Deep-MSP: Morphological Sentence Pattern Recognition based on ConvNet

Abstract

Sentiment analysis aims to observe and summarize a person's opinions or emotional states through textual data. Despite the demands of sentiment analysis methods for analyzing social media data, fundamental challenges still remained because user-generated data is unstructured, unlabeled, and "noisy". The morphological sentence pattern (MSP) model, an aspect-based lexicon building method, is proposed for dealing with the problems of the transitional sentiment analysis by recognizing the "aspect-expression" in a sentence. However, there are limitations on this model. Firstly, since the MSP model is based on the pattern matching, the sentences cannot be analyzed when the pattern does not exist in the lexicon. Secondly, the patterns should be continuously updated to maintain a high level of accuracy. In this paper, to compensate for the limitations of the MSP model, we proposed Deep-MSP, a deep learning approach based on multiple convolutional neural networks (ConvNet or CNN), designed to recognize whether or not the target part-of-speech has potential to be the aspect-expression from not only existing patterns but also new patterns.