Hybrid attribute based sentiment classification of online reviews for consumer intelligence

摘要

Rich online consumer reviews (OCR) can be mined to gain valuable insights, beneficial for both brands and future buyers. Recently, aspect based sentiment classification have shown excellent results for fine grained sentiment analysis of OCR. However, there are only few studies so far that rely on both explicitly deriving sentiment using syntactic features, and capturing implicit contextual word relations for the task of aspect based sentiment classification. In this paper, we propose a novel method: Hybrid Attribute Based Sentiment Classification (HABSC) with the aim to derive sentiment orientation of OCR by capturing implicit word relations and incorporating domain specific knowledge. First, we detect the most frequent bigrams and trigrams in the corpus, followed by POS tagging to retain aspect descriptions and opinion words. Then, we employ TFIDF (term frequency inverse document frequency) to represent each document, followed by automatically extracting optimal number of topics in the given corpus. All the adjectives and adverbs are labelled using domain specific knowledge and pre-existing lexicons. Lastly, we find sentiment orientation of each review under the assumption that each review is a mixture of weighted and sentiment labelled attributes. We test the efficiency of our method using datasets from two different domains: hotel reviews from TripAdvisor.com and mobile phone reviews from Amazon.com. Results show that, the classification accuracy of HABSC significantly exceeds various state-of-the-art methods including aspect-based sentiment classification and supervised classification using distributed word and paragraph vectors. Our method also exhibits less computational time as compared to distributed vectorization schemes.