This is simple text classifier library that use algorythm like next:

• use pretrained Word2Vec vectors (each word presented as word_POS, so to get part of speech - I use pymorphy2)
• use clusters that pre-extracted from Word2Vec vectors. So - many cluster will present words with similar vectors. E.g. "globally" in demo script I use 2000 clusters.
• to present text :
  • tokenize it
  • for each token:
    • get word2vec vector
    • get cosine similarity between vector and cluster vectors
    • set similarities to zero when it less then threshold
  • get sum of individual token vectors
• classifier training process is like next :
  • for each label :
    • choose text vectors for text with this label
    • get mean vector
    • get clusters with highter similarity
  • choose clusters
  • use only selected cluster similarity as features
  • normalize features
  • initialize KNearestNeighbours classifier
• classification process is like next :
  • convert text to vectors (sometimes pymorphy2 can't detect POS, so where multiple parsing versions)
  • for each vector
    • get nearest neighbours distance and indices
    • get labels attached to nearest neighbours
    • if distance is bigger then threshold - set labels for this neighbour to zero
    • get mean labels value and store prediction
  • return prediction with bigger dispersion

Installation

To use it you'll need previously get:

• pymorphy2 with dictionaries
• gensim
• nltk
• get word2vec model. I tested it on russian and used model from http://rusvectores.org/static/models/ruwikiruscorpora_0_300_20.bin.gz (unpack it and place in classifier/dataset/ruwikiruscorpora_0_300_20.bin)

Demo script

You can see demonstration in classifier notebook.