- python >= 3.6
- keras >= 2.2
- tensorflow >= 1.8
- h5py
- scikit-learn
- lxml
- nltk
ROUGE-1.5.5
Please place ROUGE-1.5.5 root folder (containing ROUGE-1.5.5.pl script) next to encsum_rtv folder.
svm_rank
Please add svm_rank root folder (containing svm_rank_learn and svm_rank_classify) to $PATH.
- Training EncSum Model
- COLIEE 2018 Task 1
- SUM-HOLJ Summarization Task
path/to/coliee2018/train/corpus/dir, path/to/coliee2018/test/corpus/dir: COLIEE 2018 train, test folders containing IR.xml file.
path/to/glove/embedding/file: path to text-format word embeddings file glove.840B.300d.txt.
- Preprocessing
python -m encsum_rtv.preprocess_coliee2018 \
--mode train \
--input-dir path/to/coliee2018/training/corpus/dir \
--output-dir data/coliee2018/preprocessed/train \
--embeddings-file path/to/glove/embedding/file - Training Neural Net
python -m encsum_rtv.encsum encsum_nn \
--train \
--data-dir data/coliee2018/preprocessed/train/numeric \
--model-dir model \
--mini-epoch-factor 50 \
--nb-epochs 5 path/to/coliee2018/train/corpus/dir, path/to/coliee2018/test/corpus/dir: COLIEE 2018 train, test folders containing IR.xml file.
- Preprocessing
python -m encsum_rtv.preprocess_coliee2018 \
--mode test \
--input-dir path/to/coliee2018/test/corpus/dir \
--output-dir data/coliee2018/preprocessed/test \
--vocab-file data/coliee2018/preprocessed/train/numeric/emb_vocab.json - Generating EncSum representation
Train:
python -m encsum_rtv.encsum encsum_nn \
--infer-encsum \
--data-dir data/coliee2018/preprocessed/train/numeric \
--model-dir model \
--output-dir data/coliee2018/preprocessed/train/encsum
python -m encsum_rtv.encsum encsum_relevance \
--encsum-feature-file data/coliee2018/preprocessed/train/encsum/paras.encsum.npz \
--output-file data/coliee2018/preprocessed/train/encsum/encsum_relevance.npz \
--meta-file path/to/coliee2018/train/corpus/dir/IR.xmlTest:
python -m encsum_rtv.encsum encsum_nn \
--infer-encsum \
--data-dir data/coliee2018/preprocessed/test/numeric \
--model-dir model \
--output-dir data/coliee2018/preprocessed/test/encsum
python -m encsum_rtv.encsum encsum_relevance \
--encsum-feature-file data/coliee2018/preprocessed/test/encsum/paras.encsum.npz \
--output-file data/coliee2018/preprocessed/test/encsum/encsum_relevance.npz \
--meta-file path/to/coliee2018/test/corpus/dir/IR.xml- Extracting lexical features:
Train
python -m encsum_rtv.lexical \
--text-dir data/coliee2018/preprocessed/train/text \
--numeric-dir data/coliee2018/preprocessed/train/numeric \
--meta-file path/to/coliee2018/train/corpus/dir/IR.xml \
--cpu-count $CPU_COUNTTest
python -m encsum_rtv.lexical \
--text-dir data/coliee2018/preprocessed/test/text \
--numeric-dir data/coliee2018/preprocessed/test/numeric \
--meta-file path/to/coliee2018/test/corpus/dir/IR.xml \
--cpu-count $CPU_COUNT--cpu-count $CPU_COUNT: accelerating with multiprocessing.
- Train and Evaluate Ranker
python -m encsum_rtv.l2r train \
--model-file model/svm_rank.dat \
--feature-files data/coliee2018/preprocessed/train/encsum/encsum_relevance.npz \
data/coliee2018/preprocessed/train/numeric/rouge*.npz \
--gold-file data/coliee2018/preprocessed/train/numeric/relevance.npz
python -m encsum_rtv.l2r predict \
--model-file model/svm_rank.dat \
--feature-files data/coliee2018/preprocessed/test/encsum/encsum_relevance.npz \
data/coliee2018/preprocessed/test/numeric/rouge*.npz \
--output-file-prefix output/coliee2018/test \
--select-top 10 \
--meta-file path/to/coliee2018/test/corpus/dir/IR.xml
python -m encsum_rtv.l2r evaluate \
--submission-file output/coliee2018/test.submission.txt \
--gold-file path/to/task1_true_answer/file- Preprocessing
python -m encsum_rtv.preprocess_holj \
path/to/raw/corpus/dir \
data/holj/preprocessed- Generating Summary
python -m encsum_rtv.generate_summary \
--corpus-dir data/holj/preprocessed \
--output-dir output/holj \
--config-file model/score_model.yaml \
--weight-file model/model_weights.hdf5 \
--vocab-file model/emb_vocab.json \
--summary-mode sentence_selection \
--top-anchor 0.10--top-anchor 0.10: 10% of #sentences as summary.
- Evaluation
python -m encsum_rtv.evaluate_summary \
--corpus-dir data/holj/preprocessed \
--predict-dir output/holj- Preprocessing
python -m encsum_rtv.preprocess_docs \
--corpus-dir path/to/corpus/dir \
--embedings-file path/to/glove/embedding/file \
--output-data data/other/task/preprocessed
path/to/corpus/dir contains documents with filename format (docname.sentences, docname.summary), for example:
doc1.sentences
doc1.summary
doc2.sentences
doc2.summary
...
- Train Neural Net
python -m encsum_rtv.encsum encsum_nn \
--train \
--doctypes "{'summary': 'summary', 'sentences': 'sentences'}"
--data-dir data/other/task/preprocessed \
--model-dir model \
--mini-epoch-factor 50 \
--nb-epochs 5 - Generating Summary
python -m encsum_rtv.generate_summary \
--corpus-dir path/to/corpus/dir \
--output-dir output/summary/dir \
--config-file model/score_model.yaml \
--weight-file model/model_weights.hdf5 \
--vocab-file model/emb_vocab.json \
--summary-mode sentence_selection \
--top-anchor 0.10--top-anchor 0.10: 10% of #sentences as summary.
- Evaluation
python -m encsum_rtv.evaluate_summary \
--corpus-dir path/to/corpus/dir \
--predict-dir output/summary/dirThis code is an implementation of this paper.
Tran, V., Le Nguyen, M., Tojo, S. et al. Encoded summarization: summarizing documents into continuous vector space for legal case retrieval. Artif Intell Law (2020). https://doi.org/10.1007/s10506-020-09262-4