GloVe: Global Vectors for Word Representation _____________________________________________________ We provide an implementation of the GloVe model for learning word representations, as well as some auxiliary tools to construct word-word cooccurrence matrices from large corpora. Please see the project page for more information: http://nlp.stanford.edu/projects/glove/ This package includes four main tools: 1) vocab_count Constructs unigram counts from a corpus, and optionally thresholds the resulting vocabulary based on total vocabulary size or minimum frequency count. This file should already consist of whitespace-separated tokens. Use something like the Stanford Tokenizer (http://nlp.stanford.edu/software/tokenizer.shtml) first on raw text. 2) cooccur Constructs word-word cooccurrence statistics from a corpus. The user should supply a vocabulary file, as produced by 'vocab_count', and may specify a variety of parameters, as described by running './cooccur'. 3) shuffle Shuffles the binary file of cooccurrence statistics produced by 'cooccur'. For large files, the file is automatically split into chunks, each of which is shuffled and stored on disk before being merged and shuffled togther. The user may specify a number of parameters, as described by running './shuffle'. 4) glove Train the GloVe model on the specified cooccurrence data, which typically will be the output of the 'shuffle' tool. The user should supply a vocabulary file, as given by 'vocab_count', and may specify a number of other parameters, which are described by running './glove'. The package also provides a demo script 'demo.sh'. It downloads a small corpus, consisting of the first 100M characters of Wikipedia, collects unigram counts, constructs and shuffles cooccurrence data, and trains a simple version of the GloVe model. It also runs a word analogy evaluation script in python. All work contained in this package is licensed under the Apache License, Version 2.0. See the include LICENSE file.