc1lZmnaFJVKJV+6dIkrVarExsbGYopQVZO/cOECN2rUiG1sbLh58+bcqVMntrKyUvsepbJnzx42NjbmZcuWFWrKePjwIdeoUYNbtmzJcXFxkh63qCB1+fJlHjVqFNetW1ct4O7atYs/+eQTrlGjhuSfgaoceXl5hb7bKVOmsIGBQaEaHrO0fS1ez8QV/B5WrVrFTk5OPH36dHFdTk4O9+vXj//44w92cHDQ2k362bNn2c3NjYcPH87Xr18vdB5OnjyZGzduLOkNkMrx48fZxMSEV65cyceOHeOdO3eyhYUF9+rVi+/cucN5eXl86NAhHj58OHfv3l3yoP/bb7+xIAhcsWJFPnz4sLg+KSmJv/nmGzYzMxP7AxWkafB/m86EYWFhzJyfBVRdN93c3CS9RpUkVeDfFn2d9yWkabRsi76OwH/jxg2uVq0a+/n58fXr13nnzp1sbW3Nu3bt4o0bN/KECRNYX1+f582bJ77myZMnHB4ezk5OTmKqWco72e3bt3OFChW4d+/e3KVLF/7oo4/Y39+f8/LyOCQkhJ2dnVlfX5/r16/PlSpV0kqw8/f350aNGvHQoUO5efPmrFAoeO7cuaxUKvnJkye8f/9+7tevH3/55Zc8f/58Mc0sZS/ipKQk1tfX5/Lly6tlGFTy8vJ46dKlHBAQwGvWrNFKp5379+9zy5YtOSgoiJnzMxF3797ln376SbzYPXz4kE1NTdnDw0OyJgbVxe3kyZMcGhrKX3/9NZ8/f55fvHjBycnJYs1/w4YNzJzf0e7777/na9euSXL818tx5MgR9vHx4c6dO3NAQIBaX5PJkyezgYFBoRqeNkybNo0dHR25cuXKPHPmTL5x4wY/e/aMAwICuH79+ty6dWueMWMGu7q6ctOmTTk3N5d79+7N3t7eWitTXFwcN23alIcNGyaO+jh69CgPHz6cK1WqxOfOndPKcadNm8ZeXl5q686dO8eVKlXicePGqa3XRn+kJ0+ecGBgIBsZGYnNoCq3b9/mCRMmsCAIajcFmnrbzoT169cX+32cPn2ag4KCtHKNKikI/Fpw7tw5dnJy4rFjx/KXX34pXuSZ81Psy5cv5w8++KBQr9StW7eyiYmJmFqTwuXLl7l69eocGhrKzPk1WzMzM7FX5atXrzgnJ4e3bdvGBw8e5Js3b0p2bJXdu3ezmZkZx8bGinfG33//PQuCwHPnzuWnT58W+Tqp7qIL3tFfu3aNz58/zzY2Nuzp6Sn+TXXnrm0vXrxgV1dXnjdvHmdlZfHkyZO5VatWbGNjw4IgiCn3jIwMyYPu9u3buWLFity7d2/u0KED16hRg8ePH8+vXr0S2/wbNmyo9b4mERERbG5uzr6+vrxo0SL+4IMP2NfXVy19PW3aNBYEQa2TnaaUSqXahf5///sf29jY8E8//cRz587l6tWr88CBA/ny5cuck5PDkZGR3L17d+7SpQsPGjRIDHafffYZf/3115yXl6e1zyg+Pp6bNWvGw4YN4/379/Pq1au5U6dOkteyVZRKJfv6+nKnTp2YOT8gqjrubdq0iS0tLfn27dvi56fp+y74PRTc16tXr3jGjBksCIJax1Pm/ErVsmXLtBJo37Yz4euZqLJW01dRBf5fTl/n/X+labT8chqBXxQXF8fOzs5csWJFnjt3rtrfHj16xD179uSvv/6amf+5Y8zKyuImTZqI7XlSOHLkCDs5OTFzfgel6tWr84gRI8S/l8T45J9//pkbNmzIGRkZaif8nDlz2MDAgIOCgrSSulRdUB4+fMhPnz4VbzBevnzJJ0+eZBsbG+7SpYu43fLly3nhwoVqr5WqDLdu3RIvnKNHj+ZmzZqxkZER9+zZk1etWsWPHj0SO29po6f0xYsXuUaNGmKNPjs7mwVBUOutf+3aNfbx8eHmzZvzkydPtBLUEhISuHbt2mqjN6pUqcIKhYI7d+6s1p9i9uzZWunbwJxfa5swYQJv3rxZXLd//35u2LAhDxgwoMhx869eveIJEyawhYWFpP0+3iQ+Pp7d3Nx44MCBfPz4cbWsiKYKnhvZ2dnMzLxz5042MjISO/WqztWIiAiuX7++2KlPUwWvASEhITxy5Eju3r07r1mzRjzGzJkzWRCEN/ba1zT4v2tnwv79+2t03NJCFfi3n/6bD/51X6Nl++m/EfgLunDhAtvZ2XGzZs0Kpc+HDh3KXbp0UVv37bffsiAIfOfOHcnK8Pvvv7OnpyefO3eObW1tefjw4eJd6tmzZ3ns2LFiz3IpFBUstm3bxgYGBmImQ1VzunjxIn/wwQdsbGzMK1asYGbpmjgKDsdr1aoVOzk5saOjo1qa9OTJk2xra8sNGjTgQYMGsb6+vqQ1qoId+Zo3b86rVq3ivLw8vnXrFh89epS3bt2qlmnw9vZmf39/yY5f0MmTJ9nV1ZWZ83v0V69enYcNGyb+XZVduHr1Kt+7d08rZWDOb8edPXs2M+c3J9jZ2fG4ceP4/Pnz/MEHH/CgQYMkr9WOHDlS7NeSl5fH8fHxbGxszEZGRoUmadq/fz87ODiwj4+P2g14YmIijxs3juvVq6eVprA3OXPmDLdv314rE8VERERwy5YtuU6dOjxr1iw+ePAgjx07lu3t7dXS6VOmTGEnJ6dCoy00NWnSJK5SpQoHBgby6NGjuU6dOty3b19++fIlZ2VlcUBAgNrkQdpQGjoT6gICv5b9+eef3KRJE/bx8RGDTmZmJrdo0aJQB6Hff/+dY2Nj3/lYqkBz4cIFsaf4pUuX2NramvX19dVq+sz58wd4enpKdkIXPCkK1lrz8vK4Q4cO7OrqyikpKeL6v//+mydNmsSzZ89mIyMjsT1TKnv27OHy5cvzvHnz+NChQ9yzZ0+2srLiffv2idvcuHGDBw4cyF988YXabHZS2b17N5uYmPCSJUveePG+d+8eT5kyhStXrqy1Gu62bdu4WbNm/OjRI7azs2M/Pz/x+4qKiuKvvvpKsjkT/k1GRgYnJCRwXl4e9+7dmwcPHszPnz9npVLJbm5uLAgC9+nTR7KmlwcPHvA333xTaH8///wzV6lShfv06VNoxMLBgwe5SpUq4hBGlVOnTkl6U/62tDGEMS4ujs3NzXnOnDk8duxYdnJy4n79+nFQUBB/8803bGBgwK6urtyqVSuuUKGC5Dc7J06c4Dp16ojDJvft26fW6ZY5P8MyduxYbtmypVayT7rqTFgaqAL/jjN/c+TF+xotO84g8BcpPj6eGzRowB9++CF369aNe/fuLXYUYs7/IUnVMzUiIoJtbGx49uzZ4jC9HTt2sCAIPG3aNI6Pj+fExEQeP348V6hQQbLaVcETc9myZdynTx8eO3Ys//bbb8ycf6K3bduWGzRowIcOHeLIyEju3Lkzd+3alTMyMtjW1lbSWepu3brFLVu25KVLlzJzfs/tWrVqcc2aNblcuXKFJt3QRmeltLQ0bt68udjH4/nz53z//n3esWMHHz9+nJnze8/7+PhwrVq1JO+09ddff4n7fPbsGdvb27MgCPzll1+qbTdhwgT++OOPJW9uUf0m0tPT+dWrV2oBLDs7m93d3XnNmjXiunHjxnFUVJRkQwdfDxY//vijmFlS/btq1ar8zTffiGPWVaKjo8XM2PtwoS/o+vXrPHfuXLUOxnv27GEPDw/u06cP7969m48fP86TJ0/mhQsX8tWrVzU+5uuf4c6dO9nR0ZGZmX/55Rc2NTUVm3+ysrL44MGD/PLlS87JyRG/R6mDvy46E5YWqsC/M+ZvPpR4X6NlZwwC/xup2jYdHBz4f//7n3hRkbKjyt69e9nExITXrFnDycnJan9bt24dW1pasqWlJTs4OHDjxo0lCzQFT8jAwEA2MzPjUaNGcc2aNbl169biMLH4+Hj+7LPPuHz58ly7dm1u2bKleHI7ODhIOgWn6uKWlZXFKSkpXLduXR42bBhnZ2dzx44duVq1apJODlSUrKwsbtmyJS9fvpwfPnzI06ZN49atW7OVlZXYuTM9PZ03bdrEt27dkuy4SqWS09LSuG7duty/f3/xe96yZQvb29vzgAED+PHjxxwbG8uTJ09mc3NzrXUa27VrFzdt2pRdXFzY399f7Dz64MEDtrGx4aFDh/KpU6d4ypQpbGNjI1k7MrN6x6ucnBzu2LEju7u7i/0cmPPPCxsbmyKD/+v7eB9kZGSws7MzW1paqs2Oxpwf/Nu3b8+ffvqp1kYOxMTEMHN+k4qnpyfv3LmTTU1Nxc7HzMwHDhzgkSNHqs1k+L51JtQ1BP4SFBMTozYdrJQ1iezsbO7ZsyfPnDmTmfNreElJSfz999+LM0ulpKTwmTNn+MKFC1rpTBcXF8fDhg3jY8eOMXN+Cvvzzz9nd3d3tTnHExMTOTU1Vfwcpk6dyrVr15Z0NAMzi0Fm3Lhx3L17d7Fz3/Dhw9nExISrVq1a7Gkfi+Phw4fcu3dvdnd3Z2NjY+7VqxevXLmSb968yb169RJr3pr+Dt7U6Wvr1q3cqFEjscd8Tk4Or1+/nmvVqsXm5uZcv359btasmdYu8gkJCWxhYcELFizgkSNH8scff8xt2rQRe+8fOnSIjY2NuVatWlytWjVJ08kFO98FBwdzbGws37t3j3v37s1t2rRRm7Bo/fr1XL16dR4yZIhOUvklLT4+nuvWrcstW7Ys1Ly2f/9+dnR05IEDB3J2drbGAfeXX34RhwKOGzeO27Rpw8+ePePMzExxJEvBrM/z58+5S5cuPGDAAMlq+KWhM2Fpowr8ETE3+HDiA42WiJgbCPz/RRtBnzk/bdW4cWOeNm0aP378mMeOHctt2rRhKysrtrS05EWLFkl6vNdt2bKFmzdvzo0bN1YbEpicnMyff/45t2zZUi3Nypx/IzRq1CiN5w5QfaZ///03JyYmqt3U5ObmspeXF0+YMEFc99VXX/Gvv/4q9oPQlFKpVOu9/9dff4lzADx48IB3797NGzduVEt1f/rppzxx4kS18r8LPz8/9vX1FftUvD6f/fbt29ne3p59fX3FlG1OTg4fOnSIr1y5ItlnoFLwvcTExKh1Vty7dy936tSJW7RoIQb/27dv84ULFyTtX5CYmMiCIPDq1at54sSJXLFiRfF4KSkp3KtXr0LBf9myZdyjR4/3Lq3/Jn/++Sc7Ojry8OHDCwX/Q4cOSZKBysvL459++okFQWB3d3c2NTVV60dz6tQptra25p49e/KOHTt427Zt3LFjR3ZwcCj0PA0plIbOhKWFKvDvPnuDf730QKNl91kE/rci5Y/53LlzYqexH374gQ0NDdnc3Jx79eolpjT9/PzE8bna8ueff/LHH3/MpqamhZ4odufOHR48eDDXq1dPLb1++fJlDgkJkaRNd/v27Vy9enWuUKECf/LJJ2p37sOHD+cqVarw6tWreejQoVy5cuUi07rFpappF5zutEaNGvzRRx+xgYEBDxo0SHyUp8qjR494ypQpbGFhofG0q1u3buUqVaqINfY//viDZ8yYUWiUxi+//MI2NjY8cOBArT7+WfU5HD9+XOwkNnToULVt9u3bx506deI2bdporXkhJyeHV65cyQYGBmxubi7eiKr61KiCf9u2bdXS/tq6KS+tCs4VcPHiRa0dp3379iwIgto0yKo5FU6fPs1Nmzbl2rVrs4uLC3t7e2tlNrzS0JmwNEHgL8Pu3LnDLVq04M6dO4s1ptOnT4tP/1OdOCNHjuShQ4dKNj78TRfGK1eucKdOnbhDhw6Fph5OSkriOXPmFDqZpTi5U1JSuEmTJrxu3Tret28f9+3bl1u0aCF2qnv69Cn36dOH7e3t2dXVVZLUtp+fHw8ZMkSsmZw4cYI/+OADXr58OV+6dIm3bdvG7dq1Yy8vLz569Cgz5891PmDAAK5Zs6Ykae3vv/+e7e3tmTn/UbtVq1blypUr87Rp0wrV2JYvX86mpqbcr18/tTHqUtu1axebmJhww4YN2dbWlitUqFDoRuTAgQPs5ubGnp6enJubq5WUruohN68/4Ej1faWkpPBnn33G9erV47179zKzeuZGLuLj49nFxYX79eunlWdzvHr1ioOCgvjbb79lIyMjHj9+vPg31XeRk5PD9+/f5/v374ufv5Tj9JlLR2fC0kQV+PecvclRl9I1WvacvYnAX9JWrVrF7dq1408//VRtqBxz/tjs6dOnS9p5q+AJ9eeff/LRo0c5JSVFbD+/cOECd+zYkTt27PjG5w5oGuxfv0Cnp6dzv379xIlIkpKSePjw4ezq6qo2UiAlJUWSNv3Xa9rM+Q8f6tixo9p2x48f51atWolt+bdu3eLQ0FDJpgKOiYnhevXqcfv27VmhUPCJEyd49erVbGNjw1OmTFFrcgkPD+fGjRtzq1attDImnDn/Ajpz5kxev3495+Xl8W+//cYdOnRgOzu7QhmWw4cPS9qno+Dv8urVq5yens53797lFStWsCAIHBwcXGi7+/fv84wZM967DnzFFRMTw23btpXkd1Hw8920aRPv3r1bPOd++uknNjIyKjRXhapPkIqUAVdXnQlLO1Xg3xt7k49eTtdo2RuLwK9VBWcQK2jDhg3cunVr/vTTT8XJV06dOsV9+/blevXqaaX3/tSpU7lu3bpcsWJFdnd358mTJ4vtxQkJCdypUyfu3LlzkQ9bkaocBw4c4M8++4w///xz7tChg9o2quDfsmVLtWFLUihY0961axcvXbpUfMJiwRoDc/60sMbGxmKnMakvKKNGjWJBEMTJeZjzO7Opgr+qTXX69OkcGhoqPlJUarGxsVyhQgV2c3MThyqq1nt6erKdnZ1Wnn3ArB5sZsyYwV26dOFt27axUqnk7OxsXrx4MQuCwMuXLxe3mzBhgtpDkOQe/KWYK6Dgb3vSpElsZWXFGzduFB8hnJuby5s2bWIjIyMeM2YM37hxg728vLhHjx6SnReloTNhWaAK/Ptib/Kxyw81WvYh8Gvf6dOnedSoUYU+5A0bNrCTkxP37dtX7Kkqda1KZf78+WxtbS3OdjVo0CC2srLiYcOGiU0OCQkJ3KxZs0IP95DK8ePHWaFQcP/+/blZs2ZsYGCg9jQ15vzgP2DAAO7YsaOkw8RUNe0OHTqwIAi8a9cuDg8PZ319fbWgx5x/A1a/fn212oRUnj17xh06dOBhw4ZxgwYNuF+/fuLfli9fzg0bNuTatWtzixYtuHz58lprU2fO78TZo0cPFgRBbGpSiY2NZS8vLzYzM9PKsyBUZsyYwZUrV+YDBw6odVp8/vw5L1myhAVBYB8fH27dujXXq1fvveupXVoEBQWxtbV1oenAVW3327Zt43LlyrG9vb3afCZSDNkrTZ0JSzME/jJm3rx57ODgwF9//XWhYVzjx49nY2NjtTZ/qV25coVbtWolToF6+PBhLl++PH/66adcr149/vLLL8WL7t9//y1ZR6mC+7ly5QpHRESI063euXOHZ82axQ0aNOCAgAC11yUnJ2tlClpVTdvNzU1cN2DAAK5cuTJHRUWJNesJEyawg4ODpDceBamaN9avX8/16tVTC/6HDx/m4OBgnj59utbmlr948aL4XpOTk7lbt25Fdlw8ffo0f/bZZ5I/eEglMTGRGzVqJA5dVSnYXhseHs6dO3dmPz+/Mvs41dJIdW6qHrPcp08fcY6AGzdu8M6dO9nT05N9fHz49OnTzJx/zh47dkx8rZQ3YaWhM2Fppwr8B+Ju8m9XHmq0HIhD4Ne63Nxc/u6779jNzY1Hjx6tlrpVPdq3b9++hSbwkcrLly95+/bt/PDhQz558iRbW1vz6tWrmZm5V69eXKlSJf7000/VhtRpEvy//fZbtb4Lt2/fZgsLi0LtdCkpKRwQEMD29vaFplqV2us1bdWDO169esWDBg1iIyMjdnBwYHd3d6095vh1WVlZvGHDBra3ty+xB4mkp6eznp4e+/j4iCd9SkoKd+nShatUqVIo+GtjdkSVCxcusJWVFUdHRxf6W05OjtjOXHC4I2r80lLd1PXq1Yt79+7NS5YsYU9PT+7SpQv37NmTu3Tpwp6enoUChKYBtzR0JixrVIH/YNwtPnHlkUbLwbhbZSLwK6iMYGYiInr8+DE9e/aMHj9+TAYGBjR+/Hj65JNPKD4+nqZOnUoZGRlERHThwgXy8vKi1atXU7Vq1TQ+fl5eXqF1+vr69Mknn1ClSpVo69at1K1bNxoyZAgREdWtW5fq1q1LtWrVoooVK4qvUSje7SN/9OgR/fHHH/T48WNxnYmJCc2YMYNMTU0pJiZGXF+1alUaMWIEDRgwgEJCQmjhwoXvdMy3YWJiQnv37qW1a9fS+PHjKT4+ngYNGkR6enr0008/0ebNm2n48OHk7e1NMTEx1LRpU62VRaV8+fLUt29fmjRpEiUkJFD37t21fszKlSvTrl27aMeOHTRp0iTKzMykqlWr0rp168jZ2Zk+/vhjunjxori9kZGR1sry7NkzevbsGb148YKIiHJzc8W/RUdH0y+//EIvXrwgExMTIso/t/T19bVWHrnZs2cP9e3bl3JycmjIkCGUnp5OwcHB1Lp1a5o9ezZFRERQmzZtyNjYmMzMzNReq6en987HVSqV4vVl8+bNtH//fvLz86NZs2bR2rVracWKFTR+/HgiIvH7PnXqFFWpUoWqVKlCgiDgtyATZeIbZmYSBIH2799Py5Yto7t371L9+vVp0KBB1K1bN5o0aRLp6+vT7t27qW7duuTk5ETHjx+n+Ph4qlChgsbHJvrnhNy9ezdlZGSQlZUVeXp6krGxMRERpaenU1ZWFimVSiIiunHjBo0cOZIGDRpEgiConZTvolKlSrRv3z4yMDCgY8eOUe3atcnW1pY+//xz0tfXpylTppCFhQUtWrSIiIg+/PBDGjJkCBkaGlLv3r01+Qj+U7ly5YiIqG/fviQIAi1cuJAGDBhAW7Zs0fqx3+SDDz6gvn370osXL2jjxo109+5dqlq1qmT7V/0mC/67a9eutGPHDvFGY+HChWLw79OnD/Xq1YsuXrxIBgYGkpThTb8pV1dX8vT0pP79+9OZM2eoevXqRET04sULCgwMpPr165Ovr6+4fcH3AZpTKpWUkpJCV65coa5du5KrqysxM1laWorbnDhxQu3fmmJm8bcwefJk+t///kcLFy6kZ8+eUfny5alfv34kCAINGzaMcnNzyd/fn8aMGUMGBgbUtm1b8Tcg59+CIOQvmu6jLBBYFdlKuT179lD//v1pxowZVK1aNTp69Cjt2rWL1q5dS71796aXL1/SmTNnKDIykl6+fEm+vr5kb2+v0TF79+5NtWvXpgULFhAR0ZQpUygkJIRq1qxJf/31F02YMIGmTp1KFStWpMWLF1NYWBiZmprS8+fPKTMzkxISEkhPT0+joB8YGEhZWVk0f/58EgSBnj9/Th4eHnTp0iW6cOECVatWjR49ekRhYWE0c+ZMGjJkiBj8ifIzFZrUIoorOzubtm3bRkFBQVSzZk3as2dPiR27KM+ePaOXL1+Subm55Ps+dOgQnT17lqZNm0YKhUK8GYiMjKTu3bvTiBEj6Ntvv6VKlSrRvXv36NWrV2RrayvJsQv+prZv3063b98mQ0ND8vLyolq1atHVq1fpyy+/pNjYWJo+fTrl5OTQ77//TqmpqXTu3DnU6iRS8HsoeDPYvXt3evbsGR08eFC80cvIyKBTp07RihUr6Pbt2xQfH08GBgaFbiI1sXTpUvr+++9p79695OzsLK5/+fIlGRgY0C+//EJffPEFVa9enUxMTOjMmTOSl6GsyczMJHNzczoUf4s+KG/23y/4F9lPM8mzmR1lZGQUyuaUKjpoXii2a9eusbOzs9iOnZaWxtWqVeP69etz+fLledu2bWrbS9WJbunSpayvr8/z58/nK1eusJubG8fFxYlTzxoaGvKIESM4NzeXc3NzOSgoiL/88kseNWqU2E6maZvdsmXLWBAEDgwMFNvgrl+/zu3bt+fq1auL/RfS09M5JCSErayseOTIkZq9cQ09ffqUQ0ND2cXFpdB8Cu+T0NBQFgSBv/vuO7VOXczMixcvZoVCwSNGjNBqe9+kSZPY0tKSu3btytWrV+fOnTtzREQEM+fPpjhhwgR2cXHhtm3b8pAhQ8TOW3Jrx9W21z/PHTt2sJubm9rskAkJCdy1a1fu06ePuL1Uk/OUls6EZZGqjf9Q/C0+efWRRsuh+LLRxl9qA78qyOXk5PDDhw/5q6++4vT0dE5OTua6devy8OHD+cqVK9y6dWsuX758kc+K1oTqpFizZg0rFAoeOnQof/HFF2onSWRkJBsaGvLw4cOLfGa6pieU6jNYu3YtKxQKtY56N27c4NatW6sF/4cPH/LixYu5Vq1anJaWptPhONnZ2VobJ68LSqVSvIl7+PChOEnT5s2bxe+m4A3n+vXrxSe/aWuSoGXLlrGtra04VOzHH39kQRC4VatWak93fPz4sdpvQe4Xeim8PjlP5cqVOSQkRBwyl5WVxQ4ODjx8+HC11yUlJWkl4OqqM+H7QBX4j8Qn8alrjzVajsQnIfC/K9VF6siRIzxu3Di+ceOGOFxv3Lhx/Nlnn4k9k1Vzz1evXp0zMjIkCXYFT4Znz55xWFgY6+npccOGDfnx48dqZYyMjGQTExP29vaW9MsuOCOfUqkUA8zcuXPFbd4U/B89eiRZOeRu//79arW2nTt3cosWLbhOnTocEBDAly5d4vDwcNbT0+M5c+aIIzimT5/OP/74Y6EHBUnl6dOnPGHCBPFhTzt27OAKFSrw7Nmz2dXVlRs1alTkzbBcxmZrU8Ggv3PnTo6OjmZ/f3/u0KEDW1hY8LRp0zghIYGPHDnCTZs2LTSG//V9aGr37t3ctGlTfvHiBe/du5fbtWvHtra2PG/ePHE+/sDAQO7Ro4dkx3yfqAL/r/FJHH3tsUbLrwj8mtmxYwebmJjwnDlzxBMnNzeX27Vrx2PHjhW3Gz16NK9du1ayseEFT8jFixfzyJEj+dKlS7xlyxZWKBT87bffituoLqK7du3itm3bSnoyF7z5+eabb/jSpUu8cePGIoN/+/btuXz58u91Wl0XUlNTuWbNmuzr68vXr1/nS5cucYUKFXju3Lk8duxYbtq0KX/66ad89uxZjoiIYIVCwU2aNOGWLVtKOjX061TjsC9cuMBpaWmcmJjIderU4aVLlzIz88GDB7l8+fLs6OjIBw8e1EoZ5KrgjdOUKVPYxsZGfP5BcnIyb9myhd3d3blZs2Zsb2/Ptra24sOPtPXAo4iICLa0tBSfO3H//n1xhkCVLl268ODBg7Vy/LIOgb+UuHLlCtesWVNtbLrKxIkTuVatWhwaGspfffUVf/jhh1qZ/nTSpElsYWHBW7ZsEfevSvvPmzevUPBXkfLkVt38zJ07V7z5UZWhYPC/du0ae3l5iY+aBenExcWxs7MzjxkzhufOnav2ue/du5fbt2/PPXr04GvXrnFCQgJPmTKFJ06cyImJiZKV4U2/KVWqeP369ezi4iJmG7Zt28Y9evTgyZMny+bpeiVtzpw5bGFhwWfOnBGzgCr379/n6Oho/uyzz7hixYpco0YNyZp7Cn6fBa893bp1448//lityfHJkyd84MAB9vLyYgcHB8lmBXzfiIH/XBJHX3+s0fLrOQT+d3bkyBGuW7eu2lPVVD/W+Ph4HjlyJNesWZOdnJy0MiHMr7/+yjVr1uSTJ08W+tvq1atZT0+P58+fr9WL6r/d/KxevZoVCgXPnz9fXId2W+2Ji4tjFxcXrlGjBk+ePFntb3v37uW2bdtyr169ikzpaqrgRTo0NJS//vpr9vX15d9++02cqTA0NJQdHBz42LFjnJWVxd27d+fAwEDxdQj+0nr48CF7eHjw5s2bmTm/s9xvv/3GPj4+vG7dOrWpwQ8fPsytW7fmX375hZmlC7q66kz4PlIF/qhzt/n09ScaLVHnbpeJwF8qJ/B5+vQpPX/+XPy3amw8Uf7wLB8fH0pISKBff/1VKxPC3L59m8qVK0cNGzYU1/H/j3ocPnw4/fTTTzRjxgzatGmT5McuWAYDAwPy8vIS16k+h4JlUA3dw/As7WnWrBmtXbuWFAoFnTx5Um0inq5du9KkSZPo5s2bFBwcTM+fPxd/K5pSKpXiEKvJkyfT9OnTKSkpiW7evEkff/wxff/99/T48WPy9PQkfX19Gjp0KDVo0IBu3bolTtTCBcZ3gzQEQaDExES6dOkSnThxgsaPH08TJ06kK1eu0OTJk2nnzp3ith07dqRy5cqJw1rfdchcwWvg5s2bydramkJDQykhIYGIiDp16kRPnz6l0NBQcTsHBwcKCQmhsLAw0tfXp1evXuE68W+Ef8byv+tCZWREZKm8IjRp0oTS09NpzZo1RJQ/253qhNm+fTvt37+fTExMNJ6c53WqC/bz58/VZupTrWdm2r59OzVr1owiIyNp4MCBkh6/oKJuflSfwfHjx8nJyYnCw8Opa9euWisD/KNx48a0a9cuys7OpmXLlqkFfy8vL1q4cCHNnz+fTExMJBsPrQrYd+/epcePH9OhQ4do165ddOzYMQoODqZly5bRmjVrqFatWrRlyxZasGABzZ07l+Li4sjAwIBevXol27HZ2lSxYkWaM2cOhYaGUrdu3ahGjRo0f/58On36NHl5edG5c+eI6J9gXbNmTXr+/LnaDIrFUXCugIiICKpduzYNHjyYduzYQR06dKDp06fTrVu3aOnSpXT27FmKjY0VX1u9enVSKBSkVCoR9OEfukw3/Jv169ezgYEBT5w4kRMSEjgxMZEnTZrEFSpUKDTvudQSExNZT0+v0ENuVGnUgo801Vbq7MaNG2xiYsLTpk0r9Ldx48bxzJkzMRRHB+Lj47lZs2Y8bNgwvnjxotaPt2nTJi5XrhzXq1ePL1++rJYqXrx4MZuYmPD169cLvQ6/De1LSkpS61eTl5fHH3/8sdqTMePi4tjOzk4tBV8cpbEz4ftGleo/ev42x/z9RKPl6PmykeovtYE/Ly+Pt23bxhUrVuRq1apx7dq1uV69eiXykBfm/HZ0AwMD/vrrr/nw4cN8/Phx7tSpEzdu3LjE2sl0efMDbxYfH88uLi7cr18/rX8PR48e5S5durCJiYnYa1s1RDA9PZ1tbGx4586dWi0D/LusrCz+/fffuWvXrtyoUaNC14cHDx5ofAxddSaUAzHw/3mbY2480Wg5+mfZCPylNvejUCioT58+1LJlS0pKSiJBEKhmzZpkZWVVIsf38/MjKysr+vrrr2nHjh1UoUIFsrGxodjYWNLX1y+RqXC/+OILMjU1pREjRtDWrVvJ2NiY9PT06OjRoxpPRwzvrmnTprRixQqaOHGipFMBFzW1c9u2bcnQ0JAePXpE3bt3p7Nnz1KVKlWIKH/ufUEQkMLVIWam2NhYWrJkCb18+ZLi4uLE64MgCKRQKMjCwkKjYzx69IhOnDhBwcHB5OLiQikpKXThwgVav349tWnThjp27Ehubm60fft2OnLkCM2dO5f++OMP6t27t6yn4oU3KzNz9etKeno6ZWRkkFKppI8++ogUCkWJd5K5e/euTm5+4N+9ePFCfEiTpgpeoHfv3k3Pnz8nS0tL6tChAzEznTlzhsaOHUspKSk0Z84cMjY2pi1bttCdO3coLi6uRJ/HAOpycnIoMTGRmjRpopXrw+PHj8nBwYF8fX2pU6dOFBoaSjdv3iRBEOj69es0Y8YMGjdunLh9586dydLSkn766SfJyvA+U83Vf+zPZCpvqtn8+k+zMql9E9tSP1d/qezcV5pYWFjQRx99RHXq1NFZJ5mqVauSu7s7ubm5IeiXIlIFfaJ/entPmzaNBg4cSPPnz6eOHTvSnDlzSBAEcnNzox9++IFq1qxJw4YNo4MHD1Lnzp3p1KlTpKenV+Rjo6FkGBkZUdOmTbV2fSjpzoRypWmPfime7ldSkCMsJgyNAimpavrMTKmpqRQdHU3Hjx+nqlWrUlRUFA0ZMoSysrJo0aJF5ObmRoGBgTR//nwxvVyuXDl6/vw5mZiY6PqtAGnv+jB06FDq2LEj5eTkUJ06dYgoP9DfvXuX3NzcxGPHx8dTZGQk7dq1iwwNDbVSFij7EMUAdKTgEM0HDx7Qw4cPqWnTptSwYUOqWrUqDRo0iDZv3kzBwcE0adIkIiJq1aoVzZw5k6pUqUIdO3ake/fuIejLRPXq1alOnTr09OlTOnnyJPXo0YPu379Ps2fPFrdp1qwZnT17lpo0aaK7gpZRgkRLWYDAD6Ajqtrh1KlTqVOnTtS5c2fat28f3bx5U9zG29ubfv75Z1q+fDmNGDGCiIhatGhB33//PTEzffrpp6RUKiWbNAhKN1VnwoULFxbqTKhK9WvamVC2ZBT5EfgBSljB9viff/6Ztm7dSkOGDKGRI0fSzZs3aenSpZSamipu07dvX1q5ciVdvnxZfK2rqyutW7eOwsLC1Ca4gvebIAjk7u5Oc+bMoQMHDogTNenp6aEZEt4aevUDlJAFCxZQ3759qVatWkRE9Ntvv9HevXupUaNGNHjwYCIiioyMpE8++YT8/Pxo9uzZZG1tXWg/mHoVVIoaAgrFo+rVfyLhjiS9+ts0qoZe/QBAdPXqVTp//jzVqFGDiIiSk5PJy8uLgoKCKDk5Wdyuc+fOtH//flq/fj3NnTuXUlJSCu0LQR9UEPSlI6de/fjVAJSAunXr0tatW0lPT4/27dtH5cqVo99++41sbW3p999/p/Pnz4vbqoL/ypUraevWrborNICMyKiJH4EfoKQIgkCpqak0evRoGj9+PNWqVYvCwsLo0qVL9MMPP4hPWiPKf9padHS02sQsAABSQOAHKEHW1tYUERFBf/31F02cOJHs7e1p69atFBUVRUFBQfTXX3+J27q6uoqPUwUALZNRlR+BH6CENWvWjNauXUvx8fE0YcIEatCgAW3dupWOHz9OM2bMoBs3bqhtjzZ9AO0TJPqvLEDgB9CBpk2b0oYNGyg+Pp4mTpxIDRs2pA0bNpBCoSA7OztdFw8A3mMI/AA6ogr+58+fp6FDh1Lz5s1p586d4pzvAFBy0KsfAEpE06ZNKTQ0lExNTalcuXLiegzTAihZMmrix0N6AHTNxcWFmjdvToIgYEIWANA6BH6AUkD1hD4EfQAdkaLKXkaq/Aj8AKUE5tsH0B0peuWjVz8AAACUOgj8AAAge7rq1R8SEkJ2dnZkbGxMrq6uFBMT88ZtL168SJ999hnZ2dmRIAgUHBz8Tu8VgR8AAGRPF736w8PDyd/fnwICAig+Pp6aNGlCnp6edP/+/SK3f/bsGdWqVYsWLFhQ5JM73xYCPwAAgA4if1BQEPn5+ZGvry81aNCAVq1aReXKlaMNGzYUuX3z5s1p0aJF1K9fPzIyMir+e/x/CPwAAAASyszMVFtycnIKbZObm0txcXHk4eEhrlMoFOTh4UHR0dFaLR8CPwAAyJ6Uc/Xb2tqSubm5uAQGBhY6Xnp6OuXl5ZGVlZXaeisrK0pNTdXqe8VwPgAAkD0pptxVvT45OZnMzMzE9Zqk5bUBgR8AAEBCZmZmaoG/KBYWFqSnp0dpaWlq69PS0jTquPc2kOoHAADZK+m+fYaGhuTk5ERRUVHiOqVSSVFRUeTu7q7x+/k3qPEDAADoYMpef39/Gjx4MDk7O5OLiwsFBwdTdnY2+fr6EhGRj48P2djYiH0EcnNzKTExUfz/lJQUOn/+PJUvX55q16791sdF4AcAANABb29vevDgAc2aNYtSU1PJ0dGRIiMjxQ5/t2/fVnt+x927d6lp06bivxcvXkyLFy+mtm3b0vHjx9/6uAIzs2TvAgAAoAzJzMwkc3Nzir+WSuVN/71d/r88zcqkZnWsKSMj4z/b+HUJNX4AAAAJevWXkWf0oHMfAACAnKDGDwAAsqeDvn06g8APAAAgo8iPwA8AALJXcMpdTfZRFqCNHwAAQEZQ4wcAANmTcq7+0g6BHwAAZE9GTfxI9QMAAMgJavwAAAAyqvIj8AMAgOyhVz8AAAC8l1DjBwAA2RNIgl79kpRE+xD4AQBA9mTUxI9UPwAAgJygxg8AALKHCXwAAABkRT7JfgR+AACQPTnV+NHGDwAAICOo8QMAgOzJJ9GPwA8AAIBUPwAAALyfUOMHAADZk9Nc/Qj8AAAAMmrkR6ofAABARlDjBwAA2ZNRhR+BHwAAAL36AQAA4L2EGj8AAMgeevUDAADIiYwa+RH4AQBA9mQU99HGDwAAICeo8QMAgOzJqVc/Aj8AAIAEnfvKSrIfqX4AAAAZQY0fAABkT06pftT4AQAAZASBHwAAQEaQ6gcAANmTU6ofgR8AAGRPTlP2ItUPAAAgI6jxAwCA7CHVDwAAICNymqsfgR8AAEBGkR9t/AAAADKCGj8AAMienHr1I/ADAIDsyalzH1L9AAAAMoIaPwAAyJ6M+vahxg8AACBGfk2XYgoJCSE7OzsyNjYmV1dXiomJ+dftf/nlF7K3tydjY2Nq1KgRHThwoNjHROAHAADQgfDwcPL396eAgACKj4+nJk2akKenJ92/f7/I7U+dOkX9+/enoUOH0rlz56hnz57Us2dP+uuvv4p1XIGZWYo3AAAAUNZkZmaSubk5paZnkJmZmcb7srYwp4yMt9uXq6srNW/enFasWEFEREqlkmxtbemrr76iKVOmFNre29ubsrOzad++feI6Nzc3cnR0pFWrVr11OVHjBwAA2VP16td0eVu5ubkUFxdHHh4e4jqFQkEeHh4UHR1d5Guio6PVtici8vT0fOP2b4LOfQAAIHuZmZmS7eP1fRkZGZGRkZHauvT0dMrLyyMrKyu19VZWVnT58uUi95+amlrk9qmpqcUqJwI/AADIlqGhIVlbW1OdmraS7K98+fJka6u+r4CAAJo9e7Yk+5cCAj8AAMiWsbEx3bx5k3JzcyXZHzOT8FrO//XaPhGRhYUF6enpUVpamtr6tLQ0sra2LnLf1tbWxdr+TRD4AQBA1oyNjcnY2LhEj2loaEhOTk4UFRVFPXv2JKL8zn1RUVE0ZsyYIl/j7u5OUVFRNG7cOHHdkSNHyN3dvVjHRuAHAADQAX9/fxo8eDA5OzuTi4sLBQcHU3Z2Nvn6+hIRkY+PD9nY2FBgYCAREY0dO5batm1LS5YsoU8++YTCwsIoNjaW1qxZU6zjIvADAADogLe3Nz148IBmzZpFqamp5OjoSJGRkWIHvtu3b5NC8c/guxYtWtCWLVtoxowZNG3aNKpTpw7t2rWLHBwcinVcjOMHAACQEYzjBwAAkBEEfgAAABlB4AcAAJARBH4AAAAZQeAHAACQEQR+AAAAGUHgBwAAkBEEfgAAABlB4AcAAJARBH4AAAAZQeAHAACQEQR+AAAAGfk/d0NxK6+eoMUAAAAASUVORK5CYII=",
      "text/plain": [
       "<Figure size 640x480 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def show_tfidf(texts):\n",
    "    '''\n",
    "    Function to show the tf-idf scores of words in a corpus and visualize them.\n",
    "    '''\n",
    "    # We'll instantiate a term frequency - inverse document frequency vectorizer class\n",
    "    vectorizer = TfidfVectorizer()\n",
    "\n",
    "    # And apply the fit_transform method to our texts\n",
    "    X = vectorizer.fit_transform(texts)\n",
    "\n",
    "    # This gives us a matrix of shape n_texts, vocab_size\n",
    "    X.todense()\n",
    "\n",
    "    # The features each column represents can be accessed here\n",
    "    features = vectorizer.get_feature_names_out()\n",
    "    visualize_doc_vec(X, features)\n",
    "\n",
    "show_tfidf(texts)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Learning from features\n",
    "\n",
    "Support Vector machines work by finding a hyperplane in a multi-dimensional space that separates samples of different classes. The matrix we have visualised above forms our multidimensional space."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 138,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<style>#sk-container-id-11 {color: black;background-color: white;}#sk-container-id-11 pre{padding: 0;}#sk-container-id-11 div.sk-toggleable {background-color: white;}#sk-container-id-11 label.sk-toggleable__label {cursor: pointer;display: block;width: 100%;margin-bottom: 0;padding: 0.3em;box-sizing: border-box;text-align: center;}#sk-container-id-11 label.sk-toggleable__label-arrow:before {content: \"▸\";float: left;margin-right: 0.25em;color: #696969;}#sk-container-id-11 label.sk-toggleable__label-arrow:hover:before {color: black;}#sk-container-id-11 div.sk-estimator:hover label.sk-toggleable__label-arrow:before {color: black;}#sk-container-id-11 div.sk-toggleable__content {max-height: 0;max-width: 0;overflow: hidden;text-align: left;background-color: #f0f8ff;}#sk-container-id-11 div.sk-toggleable__content pre {margin: 0.2em;color: black;border-radius: 0.25em;background-color: #f0f8ff;}#sk-container-id-11 input.sk-toggleable__control:checked~div.sk-toggleable__content {max-height: 200px;max-width: 100%;overflow: auto;}#sk-container-id-11 input.sk-toggleable__control:checked~label.sk-toggleable__label-arrow:before {content: \"▾\";}#sk-container-id-11 div.sk-estimator input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-11 div.sk-label input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-11 input.sk-hidden--visually {border: 0;clip: rect(1px 1px 1px 1px);clip: rect(1px, 1px, 1px, 1px);height: 1px;margin: -1px;overflow: hidden;padding: 0;position: absolute;width: 1px;}#sk-container-id-11 div.sk-estimator {font-family: monospace;background-color: #f0f8ff;border: 1px dotted black;border-radius: 0.25em;box-sizing: border-box;margin-bottom: 0.5em;}#sk-container-id-11 div.sk-estimator:hover {background-color: #d4ebff;}#sk-container-id-11 div.sk-parallel-item::after {content: \"\";width: 100%;border-bottom: 1px solid gray;flex-grow: 1;}#sk-container-id-11 div.sk-label:hover label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-11 div.sk-serial::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: 0;}#sk-container-id-11 div.sk-serial {display: flex;flex-direction: column;align-items: center;background-color: white;padding-right: 0.2em;padding-left: 0.2em;position: relative;}#sk-container-id-11 div.sk-item {position: relative;z-index: 1;}#sk-container-id-11 div.sk-parallel {display: flex;align-items: stretch;justify-content: center;background-color: white;position: relative;}#sk-container-id-11 div.sk-item::before, #sk-container-id-11 div.sk-parallel-item::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: -1;}#sk-container-id-11 div.sk-parallel-item {display: flex;flex-direction: column;z-index: 1;position: relative;background-color: white;}#sk-container-id-11 div.sk-parallel-item:first-child::after {align-self: flex-end;width: 50%;}#sk-container-id-11 div.sk-parallel-item:last-child::after {align-self: flex-start;width: 50%;}#sk-container-id-11 div.sk-parallel-item:only-child::after {width: 0;}#sk-container-id-11 div.sk-dashed-wrapped {border: 1px dashed gray;margin: 0 0.4em 0.5em 0.4em;box-sizing: border-box;padding-bottom: 0.4em;background-color: white;}#sk-container-id-11 div.sk-label label {font-family: monospace;font-weight: bold;display: inline-block;line-height: 1.2em;}#sk-container-id-11 div.sk-label-container {text-align: center;}#sk-container-id-11 div.sk-container {/* jupyter's `normalize.less` sets `[hidden] { display: none; }` but bootstrap.min.css set `[hidden] { display: none !important; }` so we also need the `!important` here to be able to override the default hidden behavior on the sphinx rendered scikit-learn.org. See: https://github.com/scikit-learn/scikit-learn/issues/21755 */display: inline-block !important;position: relative;}#sk-container-id-11 div.sk-text-repr-fallback {display: none;}</style><div id=\"sk-container-id-11\" class=\"sk-top-container\"><div class=\"sk-text-repr-fallback\"><pre>Pipeline(steps=[(&#x27;vect&#x27;, TfidfVectorizer()),\n",
       "                (&#x27;clf&#x27;, SVC(class_weight=&#x27;balanced&#x27;, probability=True))])</pre><b>In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook. <br />On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.</b></div><div class=\"sk-container\" hidden><div class=\"sk-item sk-dashed-wrapped\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-39\" type=\"checkbox\" ><label for=\"sk-estimator-id-39\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">Pipeline</label><div class=\"sk-toggleable__content\"><pre>Pipeline(steps=[(&#x27;vect&#x27;, TfidfVectorizer()),\n",
       "                (&#x27;clf&#x27;, SVC(class_weight=&#x27;balanced&#x27;, probability=True))])</pre></div></div></div><div class=\"sk-serial\"><div class=\"sk-item\"><div class=\"sk-estimator sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-40\" type=\"checkbox\" ><label for=\"sk-estimator-id-40\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">TfidfVectorizer</label><div class=\"sk-toggleable__content\"><pre>TfidfVectorizer()</pre></div></div></div><div class=\"sk-item\"><div class=\"sk-estimator sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-41\" type=\"checkbox\" ><label for=\"sk-estimator-id-41\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">SVC</label><div class=\"sk-toggleable__content\"><pre>SVC(class_weight=&#x27;balanced&#x27;, probability=True)</pre></div></div></div></div></div></div></div>"
      ],
      "text/plain": [
       "Pipeline(steps=[('vect', TfidfVectorizer()),\n",
       "                ('clf', SVC(class_weight='balanced', probability=True))])"
      ]
     },
     "execution_count": 138,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.pipeline import Pipeline\n",
    "from sklearn.svm import SVC\n",
    "\n",
    "def build_svm(toy_texts, y):\n",
    "    '''\n",
    "    Function to create a pipeline with a tf-idf vectorizer and a support vector machine classifier.\n",
    "    '''\n",
    "    svm_classifier = Pipeline(\n",
    "        steps=[\n",
    "            (\"vect\", TfidfVectorizer()),\n",
    "            (\"clf\", SVC(probability=True, class_weight=\"balanced\")),\n",
    "        ]\n",
    "    )\n",
    "    svm_classifier.fit(toy_texts, y)\n",
    "    return svm_classifier\n",
    "    \n",
    "\n",
    "# We'll slightly expand our list of texts\n",
    "toy_texts, y = zip(\n",
    "    *[\n",
    "        (\"Climate change is impacting human systems\", 1),\n",
    "        (\"Climate change is caused by fossil fuels\", 0),\n",
    "        (\"Agricultural yields are affected by climate change\", 1),\n",
    "        (\"System change not climate change\", 0),\n",
    "        (\"higher temperatures are impacting human health\", 1),\n",
    "        (\"Forest fires are becoming more frequent due to climate change\", 1),\n",
    "        (\"Machine learning can read texts\", 0),\n",
    "        (\"AI can help solve climate change!\", 0),\n",
    "        (\"We need to save gas this winter\", 0),\n",
    "        (\"More frequent droughts are impacting crop yields\", 1),\n",
    "        (\"Many communities are affected by rising sea levels\", 1),\n",
    "        (\"Global emissions continue to rise\", 0),\n",
    "        (\"Ecosystems are increasingly impacted by rising temperatures\", 1),\n",
    "        (\"Emissions from fossil fuels need to decline\", 0),\n",
    "        (\"Anthropogenic climate change is impacting vulnerable communities\", 1),\n",
    "    ]\n",
    ")\n",
    "\n",
    "# Now lets build a simple SVM classifier on this toy data\n",
    "\n",
    "svm_classifier = build_svm(toy_texts, y)\n",
    "svm_classifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 139,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "predictions: [1 0 1]\n",
      "probabilities:\n",
      " [[0.47595172 0.52404828]\n",
      " [0.45574361 0.54425639]\n",
      " [0.46922348 0.53077652]]\n"
     ]
    }
   ],
   "source": [
    "# We can now use this classifier to make predictions about new texts\n",
    "\n",
    "# Here are a couple of easy examples (easy because they are similar to our examples)\n",
    "new_texts = [\n",
    "    \"climate change is impacting terrestrial ecosystems\",\n",
    "    \"Machine Learning will solve climate change\",\n",
    "    \"Coca Cola is a company that produces sugary drinks\",\n",
    "]\n",
    "# We can make predictions about which class the examples belong to\n",
    "print('predictions:', svm_classifier.predict(new_texts))\n",
    "# We can also get these as probabilities (which are very unstable with small datasets, and not guaranteed to match the predictions)\n",
    "print('probabilities:\\n', svm_classifier.predict_proba(new_texts))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 140,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([0.52404828, 0.54425639, 0.53077652])"
      ]
     },
     "execution_count": 140,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Each row contains predicted probabilites for each class, in this case 0, and 1\n",
    "\n",
    "# We can get predicted probabilities of being in the INCLUDE class like this\n",
    "# One for each of the three examples\n",
    "svm_classifier.predict_proba(new_texts)[:, 1]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 141,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train size: 2000\n",
      "Test size: 500\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "           0       0.84      0.78      0.81       235\n",
      "           1       0.82      0.87      0.84       265\n",
      "\n",
      "    accuracy                           0.83       500\n",
      "   macro avg       0.83      0.83      0.83       500\n",
      "weighted avg       0.83      0.83      0.83       500\n",
      "\n"
     ]
    }
   ],
   "source": [
    "from sklearn.metrics import classification_report\n",
    "\n",
    "def train_eval_svm(df):\n",
    "    # load the train/test splits\n",
    "    x_train, y_train, x_test, y_test = train_test_split_df(df)\n",
    "    # build the classifier\n",
    "    svm_classifier = build_svm(x_train, y_train)\n",
    "    # get the predictions on the test set\n",
    "    test_pred = svm_classifier.predict(x_test)\n",
    "    # Get the classification evaluation metrics\n",
    "    # This will be covered in more detail later.\n",
    "    print(classification_report(y_test, test_pred))\n",
    "\n",
    "# So far this was an illustrative example using toy data. Let's now apply this to our real data.\n",
    "# takes around 20 seconds to run\n",
    "train_eval_svm(df)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Multilabel\n",
    "\n",
    "Multilabel problems are those where any sample can belong to any of the classes (as opposed to multiclass problems where each sample must belong to **exactly 1** class. In our dataset, most documents which were marked as relevant are assigned to one class, with some assigned to two, and very few assigned to three."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 142,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# This plot shows that certain documents have multiple labels for different impacts.\n",
    "plt.hist(df[df[\"INCLUDE\"] == 1][impacts].sum(axis=1));"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To prepare our target for the model, we simply create a matrix with a row for each document and a column for each class"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 143,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0., 0., 1., 0., 0.],\n",
       "       [0., 0., 0., 1., 0.],\n",
       "       [0., 0., 0., 1., 0.],\n",
       "       ...,\n",
       "       [1., 0., 0., 0., 1.],\n",
       "       [0., 0., 0., 1., 0.],\n",
       "       [0., 0., 0., 0., 1.]])"
      ]
     },
     "execution_count": 143,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# create an x and y variable with a sample of 200 relevant documents\n",
    "samples = df[df[\"INCLUDE\"] == 1].sample(200).index\n",
    "X = df.iloc[samples][\"abstract\"]\n",
    "y = df.iloc[samples][impacts].values\n",
    "np.set_printoptions(threshold=5)\n",
    "y"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we just need to slightly amend our pipeline by defining how we want to approach the multilabel problem.\n",
    "\n",
    "We use a One vs Rest classifier which simply creates a binary classifier for each class"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 144,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<style>#sk-container-id-12 {color: black;background-color: white;}#sk-container-id-12 pre{padding: 0;}#sk-container-id-12 div.sk-toggleable {background-color: white;}#sk-container-id-12 label.sk-toggleable__label {cursor: pointer;display: block;width: 100%;margin-bottom: 0;padding: 0.3em;box-sizing: border-box;text-align: center;}#sk-container-id-12 label.sk-toggleable__label-arrow:before {content: \"▸\";float: left;margin-right: 0.25em;color: #696969;}#sk-container-id-12 label.sk-toggleable__label-arrow:hover:before {color: black;}#sk-container-id-12 div.sk-estimator:hover label.sk-toggleable__label-arrow:before {color: black;}#sk-container-id-12 div.sk-toggleable__content {max-height: 0;max-width: 0;overflow: hidden;text-align: left;background-color: #f0f8ff;}#sk-container-id-12 div.sk-toggleable__content pre {margin: 0.2em;color: black;border-radius: 0.25em;background-color: #f0f8ff;}#sk-container-id-12 input.sk-toggleable__control:checked~div.sk-toggleable__content {max-height: 200px;max-width: 100%;overflow: auto;}#sk-container-id-12 input.sk-toggleable__control:checked~label.sk-toggleable__label-arrow:before {content: \"▾\";}#sk-container-id-12 div.sk-estimator input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-12 div.sk-label input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-12 input.sk-hidden--visually {border: 0;clip: rect(1px 1px 1px 1px);clip: rect(1px, 1px, 1px, 1px);height: 1px;margin: -1px;overflow: hidden;padding: 0;position: absolute;width: 1px;}#sk-container-id-12 div.sk-estimator {font-family: monospace;background-color: #f0f8ff;border: 1px dotted black;border-radius: 0.25em;box-sizing: border-box;margin-bottom: 0.5em;}#sk-container-id-12 div.sk-estimator:hover {background-color: #d4ebff;}#sk-container-id-12 div.sk-parallel-item::after {content: \"\";width: 100%;border-bottom: 1px solid gray;flex-grow: 1;}#sk-container-id-12 div.sk-label:hover label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-12 div.sk-serial::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: 0;}#sk-container-id-12 div.sk-serial {display: flex;flex-direction: column;align-items: center;background-color: white;padding-right: 0.2em;padding-left: 0.2em;position: relative;}#sk-container-id-12 div.sk-item {position: relative;z-index: 1;}#sk-container-id-12 div.sk-parallel {display: flex;align-items: stretch;justify-content: center;background-color: white;position: relative;}#sk-container-id-12 div.sk-item::before, #sk-container-id-12 div.sk-parallel-item::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: -1;}#sk-container-id-12 div.sk-parallel-item {display: flex;flex-direction: column;z-index: 1;position: relative;background-color: white;}#sk-container-id-12 div.sk-parallel-item:first-child::after {align-self: flex-end;width: 50%;}#sk-container-id-12 div.sk-parallel-item:last-child::after {align-self: flex-start;width: 50%;}#sk-container-id-12 div.sk-parallel-item:only-child::after {width: 0;}#sk-container-id-12 div.sk-dashed-wrapped {border: 1px dashed gray;margin: 0 0.4em 0.5em 0.4em;box-sizing: border-box;padding-bottom: 0.4em;background-color: white;}#sk-container-id-12 div.sk-label label {font-family: monospace;font-weight: bold;display: inline-block;line-height: 1.2em;}#sk-container-id-12 div.sk-label-container {text-align: center;}#sk-container-id-12 div.sk-container {/* jupyter's `normalize.less` sets `[hidden] { display: none; }` but bootstrap.min.css set `[hidden] { display: none !important; }` so we also need the `!important` here to be able to override the default hidden behavior on the sphinx rendered scikit-learn.org. See: https://github.com/scikit-learn/scikit-learn/issues/21755 */display: inline-block !important;position: relative;}#sk-container-id-12 div.sk-text-repr-fallback {display: none;}</style><div id=\"sk-container-id-12\" class=\"sk-top-container\"><div class=\"sk-text-repr-fallback\"><pre>Pipeline(steps=[(&#x27;vec&#x27;, TfidfVectorizer()),\n",
       "                (&#x27;clf&#x27;,\n",
       "                 OneVsRestClassifier(estimator=SVC(class_weight=&#x27;balanced&#x27;,\n",
       "                                                   probability=True)))])</pre><b>In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook. <br />On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.</b></div><div class=\"sk-container\" hidden><div class=\"sk-item sk-dashed-wrapped\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-42\" type=\"checkbox\" ><label for=\"sk-estimator-id-42\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">Pipeline</label><div class=\"sk-toggleable__content\"><pre>Pipeline(steps=[(&#x27;vec&#x27;, TfidfVectorizer()),\n",
       "                (&#x27;clf&#x27;,\n",
       "                 OneVsRestClassifier(estimator=SVC(class_weight=&#x27;balanced&#x27;,\n",
       "                                                   probability=True)))])</pre></div></div></div><div class=\"sk-serial\"><div class=\"sk-item\"><div class=\"sk-estimator sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-43\" type=\"checkbox\" ><label for=\"sk-estimator-id-43\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">TfidfVectorizer</label><div class=\"sk-toggleable__content\"><pre>TfidfVectorizer()</pre></div></div></div><div class=\"sk-item sk-dashed-wrapped\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-44\" type=\"checkbox\" ><label for=\"sk-estimator-id-44\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">clf: OneVsRestClassifier</label><div class=\"sk-toggleable__content\"><pre>OneVsRestClassifier(estimator=SVC(class_weight=&#x27;balanced&#x27;, probability=True))</pre></div></div></div><div class=\"sk-parallel\"><div class=\"sk-parallel-item\"><div class=\"sk-item\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-45\" type=\"checkbox\" ><label for=\"sk-estimator-id-45\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">estimator: SVC</label><div class=\"sk-toggleable__content\"><pre>SVC(class_weight=&#x27;balanced&#x27;, probability=True)</pre></div></div></div><div class=\"sk-serial\"><div class=\"sk-item\"><div class=\"sk-estimator sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-46\" type=\"checkbox\" ><label for=\"sk-estimator-id-46\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">SVC</label><div class=\"sk-toggleable__content\"><pre>SVC(class_weight=&#x27;balanced&#x27;, probability=True)</pre></div></div></div></div></div></div></div></div></div></div></div></div>"
      ],
      "text/plain": [
       "Pipeline(steps=[('vec', TfidfVectorizer()),\n",
       "                ('clf',\n",
       "                 OneVsRestClassifier(estimator=SVC(class_weight='balanced',\n",
       "                                                   probability=True)))])"
      ]
     },
     "execution_count": 144,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.multiclass import OneVsRestClassifier\n",
    "\n",
    "clf = Pipeline(\n",
    "    steps=[\n",
    "        (\"vec\", TfidfVectorizer()),\n",
    "        (\"clf\", OneVsRestClassifier(SVC(probability=True, class_weight=\"balanced\"))),\n",
    "    ]\n",
    ")\n",
    "\n",
    "clf.fit(X, y)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In the same way as before, we can make predictions with some sample texts"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 145,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>12 - Coastal and marine Ecosystems</th>\n",
       "      <th>12 - Human and managed</th>\n",
       "      <th>12 - Mountains, snow and ice</th>\n",
       "      <th>12 - Rivers, lakes, and soil moisture</th>\n",
       "      <th>12 - Terrestrial ES</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   12 - Coastal and marine Ecosystems  12 - Human and managed   \n",
       "0                                   0                       0  \\\n",
       "1                                   0                       0   \n",
       "\n",
       "   12 - Mountains, snow and ice  12 - Rivers, lakes, and soil moisture   \n",
       "0                             0                                      0  \\\n",
       "1                             0                                      0   \n",
       "\n",
       "   12 - Terrestrial ES  \n",
       "0                    1  \n",
       "1                    0  "
      ]
     },
     "execution_count": 145,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "example_texts = [\n",
    "    \"Climate change has affected butterfly phenology in North America. \"\n",
    "    + \"In an experiment in the butterfly national park, we found that \"\n",
    "    + \"mean dates of appearance had shifted by two weeks. Species traits, \"\n",
    "    + \"especially ecological traits, were found to explain some part of \"\n",
    "    + \"different phenological responses\",\n",
    "    \"Rising temperatures have contributed to increasing heat related mortaility \"\n",
    "    + \"in cities across Europe. \",\n",
    "]\n",
    "y_pred = clf.predict(example_texts)\n",
    "\n",
    "pd.DataFrame(y_pred, columns=impacts)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Transfer learning with transformers!\n",
    "\n",
    "A more powerful way to do classification is to fine-tune a langauge model which has been pretrained on large amounts of text. For an introduction to how these work (which is outside the scope of this tutorial) check out [The Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/).\n",
    "\n",
    "The [Transformers](https://huggingface.co/docs/transformers) library makes doing this really easy, and there are a wealth of models available with different features and which have been pretrained on different corpora. We are going to use [ClimateBert](https://huggingface.co/climatebert), which has undergone additional pretraining on texts from the climate domain (included scientific articles)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 146,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Let's take our texts and our labels again\n",
    "# Each example is a tuple of (text, label)\n",
    "# Where a label is 1 if the text is relevant and should be included in a synthesis report and 0 otherwise\n",
    "texts, y = zip(\n",
    "    *[\n",
    "        (\"Climate change is impacting human systems\", 1),\n",
    "        (\"Climate change is caused by fossil fuels\", 0),\n",
    "        (\"Agricultural yields are affected by climate change\", 1),\n",
    "        (\"System change not climate change\", 0),\n",
    "        (\"higher temperatures are impacting human health\", 1),\n",
    "        (\"Forest fires are becoming more frequent due to climate change\", 1),\n",
    "        (\"Machine learning can read texts\", 0),\n",
    "        (\"AI can help solve climate change!\", 0),\n",
    "        (\"We need to save gas this winter\", 0),\n",
    "        (\"More frequent droughts are impacting crop yields\", 1),\n",
    "        (\"Many communities are affected by rising sea levels\", 1),\n",
    "        (\"Global emissions continue to rise\", 0),\n",
    "        (\"Ecosystems are increasingly impacted by rising temperatures\", 1),\n",
    "        (\"Emissions from fossil fuels need to decline\", 0),\n",
    "        (\"Anthropogenic climate change is impacting vulnerable communities\", 1),\n",
    "    ]\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 147,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                  \r"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'text': 'Climate change is impacting human systems',\n",
       " 'label': 1,\n",
       " 'input_ids': [0,\n",
       "  40466,\n",
       "  464,\n",
       "  16,\n",
       "  18081,\n",
       "  1050,\n",
       "  1437,\n",
       "  50278,\n",
       "  2,\n",
       "  1,\n",
       "  1,\n",
       "  1,\n",
       "  1,\n",
       "  1,\n",
       "  1],\n",
       " 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0]}"
      ]
     },
     "execution_count": 147,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# To use these with transformers, we are going to need to get them into the right format.\n",
    "from datasets import Dataset\n",
    "from transformers import AutoTokenizer\n",
    "\n",
    "# First we'll put them into a HuggingFace Dataset object\n",
    "dataset = Dataset.from_dict({\"text\": texts, \"label\": y})\n",
    "\n",
    "# And now we need to tokenize the texts, using the pretrained tokenizer from climatebert\n",
    "model_name = \"climatebert/distilroberta-base-climate-f\"\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
    "\n",
    "\n",
    "def tokenize_function(examples):\n",
    "    return tokenizer(examples[\"text\"], padding=\"longest\", truncation=True)\n",
    "\n",
    "\n",
    "tokenized_dataset = dataset.map(tokenize_function, batched=True)\n",
    "# input_ids are the token ids for each token in the text\n",
    "# attention_mask is a mask that tells the model which tokens to pay attention to and which to ignore\n",
    "# Note the attention mask has 0s for padded tokens.\n",
    "tokenized_dataset[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 148,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                  \r"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'text': 'Climate change is impacting human systems',\n",
       " 'label': 1,\n",
       " 'input_ids': [0,\n",
       "  40466,\n",
       "  464,\n",
       "  16,\n",
       "  18081,\n",
       "  1050,\n",
       "  1437,\n",
       "  50278,\n",
       "  2,\n",
       "  1,\n",
       "  1,\n",
       "  1,\n",
       "  1,\n",
       "  1,\n",
       "  1],\n",
       " 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0]}"
      ]
     },
     "execution_count": 148,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# We can wrap this into one function that turns any set of texts (and optional labels)\n",
    "# into a tokenized huggingface dataset\n",
    "def datasetify(x, tokenizer, y=None):\n",
    "    data_dict = {\"text\": x}\n",
    "    if y is not None:\n",
    "        data_dict[\"label\"] = y\n",
    "    dataset = Dataset.from_dict(data_dict)\n",
    "\n",
    "    def tokenize_function(examples):\n",
    "        return tokenizer(examples[\"text\"], padding=\"longest\", truncation=True)\n",
    "\n",
    "    return dataset.map(tokenize_function, batched=True)\n",
    "\n",
    "\n",
    "tokenized_dataset = datasetify(texts, tokenizer, y)\n",
    "tokenized_dataset[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 149,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at climatebert/distilroberta-base-climate-f were not used when initializing RobertaForSequenceClassification: ['lm_head.layer_norm.weight', 'lm_head.dense.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias']\n",
      "- This IS expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
      "Some weights of RobertaForSequenceClassification were not initialized from the model checkpoint at climatebert/distilroberta-base-climate-f and are newly initialized: ['classifier.dense.bias', 'classifier.out_proj.bias', 'classifier.out_proj.weight', 'classifier.dense.weight']\n",
      "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n",
      "/home/dspokoyn/.cache/pypoetry/virtualenvs/tutorial-IHiCU4EY-py3.8/lib/python3.8/site-packages/transformers/optimization.py:391: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning\n",
      "  warnings.warn(\n",
      "/home/dspokoyn/.cache/pypoetry/virtualenvs/tutorial-IHiCU4EY-py3.8/lib/python3.8/site-packages/torch/nn/parallel/_functions.py:68: UserWarning: Was asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.\n",
      "  warnings.warn('Was asked to gather along dimension 0, but all '\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='3' max='3' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [3/3 00:00, Epoch 3/3]\n",
       "    </div>\n",
       "    <table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       " <tr style=\"text-align: left;\">\n",
       "      <th>Step</th>\n",
       "      <th>Training Loss</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "  </tbody>\n",
       "</table><p>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": [
       "TrainOutput(global_step=3, training_loss=0.681313673655192, metrics={'train_runtime': 0.7791, 'train_samples_per_second': 57.76, 'train_steps_per_second': 3.851, 'total_flos': 174639636900.0, 'train_loss': 0.681313673655192, 'epoch': 3.0})"
      ]
     },
     "execution_count": 149,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Now we want to load our model, and instantiate a Trainer class\n",
    "from transformers import AutoModelForSequenceClassification, Trainer, logging\n",
    "import torch\n",
    "from torch.nn import functional as F\n",
    "\n",
    "logging.set_verbosity_warning()\n",
    "\n",
    "# We set num_labels to 2 for binary classification, as we have two classes - positive and negative\n",
    "model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=2)\n",
    "# The trainer class needs to be supplied with a model, and a dataset (and will also accept TrainingArguments and validation data)\n",
    "\n",
    "trainer = Trainer(model=model, train_dataset=datasetify(texts, tokenizer, y))\n",
    "# Once this has been instantiated we can apply the train() method\n",
    "trainer.train()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 150,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                 "
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\r"
     ]
    },
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": [
       "PredictionOutput(predictions=array([[0.04632454, 0.16618322],\n",
       "       [0.07823081, 0.13172969],\n",
       "       [0.04848875, 0.134009  ]], dtype=float32), label_ids=array([1, 0, 0]), metrics={'test_loss': 0.6973627209663391, 'test_runtime': 0.0827, 'test_samples_per_second': 36.295, 'test_steps_per_second': 12.098})"
      ]
     },
     "execution_count": 150,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# To generate predictions, we just need to supply a dataset to the predict method\n",
    "new_texts = [\n",
    "    \"climate change is impacting terrestrial ecosystems\",\n",
    "    \"Machine Learning will solve climate change\",\n",
    "    \"Supreme court made a series of decisions on personal freedom\",\n",
    "]\n",
    "\n",
    "\n",
    "pred = trainer.predict(datasetify(new_texts, tokenizer, [1, 0, 0]))\n",
    "pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 151,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/tmp/ipykernel_9484/30200511.py:10: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.\n",
      "  activation(tensor(pred.predictions))\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "tensor([[0.4701, 0.5299],\n",
       "        [0.4866, 0.5134],\n",
       "        [0.4786, 0.5214]])"
      ]
     },
     "execution_count": 151,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# However, the model output gives us logits. If these are negative, then the prediction\n",
    "# is negative, if they are positive, the prediction is positive.\n",
    "# We can turn these into probabilities with an activation function\n",
    "from torch import tensor\n",
    "from torch.nn import Sigmoid, Softmax\n",
    "\n",
    "activation = (\n",
    "    Softmax()\n",
    ")  # Since we have two exclusive classes, we use the Softmax function\n",
    "activation(tensor(pred.predictions))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 152,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                  \r"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='3' max='3' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [3/3 00:00, Epoch 3/3]\n",
       "    </div>\n",
       "    <table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       " <tr style=\"text-align: left;\">\n",
       "      <th>Step</th>\n",
       "      <th>Training Loss</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "  </tbody>\n",
       "</table><p>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                 \r"
     ]
    },
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": [
       "array([[0.42267254, 0.57732743],\n",
       "       [0.49319434, 0.50680566],\n",
       "       [0.4695467 , 0.53045326]], dtype=float32)"
      ]
     },
     "execution_count": 152,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# If we want to always get probabilities, we can subclass Trainer and add a new predict_proba method\n",
    "\n",
    "from transformers.trainer_utils import PredictionOutput\n",
    "\n",
    "\n",
    "class ProbTrainer(Trainer):\n",
    "    def predict_proba(self, test_dataset: Dataset) -> PredictionOutput:\n",
    "        logits = self.predict(test_dataset).predictions\n",
    "        if logits.shape[1] > 2:\n",
    "            activation = Sigmoid()\n",
    "        else:\n",
    "            activation = Softmax(dim=1)\n",
    "        return activation(tensor(logits)).numpy()\n",
    "\n",
    "\n",
    "trainer = ProbTrainer(model=model, train_dataset=datasetify(texts, tokenizer, y))\n",
    "trainer.train()\n",
    "\n",
    "pred = trainer.predict_proba(datasetify(new_texts, tokenizer))\n",
    "pred\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 153,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train size: 2000\n",
      "Test size: 500\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at climatebert/distilroberta-base-climate-f were not used when initializing RobertaForSequenceClassification: ['lm_head.layer_norm.weight', 'lm_head.dense.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias']\n",
      "- This IS expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
      "Some weights of RobertaForSequenceClassification were not initialized from the model checkpoint at climatebert/distilroberta-base-climate-f and are newly initialized: ['classifier.dense.bias', 'classifier.out_proj.bias', 'classifier.out_proj.weight', 'classifier.dense.weight']\n",
      "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n",
      "                                                                 \r"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='189' max='189' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [189/189 01:08, Epoch 3/3]\n",
       "    </div>\n",
       "    <table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       " <tr style=\"text-align: left;\">\n",
       "      <th>Step</th>\n",
       "      <th>Training Loss</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "  </tbody>\n",
       "</table><p>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                   \r"
     ]
    },
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "test_pred 500 [[9.9961412e-01 3.8582133e-04]\n",
      " [9.9911886e-01 8.8113756e-04]\n",
      " [4.7982158e-04 9.9952018e-01]\n",
      " ...\n",
      " [2.5812304e-04 9.9974185e-01]\n",
      " [9.9651754e-01 3.4824908e-03]\n",
      " [3.1496082e-02 9.6850395e-01]]\n",
      "              precision    recall  f1-score   support\n",
      "\n",
      "           0       0.90      0.80      0.85       235\n",
      "           1       0.84      0.92      0.88       265\n",
      "\n",
      "    accuracy                           0.86       500\n",
      "   macro avg       0.87      0.86      0.86       500\n",
      "weighted avg       0.87      0.86      0.86       500\n",
      "\n"
     ]
    }
   ],
   "source": [
    "def one_hot_encode(targets, num_labels):\n",
    "    '''Helper function to one-hot encode the labels'''\n",
    "    return F.one_hot(targets, num_labels)\n",
    "\n",
    "\n",
    "def train_eval_transformer(df):\n",
    "    ''' Trains a pretrained transformer model for the binary classification task\n",
    "        of predicting whether a study abstract should be included in a synthesis report.\n",
    "    '''\n",
    "    # load the train/test splits\n",
    "    x_train, y_train, x_test, y_test = train_test_split_df(df)\n",
    "    # build the classifier\n",
    "    # We set num_labels to 2 for binary classification, as we have two classes - positive and negative\n",
    "    model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=2)\n",
    "    y_train_one_hot = one_hot_encode(torch.tensor(y_train), 2)\n",
    "\n",
    "    # trainer = Trainer(model=model, train_dataset=datasetify(x_train, tokenizer, y_train_one_hot))\n",
    "    trainer = ProbTrainer(model=model, train_dataset=datasetify(x_train, tokenizer, y_train_one_hot))\n",
    "\n",
    "\n",
    "    # Once this has been instantiated we can apply the train() method\n",
    "    trainer.train()\n",
    "\n",
    "    # get the predictions on the test set\n",
    "    test_pred = trainer.predict_proba(datasetify(x_test, tokenizer))\n",
    "    print('test_pred', len(test_pred), test_pred)\n",
    "\n",
    "    # Get the classification evaluation metrics\n",
    "    # This will be covered in more detail later.\n",
    "    print(classification_report(y_test, test_pred.argmax(axis=1)))\n",
    "    # f1-score: should be around 88% when trained with 80% of the training data.\n",
    "    \n",
    "\n",
    "# So far this was an illustrative example using toy data. Let's now apply this to our real data.\n",
    "# takes around 6 hours to run\n",
    "# to speed this up you can use less training data\n",
    "# or use a smaller model\n",
    "# train_eval_transformer(df)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 154,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0., 1., 0., 0., 0.],\n",
       "       [0., 0., 0., 0., 1.],\n",
       "       [0., 0., 0., 0., 1.],\n",
       "       [0., 0., 0., 0., 1.],\n",
       "       [0., 0., 0., 1., 0.]])"
      ]
     },
     "execution_count": 154,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "### To do multilabel learning, we just need the data in the same format as before\n",
    "# create an x and y variable with a sample of 200 relevant documents\n",
    "samples = df[df[\"INCLUDE\"] == 1].sample(5).index\n",
    "X = df.iloc[samples][\"abstract\"]\n",
    "y = df.iloc[samples][impacts].values\n",
    "y"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 155,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at climatebert/distilroberta-base-climate-f were not used when initializing RobertaForSequenceClassification: ['lm_head.layer_norm.weight', 'lm_head.dense.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias']\n",
      "- This IS expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
      "Some weights of RobertaForSequenceClassification were not initialized from the model checkpoint at climatebert/distilroberta-base-climate-f and are newly initialized: ['classifier.dense.bias', 'classifier.out_proj.bias', 'classifier.out_proj.weight', 'classifier.dense.weight']\n",
      "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n",
      "/home/dspokoyn/.cache/pypoetry/virtualenvs/tutorial-IHiCU4EY-py3.8/lib/python3.8/site-packages/transformers/optimization.py:391: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning\n",
      "  warnings.warn(\n",
      "/home/dspokoyn/.cache/pypoetry/virtualenvs/tutorial-IHiCU4EY-py3.8/lib/python3.8/site-packages/torch/nn/parallel/_functions.py:68: UserWarning: Was asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.\n",
      "  warnings.warn('Was asked to gather along dimension 0, but all '\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='3' max='3' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [3/3 00:00, Epoch 3/3]\n",
       "    </div>\n",
       "    <table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       " <tr style=\"text-align: left;\">\n",
       "      <th>Step</th>\n",
       "      <th>Training Loss</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "  </tbody>\n",
       "</table><p>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "                                                 \r"
     ]
    },
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": [
       "array([[0.4031779 , 0.48949987, 0.42142   , 0.48305014, 0.5281289 ],\n",
       "       [0.40742624, 0.49437627, 0.41965604, 0.48600367, 0.5311437 ],\n",
       "       [0.42072555, 0.4901937 , 0.42390627, 0.48873633, 0.5178754 ]],\n",
       "      dtype=float32)"
      ]
     },
     "execution_count": 155,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=5)\n",
    "trainer = ProbTrainer(model=model, train_dataset=datasetify(X, tokenizer, y))\n",
    "trainer.train()\n",
    "\n",
    "pred = trainer.predict_proba(datasetify(new_texts, tokenizer))\n",
    "pred"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a name=\"model-evaluation\"></a>\n",
    "\n",
    "# Model evaluation\n",
    "\n",
    "<!-- [![colab badge](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mcallaghan/NLP-climate-science-tutorial-CCAI/blob/main/C_evaluation.ipynb) -->\n",
    "\n",
    "Now we know how to build a model, we need to \n",
    "- choose the best model, \n",
    "- estimate our chosen model's performance, \n",
    "- make predictions for the documents we haven't seen \n",
    "\n",
    "First we will look at the basics of evaluation model outputs\n",
    "\n",
    "## Metrics\n",
    "\n",
    "When we have labelled data, evaluation works by comparing the labels we gave to the labels the model predicts. The closer the predicted labels are to the true labels, the better our model. **Metrics** quantify different aspects of that closeness, and we'll explore a few of the commonly used ones"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Accuracy\n",
    "\n",
    "Accuracy is a frequently used metric, because it is very easy to understand and to explain. It simply describes the proportion of predictions which are correct:\n",
    "\n",
    "$ Accuracy = \\frac{CorrectPredictions}{TotalPredictions} $"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 156,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "8 correct predictions out of 10 total predictions.\n",
      "Accuracy = 80%\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "\n",
    "# Let's examine\n",
    "y_pred = np.array([0, 0, 0, 0, 0, 1, 1, 0, 0, 0])\n",
    "y_true = np.array([0, 0, 0, 0, 1, 1, 0, 0, 0, 0])\n",
    "\n",
    "def calculate_accuracy(y_true, y_pred):\n",
    "    total_predictions = y_true.size\n",
    "    correct_predictions = (y_pred==y_true).sum()\n",
    "    accuracy = correct_predictions / total_predictions\n",
    "    print(f\"{correct_predictions} correct predictions out of {total_predictions} total predictions.\\nAccuracy = {accuracy:.0%}\")\n",
    "    \n",
    "calculate_accuracy(y_true, y_pred)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This seems pretty good, but because the classes are imbalanced, we could get exactly the same accuracy just by predicting 0 for each sample "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 157,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "8 correct predictions out of 10 total predictions.\n",
      "Accuracy = 80%\n"
     ]
    }
   ],
   "source": [
    "calculate_accuracy(y_true, np.zeros(y_true.size))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Accuracy only tells a part of the story, and under quite common conditions, this story can be misleading. \n",
    "\n",
    "Let's therefore take a look at some other metrics that give us more specific information about how the model performs, and tells us what kinds of mistakes it tends to make how frequently\n",
    "\n",
    "## Precision and Recall\n",
    "\n",
    "Predictions can have two types of errors: false positives (where a sample is negative and it is predicted to be positive) & false negatives (where a sample is positive and it is predicted to be negative). Correspondingly there are also two types of correct predictions: true positives & true negatives. In different contexts, these types of errors can have different costs. For example, in health screening applications, false negatives can carry very high costs, as they could result in a serious condition going untreated. Precision and recall are ways of describing the prevalence of each type of error.\n",
    "\n",
    "### Precision\n",
    "\n",
    "Precision describes the fraction of samples predicted to be positive that are actually positive. \n",
    "\n",
    "$ Precision = \\frac{TruePositives}{TruePositives + FalsePositives} $\n",
    "\n",
    "In other words, if a classifier identifies 100 samples as true, precision tells us how many of those samples are actually true.\n",
    "\n",
    "### Recall\n",
    "\n",
    "Recall describes the fraction of positive samples which are identified by the classifier.\n",
    "\n",
    "$ Recall = \\frac{TruePositives}{TruePositives + FalseNegatives} $\n",
    "\n",
    "In other words, if 100 samples are true, recall tells us how many of those samples our classifier will identify.\n",
    "\n",
    "## F1\n",
    "\n",
    "The F1 score is the average of the precision and recall. It can be weighted (in which case it becomes an $ F{\\beta} $ score if we value either precision or recall more highly\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 158,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1 out of 2 predicted positives are actually positive\n",
      "precision = 50%\n",
      "0 out of 0 predicted positives are actually positive\n",
      "precision = nan%\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/tmp/ipykernel_9484/4293625237.py:4: RuntimeWarning: invalid value encountered in scalar divide\n",
      "  precision = tps / positive_predictions\n"
     ]
    }
   ],
   "source": [
    "def calculate_precision(y_true, y_pred):\n",
    "    tps = (y_true+y_pred==2).sum()\n",
    "    positive_predictions = (y_pred==1).sum()\n",
    "    precision = tps / positive_predictions\n",
    "    print(f\"{tps} out of {positive_predictions} predicted positives are actually positive\")\n",
    "    print(f\"precision = {precision:.0%}\")\n",
    "\n",
    "# Our predictions achieve a much more modest precision score\n",
    "calculate_precision(y_true, y_pred)\n",
    "# Whereas predicting all zeros gives us NA result as the denominator is zero\n",
    "calculate_precision(y_true, np.zeros(y_true.size))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 123,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1 out of 2 true samples are identified\n",
      "recall = 50%\n",
      "0 out of 2 true samples are identified\n",
      "recall = 0%\n"
     ]
    }
   ],
   "source": [
    "def calculate_recall(y_true, y_pred):\n",
    "    tps = (y_true+y_pred==2).sum()\n",
    "    positive_samples = (y_true==1).sum()\n",
    "    recall = tps / positive_samples\n",
    "    print(f\"{tps} out of {positive_samples} true samples are identified\")\n",
    "    print(f\"recall = {recall:.0%}\")\n",
    "\n",
    "# Recall is also more modest with our predictions\n",
    "calculate_recall(y_true, y_pred)\n",
    "# And a prediction of all zeros has zero recall\n",
    "calculate_recall(y_true, np.zeros(y_true.size))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "tutorial-IHiCU4EY-py3.8",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.0"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}