This code was used to perform the lexical substitution evaluation described in the following papers:

[1] A Simple Word Embedding Model for Lexical Substitution Oren Melamud, Omer Levy, Ido Dagan. Workshop on Vector Space Modeling for NLP (VSM), 2015 [pdf].

[2] context2vec: Learning Generic Context Embedding with Bidirectional LSTM
Oren Melamud, Jacob Goldberger, Ido Dagan. CoNLL, 2016 [pdf].

• Python 2.7
• NLTK 3.0) - optional (only required for the AWE baseline and MSCC evaluation)
• Numpy
• context2vec - for the context2vec evaluation

This repository contains preprocessed data files based on the datasets introduced by the following papers:

[3] Semeval-2007 task 10: English lexical substitution task Diana McCarthy, Roberto Navigli, SemEval 2007.
(files with the prefix 'lst' under the 'dataset' directory)

[4] What substitutes tell us-analysis of an ”all-words” lexical substitution corpus. Gerhard Kremer,Katrin Erk, Sebastian Pado, Stefan Thater. EACL, 2014.
(files with the prefix 'coinco' under the 'dataset' directory)

• Download the word embeddings, context embeddings from [here]
• Preprocess the embedding files:

python jcs/text2numpy.py <word-embeddings-filename> <word-embeddings-filename>
python jcs/text2numpy.py <context-embeddings-filename> <context-embeddings-filename>

• To perform the lexical substitution evaluation run (replace the example datasets files and params below as you wish):

python jcs/jcs_main.py --inferrer emb -vocabfile datasets/ukwac.vocab.lower.min100 -testfile datasets/lst_all.preprocessed -testfileconll datasets/lst_all.conll -candidatesfile datasets/lst.gold.candidates -embeddingpath <word-embeddings-filename> -embeddingpathc <context-embeddings-filename> -contextmath mult --debug -resultsfile <result-file>

• This will create the following output files:
  • <result-file>
  • <result-file>.ranked
  • <result-file>.generate.oot
  • <result-file>.generate.best
• Run the following to compute the candidate ranking GAP score. The results will be written to <gap-score-file>.

python jcs/evaluation/lst/lst_gap.py ~/datasets/lst_all.gold <result-file>.ranked <gap-score-file> no-mwe

• Run the following to compute the OOT and BEST substitute prediction scores. The results will be written to <xxx-score-file>. score.pl was distributed in [3].

perl dataset/score.pl \<result-file\>.generate.oot datasets/lst_all.gold -t oot > \<oot-score-file\>

perl dataset/score.pl \<result-file\>.generate.best datasets/lst_all.gold -t best > \<best-score-file\>

• See context2vec for how to download or train a <context2vec-model>.
• To perform the lexical substitution evaluation run (replace the example datasets files and params below as you wish):

python jcs/jcs_main.py --inferrer lstm -lstm_config \<context2vec-model\>.params -testfile datasets/lst_all.preprocessed -testfileconll datasets/lst_all.conll -candidatesfile datasets/lst.gold.candidates -contextmath mult -resultsfile <result-file> --ignoretarget --debug

• From here, follow the same instructions as in the previous section.

Apache 2.0