Thai Sentiment Analysis via Bidirectional LSTM-CNN Model with Embedding Vectors and Sentic Features

Abstract

Sentiment analysis is one of the most frequently performed tasks in Natural Language Processing that plays an important role in marketing research. It allows us to understand customer sentiment. The outcomes from this kind of analysis can be used to improve products and services. Recently, a Word2Vec model, a technique for word embedding (converting text into a number) has been developed and used successfully to a degree to get the sentiment of customers from the text responses that they provided. This work attempted to incorporate two more features-part-of-speech and sentic features-to make the analysis more accurate. The part-of-speech feature identifies the type of words that better convey various sentiments, while the sentic feature identifies the emotion underlying certain words. Combining Bidirectional Long Short-term Memory and Convolutional Neural Networks models with several combinations of the features mentioned, we performed a sentiment analysis of Thai children stories and found that the combination of all three features gave the best result at 78.89 % F1-score.