mssammon / ewise

ACL 2019: Zero-shot Word Sense Disambiguation using Sense Definition Embedding

EWISE

Reference code for ACL2019 paper "Zero-shot Word Sense Disambiguation using Sense Definition Embeddings". EWISE[1] (Extended WSD Incorporating Sense Embeddings) is a principled framework to learn from a combination of sense-annotated data, dictionary definitions and lexical knowledge bases.

We have used the WSD evalauation framework[2] for training and evaluation.

Dependencies

The code was written with, or depends on:

• Python 2
• Pytorch 0.4.0
• NLTK 3.2.5
• WSD evalauation framework[2]

Running the code

As of now, the code requires a working directory with the following files:

• i_id_to_i_token : dictionary (pickle file) -- input token index to input token
• i_id_embedding : numpy array (pickle file) -- input embedding matrix
• o_id_to_o_token : dictionary (pickle file) -- output token index to output token (for train tokens; output tokens indices include and begin with input token indices)
• o_id_remainingWordNet_to_o_token : dictionary (pickle file) -- output token index to output token (for output tokens outsode the training set, obtained from WordNet)
• i_id_to_candidate_wn_o_id : dictionary (pickle file) -- input token index to a list of candidate output token indices (obtained from WordNet)
• i_id_to_candidate_train_o_id : dictionary (pickle file) -- input token index to a list of candidate output token indices (obtained from training set)
• {semcor, semeval2007, senseval2, senseval3, semeval2013, semeval2015, ALL}_indexed.json : list (pickle file) -- list of indexed train/test examples (a dictionary for each sentence) created from the WSD evalauation framework[2] with the following fields:
  • original : input token indices
  • annotated : output token indices
  • offsets : offsets in original/annotated for annotated tokens
  • pos : POS tags for annotated words (equal in length to offsets)
  • stems : Stems for annotated words (equal in length to offsets)
  • doc_offset : Tag used for making predictions for words in the sentence

Additionally, depending on the encoding method used, it requires:

• $output_embeddings : numpy array (npz) -- output embedding matrix

Finally, the Scorer script from the WSD evalauation framework[2] needs to be present on the machine. The directory path to the Socrer needs to be provided.

To test an existing model:

python wsd_main.py --cuda --input_directory $working_directory --scorer $path_to_scorer --output_embedding customnpz-$output_embeddings --train semcor --val semeval2007 --predict_on_unseen --evaluate --pretrained $model_path

To train a model:

python wsd_main.py --cuda --dropout 0.5 --epochs $num_epochs --input_directory $working_directory    --scorer ./ --output_embedding customnpz-$output_embeddings  --train semcor --val semeval2007 --lr 0.0001 --predict_on_unseen  --save $save_path

Citation

If you use this code, please consider citing:

[1] Sawan Kumar, Sharmistha Jat, Karan Saxena and Partha Talukdar. 2019. Zero-shot Word Sense Disambiguation using Sense Definition Embeddings. Accepted to the 57th Annual Meeting of the Association for Computational Linguistics (ACL) (proceedings in press).

References

[2] Alessandro Raganato, Jose Camacho-Collados, and Roberto Navigli. 2017. Word sense disambiguation: A unified evaluation framework and empirical comparison. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 99–110, Valencia, Spain. Association for Computational Linguistics.

Contact

For any clarification, comments, or suggestions please create an issue or contact [email protected]