Athena is a word embedding program based on the original paper Efficient Estimation of Word Representations in Vector Space published by Tomas Mikolov in January 2013.
The application will generate embeddings using either the Continuous Bag of Words (CBOW) or Skip Gram with Negative Sampling (SGNS) models.
This is a GPU accelerated C# implementation, which provides a full environment to manage a large text corpus and subsequently learn and query word embeddings. To the best of my knowledge, this is the only implementation of word2vec on a GPU using C#.
Athena works with either CUDA or OpenCL devices and will auto-select depending on your hardware. A reference to Cudafy.NET is required, which can be downloaded from GitHub.
To get started, select the Load [L] option which will load your corpus file located in the data folder – I use a 6GB full text dump of Wikipedia - this file must be called corpus.txt.
Athena then cleans up the corpus file, transforming text to lower case, standardising diacritics and converting numerics to tokens - the resulting text will be saved as corpus_0.txt.
Next, Athena will identify common recurring bigrams and concatenate these into phrases – the result will be saved as corpus_1.txt - this process is heuristic and may therefore omit some bigrams. To guarantee the occurrence of particular bigrams in your corpus it is possible to force addition - create a bigrams.txt file and populate each row with a required bigram in the form of hillary_clinton, donald_trump, etc.
The proceeding steps, which will take several hours to run, need only be executed once – the first round of training will now commence automatically.
The word embedding model will be stored in model.bin.
Once an intial round has completed, you may wish to tweak the hyperparameters and try again. Do this by selecting the Retrain [R] option.
To query the model, simply select the Query [Q] option and type in a word or phrase.
For example, typing in the word paris will return cities similar to Paris:
? paris Nearest ------------------- 1.00 paris 0.81 brussels 0.73 vienna 0.71 munich 0.71 marseilles 0.70 turin 0.69 strasbourg 0.69 madrid 0.69 geneva 0.68 london
It is also perform vector subtraction by appending a colon to the end of the word to negate. For example paris: france oslo is the equivalent of asking Athena Paris is to France as Oslo is to...? – this should return Norway as the top answer:
? paris: france oslo Nearest ------------------- 0.87 norway 0.78 finland 0.75 denmark 0.75 iceland 0.74 sweden 0.73 finnmark 0.72 trondheim 0.71 tromso 0.71 estonia 0.71 oslo
To carry out an exhaustive test of the model, select the Test [T] option. This will load analogies from the test.txt file. These should be in the form of athens greece baghdad iraq.
Follow me on Twitter for more updates like this.