Knowledge-Based Approach to Concept-Level Opinion Mining

Abstract

Present day, World Wide Web offers a great means to share knowledge, becoming one large repository of valuable opinions of different people on numerous products and services. It is not straightforward to access these opinions because the available number is vast, and understanding what is actually meant has always been problematic with the complexity of human language. These texts could contain various attributes (or features) forming complex opinion relationships hidden inside large numbers of sentences. Most of the existing approaches for opinion mining are based on word-level analysis of texts and are able to detect only explicitly expressed opinions. Concept-level opinion mining attempts to go beyond mere word level analysis, providing a more semantic inspection of text through the use of semantic networks, empowering novel approaches to sentiment analysis in potentially any domain. Commonsense knowledge, spans a huge portion of human experience, encompassing knowledge about the spatial, physical, social, temporal, and psychological aspects of everyday life. This research proposes a novel approach to concept-level opinion mining by relying on publicly available commonsense knowledge, with the assumption that closely related concepts in the knowledge-base are likely to share similar sentiments to each other. A sentiment concept dictionary is developed by blending sentiment seed sources and a commonsense knowledge-base to provide a general sentiment classification model, which is potentially capable of handling any domain. This study tries to simulate the cognitive ability of human brain, up to some extent by identifying concepts, process those with the aid of large semantic knowledge-bases and machine learning with a Support Vector Machines classification model. Concept extraction from natural language text achieved in two steps, part-of-speech based extraction and dependency based extraction. Bag-of-concepts features with novel weighting function were generated along with other syntactical and lexical features for the classification with SVM model. Successful results were achieved by evaluating the proposed method using online customer review data spread over five diverse domains with the purpose of conducting domain-independent sentiment analysis. These experiments show state of the art promising results, proving that the proposed approach suits both domain dependent and domain independent sentiment analysis.