Tools for converting lexical semantic representations

This project contains software written for my MSc thesis "Word embedding models as graphs: conversion and evaluation". The purpose of this software is to convert between graph and vector lexical semantic representations, and to evaluate the results.

Evaluation

For evaluation, we use gensim with gold standard data files from the wordvectors.org evaluation suite. These are found at https://github.com/mfaruqui/eval-word-vectors. Three of these data files use spaces as field separators. This must be changed to tabs, to work with gensim.

Word embedding models

Our software has been tested with word embeddings from http://vectors.nlpl.eu/repository/. We have mainly used model 11.

convert.py

Lexical semantic model conversion script. Takes as input a model in either word2vec or graphml format. Outputs the model converted to the opposite format.

evaluate.py

Script for evaluating the conversion results. Takes as parameter a word2vec word embedding model. This model is converted to a graph model, and then back to a word embedding model. The resulting word embeddings are evaluated with SimLex-999. The script expects to find the gold standard data files from https://github.com/mfaruqui/eval-word-vectors in the directory word-sim.

Example command line, using the threshold method:

python3 evaluate.py Gigaword/skip-gram.txt Threshold -s -max 10000

This requires an unzipped word embedding model at the given path. The evaluation results are stored in a text file.

evaluation-suite.py

Script for evaluating the conversion using all data sets from the wordvectors.org evaluation suite. Example command, evaluating the kNN method with data sets in the directory word-sim:

python3 evaluation-suite.py kNN word-sim/ -s

With large models, this might take some hours to complete. The results are stored in a text file with a name like suite-kNN-400000.txt.

SemEval-2013.py

Evaluation script for performing and evaluating word sense induction based on SemEval-2013 task 11. Can be used with either a word embedding model or a graph model. Example command line: python3 SemEval-2013.py Wikipedia/wikipedia_merged_w2v.txt.gz 20 56 -n 10000 This runs the evaluation with thresholds from 0.20 to 0.56, limiting the model to the first 10,000 word

WSI.py

Contains methods for doing word sense induction based on the HyperLex or spinglass graph clustering algorithms.

centrality.py

Script for evaluating converted graphs by performing hypernym discovery in WordNet graphs.

python centrality.py gigaword/skip-gram.txt

visualize.py

Helper script for plotting visualizations of graph models.

graph_tools.py

Some functions for working with graphs