Comparative Study of CNN and RNN for Natural Language Processing (Yin et al, 2017)*

Overview:

• Objectively compares the performance of CNN vs. GRU vs. LSTM in different tasks of NLP

  • Sentiment classification

  • Relation classification

  • Textual entailment

  • Answer selection

  • Question relation match

  • Path query answering

  • POS tagging

• Trains each model from scratch without no extra knowledge (e.g. no pre-trained word embeddings)

• Uses the most basic architecture and optimization method

Key Results:

• As our intuition, RNNs are more suitable for encoding order information and long range context depedency

• Surprisingly, GRUs can outperform CNN in text classification and semantic matching if sentiment/semantic are determined by long range semantic dependency.

Comments:

• One of key issues of deep learning papers = Not being able to replicate the results

• This paper is one of the first papers to compare the performance of CNN and RNN in equal and objective experiment setting.