nlp_project

The final project for CS224n

The attempt to replicate below paper.

Open Source in GitHub

https://github.com/peter6888/nlp_project

Download data

Type the following with your current directory being our repository.

cd data
sh download.sh
unzip finished_files.zip

Trial Run

cd src
sh run_cmd.sh

Train base model from https://github.com/abisee/pointer-generator

Pre-requirement, make sure have pyrouge installed

pip install pyrouge

Train

Remember to change the name of the experiment and log_root (this is the directory where logs will be saved).

python run_summarization.py --mode=train --data_path=../data/finished_files/chunked/train_*.bin --vocab_path=../data/finished_files/vocab --log_root=/home/stonepeter/log --exp_name=baseline

Validate

Remember to change the name of the experiment and log_root (this is the directory where logs will be saved).

python run_summarization.py --mode=eval --data_path=../data/finished_files/chunked/val_* --vocab_path=../data/finished_files/vocab --log_root=/home/stonepeter/log --exp_name=baseline

Beam Search Validate

Remember to change the name of the experiment and log_root (this is the directory where logs will be saved).

python run_summarization.py --mode=decode --data_path=../data/finished_files/chunked/test_* --vocab_path=../data/finished_files/vocab --log_root=/home/stonepeter/log --exp_name=baseline

Result Example

data/sample_summary.txt

Baseline

Use PyTeaser generate the Summary From CNN https://github.com/xiaoxu193/PyTeaser save the summary.

sumy_eval lex-rank ref_summ.txt --url=https://www.huffingtonpost.com/2013/11/22/twitter-forward-secrecy_n_4326599.html

sample output

Precision: 0.000000
Recall: 0.000000
F-score: 0.000000
Cosine similarity: 0.350201
Cosine similarity (document): 0.670010
Unit overlap: 0.128788
Unit overlap (document): 0.294798
Rouge-1: 0.161616
Rouge-2: 0.000000
Rouge-L (Sentence Level): 0.066107
Rouge-L (Summary Level): 0.039251

sumy_eval lex-rank art2_sum.txt --file=/Users/peli/forgit/nlp_project/data/art2.txt --format=plaintext
Precision: 0.000000
Recall: 0.000000
F-score: 0.000000
Cosine similarity: 0.721001
Cosine similarity (document): 0.900767
Unit overlap: 0.216450
Unit overlap (document): 0.344426
Rouge-1: 0.597403
Rouge-2: 0.461538
Rouge-L (Sentence Level): 0.103124
Rouge-L (Summary Level): 0.004133

To-do List: see issues for detail tasks

Intra-Temporal Attention Decoder on Input Sequence
Intra-Decoder Attention Decoder
Token Generation and Pointer
Sharing Decoder Weights

Group sharing resources

Mendeley https://www.mendeley.com/newsfeed/#group:d48f9375-8628-3922-aa04-ac0f8832ae01

Slack https://cs224nprojectteam.slack.com

Overleaf https://www.overleaf.com/read/hykvfpwrgvsj

Software Versions

pip install tensorflow==1.4.1

python 3.5.2 (or python 3.6.x)

pyrouge

Which is a perl based tool, and use pyrouge to call https://github.com/andersjo/pyrouge

git clone https://github.com/andersjo/pyrouge.git

edit ~/.pyrouge/settings.ini

[pyrouge settings]
home_dir = /home/cs224n-team/notebooks/pyrouge/tools/ROUGE-1.5.5

Maybe want to re-build the .db for some reason

cd /home/cs224n-team/notebooks/pyrouge/tools/ROUGE-1.5.5/data/WordNet-2.0-Exceptions/
perl buildExeptionDB.pl /home/cs224n-team/notebooks/pyrouge/tools/ROUGE-1.5.5/data/WordNet-2.0-Exceptions db /home/cs224n-team/notebooks/pyrouge/tools/ROUGE-1.5.5/data/WordNet-2.0.exc.db