HSG_ExSUM_NER (extractive summarization and named entity recognition)

This repository presents and compares HeterSUMGraph and variants doing extractive summarization, named entity recognition or both.

This repository also present the influence of the summary/document ratio on performance.

HeterSUMGraph and variants use GATv2Conv (from torch_geometric).

HeterSUMGraph using GATv2Conv is the best variant of HeterSUMGraph, better than original HeterSUMGraph on NYT50, for more information see: https://github.com/Baragouine/HeterSUMGraph

The dataset is a part of general geography, architecture town planning and geology French wikipedia articles.

Warning: this code uses a French Tokenizer.

Clone project

git clone https://github.com/Baragouine/HSG_ExSUM_NER.git

Enter into the directory

cd HSG_ExSUM_NER

Create environnement

conda create --name HSG_ExSUM_NER python=3.9

Activate environnement

conda activate HSG_ExSUM_NER

Install dependencies

pip install -r requirements.txt

Install nltk data

To install nltk data:

• Open a python console.
• Type import nltk; nltk.download().
• Download all data.
• Close the python console.

Scrap, preprocessing and split articles

preprocessing mean cleaning, labeling, etc. not mean preprocessing before training.

• Run 00-00-scrap_wiki.ipynb to scrap data.
• Run 00-01-raw_dataset_to_preprocessed.ipynb to compute summarization and ner labels.
• Run 00-02-drop_article_without_body.ipynb to drop articles without body.
• Run 00-03-split_preprocessed_dataset_to_25_high_25_low_0.5.ipynb to split the previous dataset to three subsets depending of summary/article ratio (Wikipedia-0.5, Wikipedia-high-25, Wikipedia-low-25).
• Run 00-04-split_wiki_datasets_to_train_val_test.ipynb to split previous datasets to train, val and test set.
• Run python scripts/compute_tfidf_dataset.py -input data/wiki_geo_ratio_sc_0.5.json -output data/wiki_geo_ratio_sc_0.5_dataset_tfidf.json -docs_col_name flat_contents (compute tfidfs for whole dataset).
• Run python scripts/compute_tfidf_sent_dataset.py -input data/wiki_geo_ratio_sc_0.5.json -output data/wiki_geo_ratio_sc_0.5_sent_tfidf.json -docs_col_name flat_contents (compute tfidfs for each document).
• Run python scripts/compute_tfidf_dataset.py -input data/wiki_geo_low_25.json -output data/wiki_geo_low_25_dataset_tfidf.json -docs_col_name flat_contents (compute tfidfs for whole dataset).
• Run python scripts/compute_tfidf_sent_dataset.py -input data/wiki_geo_low_25.json -output data/wiki_geo_low_25_sent_tfidf.json -docs_col_name flat_contents (compute tfidfs for each document).
• Run python scripts/compute_tfidf_dataset.py -input data/wiki_geo_high_25.json -output data/wiki_geo_high_25_dataset_tfidf.json -docs_col_name flat_contents (compute tfidfs for whole dataset).
• Run python scripts/compute_tfidf_sent_dataset.py -input data/wiki_geo_high_25.json -output data/wiki_geo_high_25_sent_tfidf.json -docs_col_name flat_contents (compute tfidfs for each document).

tfidfs computing is only necessary for HeterSUMGraph based models.

Embeddings

For training you must use french fasttext embeddings, they must have the following path: data/cc.fr.300.vec

Training

Run one of the *train* notebooks to train and evaluate the associated model: The names of notebooks containing HeterSUMGraph mean that they can be used to train HeterSUMGraph. If the name contains GAT, it means that the notebook trains the original version of HeterSUMGraph. If the name contains GATv2, it means that the GAT layer has been replaced by GATv2. If it contains NER without the "Only", it means that the notebook performs summary and named entity recognition. If it contains OnlyNER, it means that the model only performs named entity recognition; if the name contains POL, it means that edge features are taken into account for the NER; finally, if instead of HeterSUMGraph we have HSGRNN, it means that the model is a combination of HeterSUMGraph and SummaRuNNer.

Result

All XPs scores

see: https://www.overleaf.com/read/gbfxvfvykxsc#77a14f

HeterSUMGraph (GATv2Conv, limited-length ROUGE Recall)

dataset	ROUGE-1	ROUGE-2	ROUGE-L
Wikipedia-0.5	29.1 ± 0.0	8.6 ± 0.0	18.9 ± 0.0
Wikipedia-high-25	23.8 ± 0.0	6.8 ± 0.0	14.9 ± 0.0
Wikipedia-low-25	33.1 ± 0.0	13.3 ± 0.0	22.9 ± 0.0

Other models on Wikipedia-0.5 (wiki_geo_ratio_sc_0.5) (limited-length ROUGE Recall)

model	ROUGE-1	ROUGE-2	ROUGE-L	BCELoss
HeterSUMGraph_GAT	31.11 $\pm$ 0.85	9.79 $\pm$ 0.73	19.59 $\pm$ 0.58	N/A
HeterSUMGraphNER_GAT	31.70 $\pm$ 0.12	10.22 $\pm$ 0.15	20.02 $\pm$ 0.12	0.926+/-0.000
HeterSUMGraphOnlyNER_GAT	N/A	N/A	N/A	0.929+/-0.001
HeterSUMGraphNERPOL_GAT	N/A	N/A	N/A	N/A
HeterSUMGraph_GATv2	31.56 $\pm$ 0.29	10.12 $\pm$ 0.30	19.91 $\pm$ 0.28	N/A
HeterSUMGraphNER_GATv2	31.66 $\pm$ 0.13	10.22 $\pm$ 0.09	20.01 $\pm$ 0.10	0.925+/-0.001
HeterSUMGraphOnlyNER_GATv2	N/A	N/A	N/A	0.930+/-0.001
HSGRNN_GATv2	30.86 $\pm$ 0.00	9.29 $\pm$ 0.00	19.59 $\pm$ 0.00	N/A
HSGRNNNER_GATv2	31.52 $\pm$ 0.10	10.06 $\pm$ 0.09	19.97 $\pm$ 0.05	0.926+/-0.000
HSGRNNOnlyNER_GATv2	N/A	N/A	N/A	0.930+/-0.001

* Wikipedia-0.5: general geography, architecture town planning and geology French wikipedia articles with len(summary)/len(content) <= 0.5.
* Wikipedia-high-25: first 25% of general geography, architecture town planning and geology French wikipedia articles sorted by len(summary)/len(content) descending.
* Wikipedia-low-25: first 25% of general geography, architecture town planning and geology French wikipedia articles sorted by len(summary)/len(content) ascending.