Mining the Opinionated Web: Classification and Detection of Aspect Contexts for Aspect Based Sentiment Analysis

摘要

Aspect Based Sentiment Analysis (ABSA) provides further insight into the analysis of social media. Understanding user opinion about different aspects of products, services or policies can be used for improving and innovating in an effective way. Thus, it is becoming an increasingly important task in the Natural Language Processing (NLP) realm. The standard pipeline of aspect-based sentiment analysis consists of three phases: aspect category detection, Opinion Target Extraction (OTE) and sentiment polarity classification. In this article, we propose an alternative pipeline: OTE, aspect classification, aspect context detection and sentiment classification. As it can be observed, the opinionated words are first detected and then are classified into aspects. In addition, the opinionated fragment of every aspect is delimited before performing the sentiment analysis. This paper is focused on the aspect classification and aspect context detection phases and proposes a twofold contribution. First, we propose a hybrid model consisting of a word embeddings model used in conjunction with semantic similarity measures in order to develop an aspect classifier module. Second, we extend the context detection algorithm by Mukherjee et al. to improve its performance. The system has been evaluated using the SemEval2016 datasets. The evaluation shows through several experiments that the use of hybrid techniques that aggregate different sources of information improves the classification performance.