Sentimental Analysis of Short Texts

Sentiment analysis of short texts such movie reviews and tweets is challenging because of the limited contextual information. To deal with limited context, the paper proposed a convolution based neural network that exploits from character- to sentence-level information to perform sentiment analysis of short texts.

The model uses character-level embeddings and word-level embeddings to capture shape and syntactic information from a sentence respectively. Once the embeddings for each character are assigned, a convolutional layer is applied to each window of k characters. A max pooling operation is then applied to calculate the word embedding. We then apply another convolutional layer to a window of words followed by a max-pool layer resulting in a sentence embedding.

The sentence embedding is then passed through a binary classification layer resulting in the overall sentiment of the sentence. For speeding up training and better accuracy, I also performed an unsupervised pre-training of word embeddings.