Vector space models.

• Replication of the original c version.

• Preserves long strings.

• Words are deliminated by ' ' and '\n'.

• Multiple training files

• Preprocess data where each word is delimited by ' ' and sentence is delimited by '\n'.

#Word2Vec.java Usage #####Enviornment Details

• Make sure not to run with the 32bit JDK, it limits heap size.
• Other Java implementations of word2vec reccomend around 10g heap size. This implementation needs more than that (details to come).
  To set these options from maven, use:

set MAVENOPTS=-Xmx<NUM_GIGS>g -d64

where NUM_GIGS is the number of gigabytes you want to allocate for the heap.

#####Command Line parameters, (each prepended with a tag '-'), most take additional arguments

• train: "path to the training file or the directory containig the training files.
• output: "output file to save the resulting word vectors.
• ext: "extension of the training files (default: "*").
• size: "size of word vectors (default: 100).
• window: "max-window of contextual words (default: 5).
• sample: "threshold for occurrence of words (default: 1e-3). Those that appear with higher frequency in the training data will be randomly down-sampled.
• negative: "number of negative examples (default: 5; common values are 3 - 10). If negative = 0, use Hierarchical Softmax instead of Negative Sampling.
• threads: "number of threads (default: 12).
• iter: "number of training iterations (default: 5).
• min-count: "min-count of words (default: 5). This will discard words that appear less than times.
• alpha: "initial learning rate (default: 0.025 for skip-gram; use 0.05 for CBOW).
• binary: "If set, save the resulting vectors in binary moded.
• cbow: "If set, use the continuous bag-of-words model instead of the skip-gram model.