Стуктура проекта:

1. Utility files
   • models.py definition of pytorch models: lstm и cnn.
   • params.py constants used by all notebooks
   • training_utils.py utility functions to facilitate sklearn-like training and cross-validation
2. Notebooks
   • [notebooks/label_hierarchies] analysis of category hierarchies .
   • notebooks/process_data data preprocessing
   • notebooks/fastText fastText based models
   • notebooks/tfidf linear model with td-idf features
   • notebooks/rnn_cross_val lstm-based models
   • notebooks/cross_val_сnn cnn-based models
   • notebooks/validation validation of selected models

Pretrained fastText embeddings from [deeppavlov.ai]:

whet -O embeddings/ft_native_300_ru_wiki_lenta_lemmatize.vec http://files.deeppavlov.ai/embeddings/ft_native_300_ru_wiki_lenta_lemmatize/ft_native_300_ru_wiki_lenta_lemmatize.vec

Results

Accuracy on validation dataset on diferent hierarchy levels:

• lstm:

1. 0.9636787056708612
2. 0.9450686386664487
3. 0.8911586860598137
4. 0.8873079751593398

• tf-idf+LinearSVM:

1. 0.9655478836411179
2. 0.9490419186141527
3. 0.8969807158032358
4. 0.8933751429972218