CorrFA_for_Summarizaion

Corr F/A evaluation metrics in paper "Xinnuo Xu, Ondrej Dusek, Jingyi Li, Yannis Konstas, and Verena Rieser. Fact-based Content Weighting for Evaluating Abstractive Summarisation" Proceedings of ACL2020 🎉 🎉 🎉 Video is available here.

Environment setup

Step1: Install pytorch env

(Install conda: wget https://repo.anaconda.com/archive/Anaconda2-2019.10-Linux-x86_64.sh; bash ~/Downloads/Anaconda2-2019.10-Linux-x86_64.sh; Reload)
conda create -n Highlight python=3.6
conda activate Highlight
conda install pytorch=1.1.0 torchvision cudatoolkit=10.0 -c pytorch
(or conda install pytorch torchvision cudatoolkit=10.0 -c pytorch; conda install pytorch=1.1.0 -c soumith)

pip install multiprocess
pip install pytorch_transformers
pip install pyrouge
pip install tensorboardX

Step2: Install allennlp

pip install allennlp
wget https://s3-us-west-2.amazonaws.com/allennlp/models/srl-model-2018.05.25.tar.gz
mv srl-model-2018.05.25.tar.gz Evaluation/

Evaluate your output with CorrF/A

Scenario1: With plain text inputs

We need three files for the evaluation, documents(SRC_PATH), gold summaries(GOLD_PATH), and model generated summaries(CAND_PATH). The format for document-file is one document per line and sentences are jointed by '\t'. The format for both gold-summary-file and generated-summary-file is one summary per line. The i-th row of document-file is paired with i-th row in gold-summary-file and generated-summary-file. The number of lines in each file should be the same. Examples are shown in ./Data/50_files, ./Data/50_files.gold, ./Data/50_files.cand. To calculate CorrF/A, run:

#!/bin/bash

SRC_PATH='./Data/50_files.src'
GOLD_PATH='./Data/50_files.gold'
CAND_PATH='./Data/50_files.cand'

python evaluate.py \
        -src_path ${SRC_PATH} \
        -gold_path ${GOLD_PATH} \
        -cand_path ${CAND_PATH}  

The Corr-F and Corr-A will be printed out. Also, the content weights referring to gold summaries and generated summaries are saved in file ./Data/cw_gold and ./Data/cw_cand respectively.

Scenario1: With Tree structured inputs

If the trees are built and saved in files, the evaluation can be run as:

#!/bin/bash

TREE_PATH='./Data/bert.tree'

python evaluate.py \
        -src_path ${SRC_PATH} \
        -gold_path ${GOLD_PATH} \
        -cand_path ${CAND_PATH} \
        -tree_path ${TREE_PATH} \
        -run_srl False \
        -run_tree False

The example file ./Data/bert.tree is generate in ./Data/ by running python full_cases_tree.py bert. The script reads processed tree MRs of documents, old summaries and generated summaries from

./Data/full_cases/bert_src.tree
./Data/full_cases/bert_gold.tree
./Data/full_cases/bert_cand.tree

respectively. The format for these three files is similar with plain text inputs. The only difference is that sentences are represented in tree MRs.

Evaluation results updates

The original data used in paper is offered in directory ./Data/.

• ./Data/50_cases/ Documents, generated summaries by model TConvS2S, PtGen, BertSumAbs and gold summaries of the 50 human-annotated examples.
• ./Data/AMT_data/ Content weights highlighted by human judges using the Amazon Mechanical Turk platform.
• ./Data/full_cases/ Documents, generated summaries by model TConvS2S, PtGen, BertSumAbs and gold summaries of the full XSum test set.

After fixing some minor problems, the experiment results are updated as below. The conclusion made in the paper is not changed.

Lower half of Table 2 in paper is updated as:

Model	Corr-F	Corr-A
TConvS2S	0.6584	0.6495
PtGen	0.6413	0.6106
BertSumAbs	0.7080	0.6798

The first column of Table 4 (CorrF/A) in paper is updated as:

Model	Corr-F	Corr-A
TConvS2S	0.6157	0.6395
PtGen	0.6008	0.6268
BertSumAbs	0.6579	0.6865

Other features

• ./Debug/ offers a tool for content weighting visualization and system attacking. The README.md in the directory explains the tool in detail.