word-vectors 1.3.0

Word Vectors

Word Vectors

A fast light-weight library for loading (and saving) word vectors.

Reading

The default way to read in word vectors is to use read. This function will sniff the file to try to determine what kind of file it is. If this sniffing fails for some reason an exception is raised. Don't worry you can still read your vectors you just need to use the read function specific to the vector file format you are using.

File Types

There are a few common types of word vectors file formats used in the NLP community. The supported formats are described here.

GloVe

A simple vector file that is a plain text file. Each line is a word followed by the vectors. Each line has the word and the elements of the vectors separated by a space. This is both slow and space inefficient.

This can be read with the read_glove function.

Word2Vec

Text

A text format this is the same as the GloVe format except the first line is two numbers, the first number is the number of elements in the vocabulary and the second is the size of the vectors. These numbers are not very helpfully because often some of these vector files have the same word at multiple lines so pre-allocating your vectors based on these numbers doesn't really work. Like GloVe this is both slow and space inefficient.

This can be read with the read_w2v_text function.

Binary

A simple binary format where the first row is the number of items in the vocab and the size of the vectors. Each line after is a word followed by the vector as a binary string separated by a space. This format is compact but slow because you need to read a byte at a time to the find the end of each word.

This can be read with the read_w2v function.

Note: The popular fastText pretrained word vectors ship in both the text and binary formats used by word2vec.

Dense

This is my new format. It is a binary file where the first 12 bytes are the vocab size, vector size, and max length of a word as unsigned, little endian, ints. Then the words and vectors follow with the words padded to the max length and then the vector. This format is a little larger than the word2vec format but it is faster because the location of each item (both the words or the vectors) can be calculated quickly. It also allows the possibility of multithreaded reading. This format is smaller than the normal glove format.

This can be read with the read_dense function.

Writing

Each format has its own writing function what takes in the destination file name, the vocab, and the vectors. The available writers are the following:

• write_glove
• write_w2v_text
• write_w2v
• write_dense