A Sentiment-aligned Topic Model for Product Aspect Rating Prediction

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1192–1202, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A Sentiment-aligned Topic Model for Product Aspect Rating Prediction Hao Wang School of Computing Science Simon Fraser University Burnaby, BC, Canada hwa63@sfu.ca Martin Ester School of Computing Science Simon Fraser University Burnaby, BC, Canada ester@sfu.ca Abstract Aspect-based opinion mining has attracted lots of attention today. In this paper, we address the problem of product aspect rat- ing prediction, where we would like to ex- tract the product aspects, and predict as- pect ratings simultaneously. Topic mod- els have been widely adapted to jointly model aspects and sentiments, but exist- ing models may not do the prediction task well due to their weakness in sentiment extraction. The sentiment topics usually do not have clear correspondence to com- monly used ratings, and the model may fail to extract certain kinds of sentiments due to skewed data. To tackle this prob- lem, we propose a sentiment-aligned topic model(SATM), where we incorporate two types of external knowledge: product- level overall rating distribution and word- level sentiment lexicon. Experiments on real dataset demonstrate that SATM is ef- fective on product aspect rating prediction, and it achieves better performance com- pared to the existing approaches. 1 Introduction Online reviews have become an important source of information for consumers. People tend to read reviews to help them compare products, and make informed decisions. As the volume of product re- views continues to grow, it is often impossible to read all of them, which calls for efficient methods for opinion mining. Nowadays, for each product, many websites aggregate the overall rating of re- views, and display its distribution. However, this cannot provide detailed information. For exam- ple, two products may have similar overall rating distributions, while people talk about different un- satisfactory aspects. This problem has inspired a new line of research on aspect-level opinion min- ing(Hu and Liu, 2004). An aspect refers to a rateable feature, such as staff and location in hotel reviews, or size and battery for digital camera reviews. In this paper, we deal with the problem of product aspect rat- ing prediction. The input is a collection of prod- ucts, and each product is associated with a set of reviews. The goal is to extract the corpus-level as- pects, and predict the aspect ratings for each prod- uct. This kind of fine-grained sentiment analysis will help users efficiently digest the reviews, and gain more insight into the product quality. The product aspect rating prediction problem usually involves two subtasks: aspect extraction and sentiment identification(Titov and McDonald, 2008b). Given some text, we would like to know what aspects it talks about, and what kind of sen- timents are expressed. For example, given a sen- tence “the room is filthy”, we would like to know that it talks about the aspect “room”. Also, “filthy” is a sentiment word, and it expresses strongly neg- ative sentiment towards the aspect “room”. Topic …

Content truncated to 500 words