Newsroom requires Python 3.6+ and can be installed using pip
:
pip install -e git+git://github.com/clic-lab/newsroom.git#egg=newsroom
There are two ways to obtain the summaries dataset. You may use the scripts described below to scrape the web pages used in the dataset and extract the summaries. Alternatively, the complete dataset is also available from https://summari.es/download/.
Newsroom contains two scripts for downloading and processing data downloaded from Archive.org. First, download the "Thin Dataset" from https://summari.es/download/.
(The "Data Builder" is this Python package.)
Download and extract thin.tar
with tar xvf thin.tar
, yielding directory thin
containing several .jsonl.gz
files.
Next, use newsroom-scrape
and newsroom-extract
to process the data, as described below.
Both of these tools have additional argument help pages when you use the --help
command line option.
The thin
directory will contain three files, train.jsonl.gz
, dev.jsonl.gz
and test.jsonl.gz
. To begin downloading the development set from Archive.org, run the following:
newsroom-scrape --thin thin/dev.jsonl.gz --archive dev.archive
Estimated download time is indicated with a progress bar. If errors occur during downloading, you may need to re-run the script later to capture the missing articles. This process is network bound and depends mostly on Archive.org, save your CPU cycles for the extraction stage!
The downloading process can be stopped at any time with Control-C
and resumed later. It is also possible to perform extraction of a partially downloaded dataset with newsroom-extract
before continuing to download the full version.
The newsroom-extract
tool extracts summaries and article text from the data downloaded by newsroom-scrape
. This tool produces a new file that does not modify the original output file of newsroom-scrape
, and can be run with:
newsroom-extract --archive dev.archive --dataset dev.dataset
The script automatically parallelizes extraction across your CPU cores. To disable this or reduce the number of cores used, use the --workers
option. Like scraping, the extraction process can be stopped at any point with Control-C
and resumed later.
All data are represented using gzip-compressed JSON lines. The Newsroom package provides an easy tool to read an write these files — and do so up to 20x faster than the standard Python gz
and json
packages!
from newsroom import jsonl
# Read entire file:
with jsonl.open("train.dataset", gzip = True) as train_file:
train = train_file.read()
# Read file entry by entry:
with jsonl.open("train.dataset", gzip = True) as train_file:
for entry in train_file:
print(entry["summary"], entry["text"])
The Newsroom package also contains scripts for identifying extractive fragments and computing metrics described in the paper: coverage, density, and compression.
import random
from newsroom import jsonl
from newsroom.analyze import Fragments
with jsonl.open("train.dataset", gzip = True) as train_file:
train = train_file.read()
# Compute stats on random training example:
entry = random.choice(train)
summary, text = entry["summary"], entry["text"]
fragments = Fragments(summary, text)
# Print paper metrics:
print("Coverage:", fragments.coverage())
print("Density:", fragments.density())
print("Compression:", fragments.compression())
# Extractive fragments oracle:
print("List of extractive fragments:")
print(fragments.strings())
[Aug 8, 2023] The evaluation pipeline does not keep track of sentence tokenization, which may result in lower-than-expected ROUGE-L scores, and that the newsroom-run -> newsroom-score -> newsroom-tables evaluation pipeline should not be used for publishable evaluation. Please see: #28
The Newsroom package contains a standardized way for running and scoring Docker-based summarization systems. For an example, see the /example
directory for a Docker image of the TextRank system used in the paper.
The package also contains a script for producing tables similar to those in the paper for compression, coverage, and density. These tables are helpful for understanding your system's performance across different difficulties of text-summary pairs.
After starting Docker and building your image (named "textrank" in the following examples), the system can be evaluated using the script:
newsroom-run \
--system textrank \ # Name of Docker image.
--dataset dev.dataset \ # Path to evaluation data.
--summaries textrank.summaries \ # Output path to write system summaries.
--keys text # JSON keys to feed Docker system.
The script runs your system Docker image, passes article text (and other requested metadata) into the container through standard input, expecting summaries to be supplied on standard output.
To score your system, run the following:
newsroom-score \
--dataset dev.dataset \ # Path to evaluation data.
--summaries textrank.summaries \ # Path to system's output summaries.
--scores textrank.scores \ # Output path to write summary scores.
--rouge 1,2,L \ # ROUGE variants to run.
--unstemmed # Or, --stemmed for Porter stemming.
The script produces a file (textrank.scores
) containing pairs of system and reference summaries, article metadata for analysis, and ROUGE scores. Additionally, overall ROUGE scores are printed on completion.
To produce ROUGE tables across Newsroom compression, density, and coverage subsets, run the following:
newsroom-tables \
--scores textrank.scores \
--rouge 1,2,L \
--variants fscore \
--bins density,compression,coverage
All command line tools have a --help
flag that show a description of arguments and their defaults.