Naver Movie Sentiment Classification

Data

All datas are from https://github.com/e9t/nsmc

Goal

Sentiment Classification: Classify Good/Bad from movie reviews

• Task: Many to One
• Use: bidirection LSTM + Self Attention + Fully Connected MLP

A STRUCTURED SELF-ATTENTIVE SENTENCE EMBEDDING: https://arxiv.org/pdf/1703.03130.pdf

Blog Explanation in Korean

Link : simonjisu.github.io

Notebooks

Date	Model	Accuracy	Link
180402	self_attn_1H_r5	0.7208	link
180402	self_attn_1H_r20	0.8460	link
180402	self_attn_3H_r5	0.8498	link

• r: n_hops of self attention
• H: number of hidden layers
• model3: self_attn_3H_r5 is not quite good at explanation. It may be overfitted, because of under reasons.
  1. It classifies labels by only using first parts and last parts of a sentences
  2. As layers go deeper, it learns from previous hidden layers just for guessing the right label and ignore whaterver the word is.

Get a Review Visualization

0. Red Blocks

Red Blocks means which words the embedding takes into account a lot, and which ones are skipped by the embedding.

1. Help

TYPE "-h" or "-help" behind "visualize

INSERT ARGUMENTS behind "visualize.py"

• First
  • [-1] model1: 1 hidden layer, r=5
  • [-2] model2: 1 hidden layer, r=20
  • [-3] model3: 3 hidden layer, r=5
• Second
  • [-sample_idx] number from 0 to 781
• Third
  • [-(file_path.html)] file path, it is an optional, default is "./figures/(file_name)[sample_idx].html"

2. Example

python3 visualize.py -2 715

Example: Sample number 715