DGL_Recursive Neural Networks_Sentiment Classification

Introduction

This project uses the DGL (Deep Graph Library) package to improve the training speed of Recursive Neural Networks(RvNN), which takes only 4-6 seconds every training epoch on the RTX server.

• References
  

  • [1] Socher R, Lin C C, Manning C, et al. Parsing natural scenes and natural language with recursive neural networks[C] . ICML-11. 2011: 129-136. https://ai.stanford.edu/~ang/papers/icml11ParsingWithRecursiveNeuralNetworks.pdf
  • [2] Socher R, Perelygin A, Wu J, et al. Recursive deep models for semantic compositionality over a sentiment treebank[C]. EMNLP. 2013: 1631-1642. https://nlp.stanford.edu/~socherr/EMNLP2013_RNTN.pdf
  • Tree-LSTM DGL Tutorial: https://docs.dgl.ai/tutorials/models/2_small_graph/3_tree-lstm.html?highlight=treelstm

• Dependencies
  

  • Python >= 3.5
    
  • PyTorch >= 0.4.1
    
  • DGL >= 0.3
    

• DGL(Deep Graph Library)
  

  • Tutorial: https://docs.dgl.ai/tutorials/basics/1_first.html
  • Github: https://github.com/dmlc/dgl

Usage

python train.py

Results

The RvNN model is trained on SST and the accuracy of predicting fine-grained sentiment labels at all phrase lengths(All) or full sentences(Root) is as follows:

	Acc_All	Acc_Root
Paper [2]	79.0	43.0
My Recurrence	75.71	47.96