Skip to content

Latest commit

 

History

History
345 lines (261 loc) · 12.3 KB

README.md

File metadata and controls

345 lines (261 loc) · 12.3 KB

Argument Mining in Scientific Reviews (AMSR)

Accompanying repository of our AAAI2021 paper "Argument Mining Driven Analysis of Peer-Reviews".

We release a new dataset of peer-reviews from different computer science conferences with annotated arguments, called AMSR (Argument Mining in Scientific Reviews).

Dataset

The dataset is available here. It contains:

  1. Raw Conference Data
  2. Cleaned Conference Data
  3. Annotated Conference Data

Requirements

The requirements of our code are given in requirements.txt. Make sure you have at least Python 3.7 with pip installed. Navigate to the root folder of the repository, where the file requirements.txt is located, and install all dependencies in the current Python environment with

pip install -r requirements.txt

Steps

In the following, we explain how to reproduce all steps from scraping the data to training the acceptance classification model.

1. Scrape Data

conferences_raw/ contains directories for each conference we scraped. The respective directory of each conference comprises multiple *.json files, where every file contains the information belonging to a single paper, such as the title, the abstract, the submission date and the reviews. The reviews are stored in a list called "review_content".

In scrape_data/executables you can find the scripts to get the data via the OpenReview-API from OpenReview. The following commands allow you to reproduce the data scraping for individual conferences. The scripts create the output directories automatically, if they are not already present.

AKBC19 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation AKBC.ws/2019/Conference/-/Blind_Submission 
--review_invitation AKBC.ws/2019/Conference/-/Paper.*/Official_Review 
--comment_invitation AKBC.ws/2019/Conference/-/Paper.*/Official_Comment 
--metareview_invitation AKBC.ws/2019/Conference/-/Paper.*/Meta_Review 
--outdir 1_scrape_data/data
--conference_name akbc19 
--invitation AKBC.ws/2019/Conference
AKBC20 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation AKBC.ws/2020/Conference/-/Blind_Submission
--review_invitation  AKBC.ws/2020/Conference/Paper.*/-/Official_Review
--comment_invitation AKBC.ws/2020/Conference/Paper.*/-/Official_Comment
--metareview_invitation AKBC.ws/2020/Conference/Paper.*/-/Decision
--outdir 1_scrape_data/data
--conference_name akbc20 
--invitation AKBC.ws/2020/Conference
ICLR19 (the execution to receive the iclr19 data may take some time, because there are a lot of reviews) 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation ICLR.cc/2019/Conference/-/Blind_Submission
--review_invitation ICLR.cc/2019/Conference/-/Paper.*/Official_Review
--comment_invitation ICLR.cc/2019/Conference/-/Paper.*/Official_Comment
--metareview_invitation ICLR.cc/2019/Conference/-/Paper.*/Meta_Review
--outdir 1_scrape_data/data
--conference_name iclr19
--invitation ICLR.cc/2019/Conference
ICLR20 (the execution to receive the iclr20 data may take some time, because there are a lot of reviews) 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation ICLR.cc/2020/Conference/-/Blind_Submission
--review_invitation ICLR.cc/2020/Conference/Paper.*/-/Official_Review
--comment_invitation ICLR.cc/2020/Conference/Paper.*/-/Official_Comment
--metareview_invitation ICLR.cc/2020/Conference/Paper.*/-/Decision
--outdir 1_scrape_data/data
--conference_name iclr20
--invitation ICLR.cc/2020/Conference
MIDL19 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation MIDL.io/2019/Conference/-/Full_Submission
--review_invitation MIDL.io/2019/Conference/-/Paper.*/Official_Review
--comment_invitation MIDL.io/2019/Conference/-/Paper.*/Official_Comment
--metareview_invitation MIDL.io/2019/Conference/-/Paper.*/Decision
--outdir 1_scrape_data/data
--conference_name midl19
--invitation MIDL.io/2019/Conference
MIDL20 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation MIDL.io/2020/Conference/-/Blind_Submission
--review_invitation MIDL.io/2020/Conference/Paper.*/-/Official_Review
--comment_invitation MIDL.io/2020/Conference/Paper.*/-/Official_Comment
--metareview_invitation MIDL.io/2020/Conference/Paper.*/-/Meta_Review
--outdir 1_scrape_data/data
--conference_name midl20
--invitation MIDL.io/2020/Conference
Graph20 
python 1_scrape_data/executables/scrap_data.py
--submission_invitation graphicsinterface.org/Graphics_Interface/2020/Conference/-/Blind_Submission
--review_invitation graphicsinterface.org/Graphics_Interface/2020/Conference/Paper.*/-/Official_Review
--comment_invitation None
--metareview_invitation graphicsinterface.org/Graphics_Interface/2020/Conference
--outdir 1_scrape_data/data
--conference_name graph20
--invitation graphicsinterface.org/Graphics_Interface/2020/Conference
NeuroAI19 
python 1_scrape_data/executables/scrap_data.py 
--submission_invitation NeurIPS.cc/2019/Workshop/Neuro_AI/-/Blind_Submission 
--review_invitation NeurIPS.cc/2019/Workshop/Neuro_AI/Paper.*/-/Official_Review 
--comment_invitation NeurIPS.cc/2019/Workshop/Neuro_AI/Paper.*/-/Official_Comment 
--metareview_invitation NeurIPS.cc/2019/Workshop/Neuro_AI/Paper.*/-/Meta_Review 
--outdir 1_scrape_data/data 
--invitation NeurIPS.cc/2019/Workshop/Neuro_AI 
--conference_name neuroai19

Important for the NeuroAI19 script is, that you are currently in the base directory of the git, called git, and that the following folder structure already exists "1_scrape_data/data/neuroai19/". Otherwise the execution will throw a FileNotFoundError. When the path inside the script is changed, it does not have to be executed from the git folder.

Note: While we scraped data from 8 conferences, we selected a subset of 6 large conferences to continue (i.e., ICLR19/20, MIDL19/20, AKBC 19/20, Graphics20, NeuroAI19).

2. Clean Data

In a second step, we remove all unwanted character sequences from the reviews. For details on how this works please refer to our paper.

The scripts, located in 2_clean_data/executables, have to be called individually for each conference, since the file format differs across conferences. They read all JSON files, and produce two CSV files: X_reviews.csv and X_papers.csv. The scripts expect scraped inputs in 1_scrape_data/data/ and output to 2_clean_data/data.

cd 2_clean_data/executables
python clean_dataset_midl19.py

3. Annotation Study

Next, the data was annotated by us using the web annotation tool webanno. All scripts and data can be found in 3_annotation_study.

4. Annotation Post-Processing

All steps to reproduce the post-processing of the annotations can be found in 4_post_processing.

The annotated data is processed with new_preparation.py and creates a file which contains all annotated sentences with the majority vote. If there is no majority, the label is None. A different person who has not previously seen this review is consulted to break the tie. These manual tie breaks are done using add_manual_decisions.py. Afterwards, we manually removed sentences with a low quality to facilitate the training. The dataset resulting from this can be found at conferences_annotated folder. For the Argumentation model and the Joint model, we take NONARG tag as majority whereas for the POS_NEG model we ignore the NON tag and take CON as majority.

In the directory 4_post_processing/exploratory_data_analysis, you can find the scripts human_accuracy_sentence.py to get the human performance on sentence level and human_accuracy_token.py for the human performance on token level.

5. Work with Models

Lastly we trained a classification layer on top of different BERT models for multiple tasks and settings. The data of this step can be found in 5_model_works.

It consists of a pre-trained BERT-model (we evaluated cased vs. uncased and BERT-base vs. BERT-large) with a classificaton layer, fine-tuned by us (fine-tuning on combinations of AURC dataset and/or AMSR dataset).

5.1. Sentence Level Models

5.1.1. Data Preprocessing at Sentence Level

Run the following command on sentences_just_one_position.csv to generate the required train-dev-test sets.

cd 5_model_works/sentence_level
python executables/s_data_preprocess.py
5.1.2. Model Training at Sentence Level

The following command can be used to train sentence-level model:

cd 5_model_works/sentence_level
python executables/s_train.py \
--max_seq_length 100 \
--batch_size 100 \
--traindata_file train.csv \
--devdata_file val.csv \
--testdata_file test.csv \
--model bert-base-cased \
--num_labels 2 \
--learning_rate 1e-5 \
--outdir_model model/ \
--save_model true \
--finetuned false \
--epochs 10 \
--patience 3 \
--task train
5.1.3. Model Prediction at Sentence Level

Use the following command to make predictions using the optimized model on the unlabeled reviews data file, all_reviews_by_sentences.csv

cd 5_model_works/sentence_level
python executables/s_predict.py

5.2. Token Level Models

5.2.1. Data Pre-processing at Token Level

Before training the models, the data folders have to be prepared using the command

cd 5_model_works/token_level
python executables/t_data_preprocess.py

Ensure that before running the command, you copy "sentence_just_one_position.csv" and "text_sentence_majority.csv" from /4_post_processing/data and save to the raw_data_folder. and you download "AURC_DATA.tsv" and "AURC_DOMAIN_SPLITS.tsv from https://github.com/trtm/AURC/tree/master/data and save them to the raw_data_folder.

5.2.2. Model Training at Token Level:

The following command can be used to train token-level model:

SEED=0
cd 5_model_works/token_level
python executables/t_modelprocessing.py \
--data_dir=data/amsr/train20 \
--models=bert-base-cased \
--task_name=classify \
--output_dir=models/classify_topic/bert-base-cased/rev20_classify_topic${SEED} \
--use_topic \
--do_train \
--use_weight \
--max_seq_length 85 \
--do_eval \
--seed ${SEED}
5.2.3. Model Prediction at Token Level:

t_predict.py script is used to generate predictions on the unlabeled data file, all_reviews_by_sentences.csv.

6 Argument Extraction

The following command can be used to extract the arguments from all reviews:

cd 6_arg_extraction
python executables/extract_topk.py

The paths for the input data and output data have been defined within the script. The same script can be used to generate k% extracted arguments at sentence level and at token level.

7 Paper Acceptance Model using ToBERT

First, we need to fine-tune a BERT model to help generate hidden states of the extracted arguments.

Step 1: Data Preparation for the BERT model:
cd 7_paper_acceptance_model
python executables/p_data_preprocess.py
Step 2: Train the BERT model:
cd 7_paper_acceptance_model 
python executables/p_modelprocessing.py \ 
--data_dir=data/k_50/ \ 
--models=bert-base-cased \
--output_dir=paper_models/k_50 \
--do_train \
--do_eval \
--seed 1 \
--num_train_epochs 10
Step 3: Train the ToBERT Model:

The following command can be used to train the ToBERT model next.

cd 7_paper_acceptance_model
python executables/ToBERT.py

The parameters and paths to the data and model have to be defined in the code directly.