Leveraging Structural and Semantic Correspondence for Attribute-Oriented Aspect Sentiment Discovery

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5528–5538, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5528 Leveraging Structural and Semantic Correspondence for Attribute-Oriented Aspect Sentiment Discovery Zhe Zhang Watson Group IBM Corporation Research Triangle Park, NC 27703-9141 zhangzhe@us.ibm.com Munindar P. Singh Department of Computer Science North Carolina State University Raleigh, NC 27695-8206 singh@ncsu.edu Abstract Opinionated text often involves attributes such as authorship and location that influence the sentiments expressed for different aspects. We posit that structural and semantic correspon- dence is both prevalent in opinionated text, es- pecially when associated with attributes, and crucial in accurately revealing its latent aspect and sentiment structure. However, it is not rec- ognized by existing approaches. We propose Trait, an unsupervised probabilis- tic model that discovers aspects and senti- ments from text and associates them with dif- ferent attributes. To this end, Trait infers and leverages structural and semantic corre- spondence using a Markov Random Field. We show empirically that by incorporating at- tributes explicitly Trait significantly outper- forms state-of-the-art baselines both by gener- ating attribute profiles that accord with our in- tuitions, as shown via visualization, and yield- ing topics of greater semantic cohesion. 1 Introduction Opinionated text is often associated with differ- ent attributes—latent variables that serve as ref- erence frames relative to which the underlying as- pects and sentiments are expressed. Common at- tributes in consumer reviews include author type (e.g., business traveler or tourist for hotel reviews; location for reviews of music (McDermott et al., 2016); culture on reviews of food (Bahauddin and Shaarani, 2015)). Whereas current approaches consider attributes in a one-off manner in each application, we posit that attributes can be sys- tematically extracted if we can properly capture the structural and semantic correspondence that is prevalent in opinionated text. We claim that ig- noring attributes may lead to biased inference on aspects and sentiments. As evidence, we demon- strate an approach that outperforms the state of the art and yields intuitive and cohesive topics. We propose Trait, a general model for discover- ing attribute-oriented aspects and sentiments from text. By incorporating attributes, Trait automat- ically generates profiles that describe attributes in terms of sentiments and aspects. To leverage struc- tural and semantic correspondence, Trait applies a Markov Random Field as regularization over sen- tences during inference. We evaluate Trait on four datasets from two domains and consider three at- tributes. Trait successfully discovers aspects asso- ciated with sentiments; the generated word clus- ters are more cohesive than the state-of-the-art baselines; the generated attribute profiles are well correlated with ground truth. Motivating Example. Figure 1 presents two ho- tel reviews from TripAdvisor. We manually as- sign aspect labels for sentences and calculate pair- wise cosine similarity between sentences using sentence embedding from a pretrained sentence encoding model (Cer et al., 2018). Review A and Review B, which mention as- pects Room, Location, and Type, exhibit structural and semantic …

Content truncated to 500 words