This reposirtory provides the implementation of the supervised catchphrase extraction model D2V-BiGRU-CRF described in the paper titled - "A Sequence Labeling Model for Catchphrase Identification from Legal Case Documents". The repository provides python codes for training a new model from a scratch and using the same for extracting catchphrases from a new unseen document. In addition we provide a pre-trained model that was trained using our data that can be readily used to extract catchphrases from unseen documents. We describe the usage of the python scripts.
Regardless of whether we want to train a new model from scratch or extract catchphrases using the trained-model, we need a pre-trained Doc2Vec model. We provide our doc2vec model (trained using gensim upon a set of 33.5K case documents from the Supreme Court of India ) that can be downloaded from the link: https://app.box.com/s/sd3v6kp1i2qtsz8r2i2dfuwvx43ri1hb. We hope to make using our model easier. One should be primarily interested in the following two scripts -
- train_on_gold_standard_catches.py for training the model and
- annotate_docs.py for predicting the catchphrases out of new unseen documents.
All other options can be provided inside the scripts themselves. And the meaning of the variables are explained within the code using appropriate commentlines. Thank You.
Thank you for using this implementation in your work, please cite our original paper: "A Sequence Labeling Model for Catchphrase Identification from Legal Case Documents", A. Mandal, K. Ghosh, S. Ghosh, S. Mandal, 2021, Journal of Artificial Intelligence and Law.