Txt2Vec
Txt2Vec is a toolkit to represent text by vector. It's based on Google's word2vec project, but with some new features, such incremental training, model vector quantization and so on. For a specified term, phrase or sentence, Txt2vec is able to generate correpsonding vector according its semantics in text. And each dimension of the vector represents a feature.
Txt2Vec is based on neural network for model encoding and cosine distance for terms similarity. Furthermore, Txt2Vec has fixed some issues of word2vec when encoding model in multiple-threading environment.
The following is the introduction about how to use console tool to train and use model. For API parts, I will update it later.
Console tool
Txt2VecConsole tool supports four modes. Run the tool without any options, it will shows usage about modes.
Txt2VecConsole.exe
Txt2VecConsole for Text Distributed Representation
Specify the running mode:
-mode <train/distance/analogy/shrink>
: train model to build vectors for words
: calculating the similarity between two words
: multi-words semantic analogy
: shrink down the size of model
: dump model to text format
: build vector quantization model in text format
Train model
With train mode, you can train a word-vector model from given corpus. Note that, before you train the model, the words in training corpus should be word broken. The following are parameters for training mode
Txt2VecConsole.exe -mode train <...>
Parameters for training:
-trainfile : Use text data from to train the model
-modelfile : Use to save the resulting word vectors / word clusters
-vector-size : Set size of word vectors; default is 200
-window : Set max skip length between words; default is 5
-sample : Set threshold for occurrence of words. Those that appear with higher frequency in the training data will be randomly down-sampled; default is 0 (off), useful value is 1e-5
-threads : the number of threads (default 1)
-min-count : This will discard words that appear less than times; default is 5
-alpha : Set the starting learning rate; default is 0.025
-debug : Set the debug mode (default = 2 = more info during training)
-cbow : Use the continuous bag of words model; default is 0 (skip-gram model)
-vocabfile : Save vocabulary into file
-save-step : Save model after every words processed. it supports K, M and G for larger number
-iter : Run more training iterations (default 5)
-negative : Number of negative examples; default is 5, common value are 3 - 15
-pre-trained-modelfile : Use which is pre-trained-model file
-only-update-corpus-word <0/1> : Use 1 to only update corpus words, 0 to update all words
Example:
Txt2VecConsole.exe -mode train -trainfile corpus.txt -modelfile vector.bin -vocabfile vocab.txt -debug 1 -vector-size 200 -window 5 -min-count 5 -sample 1e-4 -cbow 1 -threads 1 -save-step 100M -negative 15 -iter 5
After the training is finished. The tool will generate three files. vector.bin contains words and vector in binary format, vocab.txt contains all words with their frequency in given training corpus, and vector.bin.syn which is used for incremental model training in future.
Incremental Model Training
After we collected some new corpus and new words, to get these new words' vector or update existing words' vector by new corpus, we need to re-train existing model in incremental model. Here is an example:
Txt2VecConsole.exe -mode train -trainfile corpus_new.txt -modelfile vector_new.bin -vocabfile vocab_new.txt -debug 1 -window 10 -min-count 1 -sample 1e-4 -threads 4 -save-step 100M -alpha 0.1 -cbow 1 -iter 10 -pre-trained-modelfile vector_trained.bin -only-update-corpus-word 1
We have already trained a model "vector_trained.bin" before, currently, we have collected some new corpus named "corpus_new.txt" and new words saved into "vocab_new.txt". The above command line will re-train existing model incrementally, and generate a new model file named "vector_new.bin". To get better result, the "alpha" value should be usually bigger than that in full corpus and vocabulary size training.
Incremental model training is very useful for incremental corpus and new word. In this mode, we are able to generate new words vector aligned with existing words efficiently.
Calculating word similarity
With distance mode, you are able to calculate the similarity between two words. Here are parameters for this mode
Txt2VecConsole.exe -mode distance <...>
Parameters for calculating word similarity
-modelfile : encoded model needs to be loaded
-maxword : the maximum word number in result. Default is 40
After the model is loaded, you can input a word from console and then the tool will return the Top-N most similar words. Here is an example:
Txt2VecConsole.exe -mode distance -modelfile wordvec.bin
Enter word or sentence (EXIT to break):
手串
Word Cosine distance
手串 1
佛珠 0.918781571749997
小叶紫檀 0.897870467450521
手钏 0.868526208199693
菩提子 0.85667693515943
紫檀 0.855529437116288
佛珠手链 0.849541378712106
雕件 0.847901026881494
砗磲 0.842016069107114
小叶檀 0.839194380950776
星月菩提子 0.838186634277951
檀香木 0.837212392914782
沉香木 0.83575322205817
星月菩提 0.83494878072285
黄花梨 0.831824567567293
平安扣 0.831679080640205
紫檀木 0.830029415546653
小叶紫檀手串 0.82838028045219
原籽 0.823008017930358
玉髓 0.820901374489359
手链 0.819296636344601
和田玉 0.818727549748641
绿檀 0.816645224342866
和田碧玉 0.816260936449443
菩提根 0.813263703640439
血珀 0.808663718785608
海南黄花梨 0.808161968264115
天然玛瑙 0.807949867283673
紫檀佛珠 0.80467756598682
包浆 0.804356427867412
石榴石 0.803866760918248
小叶紫檀佛珠 0.803276658873406
沉香手串 0.803169094160751
绿松石 0.802508849725817
玛瑙手镯 0.802038899418962
象牙 0.800001887548549
和田玉籽料 0.799139422922168
牛毛纹 0.798210503136587
岫玉 0.797258011387281
发晶 0.797202309084502
Here is another example in English
Enter word or sentence (EXIT to break):
movie
Word Cosine distance
movie 1
film 0.885520313665965
documentary 0.724328882031764
comedy 0.697010992360393
miniseries 0.652796205327335
drama 0.650902800673727
sitcom 0.644413129252032
thriller 0.641292610371391
flick 0.632148128715013
biopic 0.62797525376413
cartoon 0.614410306167449
song 0.603244724945496
cinema 0.600059867125111
novel 0.593261651356445
book 0.58715568070338
videogame 0.585597784043403
telenovela 0.58538938018449
slasher 0.583017049013312
blockbuster 0.581119239625672
mockumentary 0.579500814540811
musical 0.577718414766068
trilogy 0.574577268735197
bollywood 0.569893531890232
docudrama 0.569531770871584
horror 0.569438628440912
featurette 0.569255166345126
telemovie 0.562765328136319
superhero 0.553818777268347
movies 0.551023619845223
story 0.543996823998135
vampire 0.542279925338312
spoof 0.539076773599035
theatrical 0.535893238188445
episode 0.535772427953206
melodrama 0.53513161843853
storyline 0.534828090692405
anime 0.530182484977031
script 0.528789220176914
comic 0.52776622526002
animated 0.526917090929355
You can get demo package in release section.
Shrink model
Sometimes, the size of encoded model maybe too big, so you can use shrink mode to reduce its size according a given lexical dictionary. In shrink mode, any words with its vectors will be filter out if the word isn't in the given lexical dictionary. Here is the usage
Txt2VecConsole.exe -mode shrink <...>
Parameters for shrinking down model
-modelfile : encoded model for shrinking down
-newmodelfile : shrinked model
-dictfile : lexical dictionary. Any words with its vector isn't in the dictionary will be filter out from the model
Dump model
Binary model format is not friendly for human to investigation, so Txt2VecConsole provides a command to dump binary model into text format. The command line as follows:
Txt2VecConsole.exe -mode dump <...>
Parameters to dump encoded model to text format
-modelfile : encoded binary model needs to be dumped.
-txtfile : dumped model in text format
Build vector quantization model
model vector quantization is another way to reduce model size by converting weights from float type to byte type. Currently, Txt2VecConsole supports model vector quantization in 8bits. That means the model size will be reduced to 1/4 original model size. The command line as follows:
Txt2VecConsole.exe -mode buildvq <...>
Parameters to build vector quantization model
-modelfile : encoded model file for vector quantization
-vqmodelfile : output vector quantized model
Demo Package
In release section, a Txt2VecConsole demo package is provided. It contains Txt2VecConsole.exe source code, binary files and two encoded models. Both of two models are word-to-vector models. One model is for Chinese and the other is for English. Please enjoy it and have fun! :)