This time, we performed Sentiment Analysis (or Sentiment Classification) of labeled tweet data. ** 2015/10/19 An additional experiment was conducted. ** ** ** 2015/12/19 The source code of SCNN has been released. hogefugabar / CharSCNN-theano ** ** 2015/12/27 Not only SCNN but also CharSCNN implementation has been released. hogefugabar / CharSCNN-theano **
This time, I tried to use the algorithm called CharSCNN described in the paper Deep Convolutional Neural Networks for Sentiment Analysis of Short Texts. For convenience, I used an algorithm called SCNN. This algorithm gives a sentence (Sentence) as input as a series of one-hot expressions of words. CharSCNN gives one-hot expressions to letters in addition to words. If my understanding is correct, SCNN seems to have an architecture close to the following.
From [UNITN: Training Deep Convolutional Neural Network for Twitter Sentiment Classification](http://alt.qcri.org/semeval2015/cdrom/pdf/SemEval079.pdf)
There was satwantrana / CharSCNN on GitHub, so I tried to use it as it is, but I fixed it myself because various codes were strange. ** 2015/12/19 The source code has been released. hogefugabar / CharSCNN-theano Please refer to here. ** **
I implemented it like To Word Embeddings → Convolution → Max Pooling → Fully-Connected → Fully-Connected → Softmax. I also use Dropout, RMSprop, etc.
I used 20,000 tweets of tweets_clean.txt in satwantrana / CharSCNN. Training data 18000 tweets and test data 2000 tweets. Each tweet is labeled 0/1 (negative / positive), so it is classified into 2 classes.
Converting the very first input to Word Embeddings It seems that the result will be better if you use the weight pre-learned with Word2Vec, so I would like to try that as well.
Since Word2Vec was included in the Chainer sample, I used the result of pre-learning with Skip-gram with Negative-Sampling. Pre-learning with Chainer and turning the Theano program lol. Thank you cPickle.
Well, it was better to pre-learn the rise, but it is better not to pre-learn the final result. .. .. What if I turn it a little longer?
Recommended Posts