syuoni / text-classification-theano

RNN, LSTM and CNN for text sentiment classification

text sentiment classification

The project homework for "Foundations of Machine Learning".

This project implements RNN, LSTM and CNN for text sentiment classification, based on Theano (0.8.2).
All models accept pre-trained word embedding inputs.

dataset

We use Large Movie Review Dataset (imdb) dataset as default dataset, and process it according to theano-tutorial.  

You can change the training data by yourself, and refer the imdb data processing script (imdb-corpus-prepare-script.py) for data processing details. Alternatively, you can use the Corpus.build_corpus_with_dic method provided by corpus-module to build corpus and dictionary simultaneously.

test error rates of models (imdb dataset)

model	pooling	ensemble CV (soft voting)	ensemble CV (hard voting)	pre-trained word embedding	randomly initialized word embedding
LSTM	max	10.10%	10.07%	10.45%	11.46%
LSTM	mean	9.54%	9.65%	9.70%	10.83%
RNN	max	N/A	N/A	12.32%	13.28%
RNN	mean	N/A	N/A	11.49%	11.58%
CNN	max	9.52%	9.62%	10.63%	10.55%
CNN	mean	11.17%	11.25%	11.90%	11.98%

text-classification-theano's Issues

你好，就文本情感分类的问题，想咨询一下

看到你是用RNN，LSTM，CNN进行的情感分类，但有些问题想要请教一下：对什么文本进行的情感分类，对于文本是怎么处理的，还有各种方法的分类精度如何，想做情感分类方面的研究，还请指导一下