Skip to content

Latest commit

 

History

History
68 lines (36 loc) · 2.69 KB

README.md

File metadata and controls

68 lines (36 loc) · 2.69 KB
Raw

CLiQS-CM

Cross-Lingual Query-Based Summarization of Crisis-Related Social Media: An Abstractive Approach Using Transformers

Data

Folder data includes tweets collected during five events and reports provided by ERCC. Every file in folder tweets (inside event directories) contains columns with tweet ids, labels (informative crisis [1] /non-crisis [0] and in some files non-informative crisis [2]), and categories ([1] - tweet labeled as category related). Folders inside reports include crisis-related .pdf files downloaded from ERCC Echo Flash and text reports used as ground truth during summaries evaluation.

Additionally, folder data includes file queries.json with queries for every dataset and category.

Features extraction

For extraction features needed for reproducing the model, you could use file features_extraction.py with arguments datafile, dataset name, category, and language (ISO 639-1 code).

python3 features_extraction.py 'taal_out_en.csv' 'taal' 'Danger' 'en'

The output files with text features, similarity features, and sentence embeddings vectors will be stored in the same directory.

Classification

For training and evaluating the classification model, you could execute classification.py script with the arguments dataset name and test language (ISO 639-1 code of the test language. The algorithm will train on all other languages allowed for the event).

python3 classification.py 'australia' 'en'

Also, it is necessary to have three files in the same directory for every language in the dataset prepared using feature_extraction.py. For Taal volcano eruption set in English it would be next files: taal_en.csv, taal_en_laser_features.csv, taal_en_text_features.to_csv.

Summarization

For executing summarization script necessary four files for every language in dataset prepared with use of feature_extraction.py. Please, don't forget files with query-similatities features. Also it necessary to have translation of tweets from languages to English (translated texts should be stored in column 'en_texts')

python3 sum.py 'fukushima' 'Casualties'

Instllation

Python 3.7.2

Paper

The pre-print of the paper on ArXiv.

If you use this code in your research, please kindly cite our paper:

@article{vitiugin2022cross,
  title={Cross-Lingual Query-Based Summarization of Crisis-Related Social Media: An Abstractive Approach Using Transformers},
  author={Vitiugin, Fedor and Castillo, Carlos},
  journal={arXiv preprint arXiv:2204.10230},
  year={2022}
}