Exploring lexico-semantic patterns for aspect-based sentiment analysis

Abstract

Web 2.0 has caused a boom in user-generated content, which contains a lot of valuable information. Analysis of these natural language data requires advanced machine learning techniques. This research focuses on determining aspect-based sentiment in consumer reviews using lexico-semantic patterns. We propose a method using a Support Vector Machine with 6 different pattern classes: lexical, syntactical, semantic, sentiment, hybrid, and surface. We show that several of these patterns, including synset bigram, negator-POS bigram, and POS bigram, can be used to better determine the aspect-based sentiment, using two widely used real-world data sets on consumer reviews. Our approach achieves 69.0% and 73.1% F1 score for the two data sets, respectively, an increase of 15.3% and 16.1% respectively compared to the considered baseline.