Leveraging Just a Few Keywords for Fine-Grained Aspect Detection Through Weakly Supervised Co-Training

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4611–4621, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4611 Leveraging Just a Few Keywords for Fine-Grained Aspect Detection Through Weakly Supervised Co-Training Giannis Karamanolakis, Daniel Hsu, Luis Gravano Columbia University, New York, NY 10027, USA {gkaraman, djhsu, gravano}@cs.columbia.edu Abstract User-generated reviews can be decomposed into fine-grained segments (e.g., sentences, clauses), each evaluating a different aspect of the principal entity (e.g., price, quality, appear- ance). Automatically detecting these aspects can be useful for both users and downstream opinion mining applications. Current super- vised approaches for learning aspect classifiers require many fine-grained aspect labels, which are labor-intensive to obtain. And, unfortu- nately, unsupervised topic models often fail to capture the aspects of interest. In this work, we consider weakly supervised approaches for training aspect classifiers that only require the user to provide a small set of seed words (i.e., weakly positive indicators) for the as- pects of interest. First, we show that current weakly supervised approaches do not effec- tively leverage the predictive power of seed words for aspect detection. Next, we pro- pose a student-teacher approach that effec- tively leverages seed words in a bag-of-words classifier (teacher); in turn, we use the teacher to train a second model (student) that is po- tentially more powerful (e.g., a neural network that uses pre-trained word embeddings). Fi- nally, we show that iterative co-training can be used to cope with noisy seed words, leading to both improved teacher and student models. Our proposed approach consistently outper- forms previous weakly supervised approaches (by 14.1 absolute F1 points on average) in six different domains of product reviews and six multilingual datasets of restaurant reviews. 1 Introduction A typical review of an entity on platforms such as Yelp and Amazon discusses multiple aspects of the entity (e.g., price, quality) in individual re- view segments (e.g., sentences, clauses). Consider for example the Amazon product review in Fig- ure 1. The text discusses various aspects of the Great price for an excellent LED TV Great Tv for the price. Easy to setup. The audio is ok for the tiny speakers. The picture is just as good as my panasonic viera 42" plasma tv. Much better than the 20" tube tv. Price Ease of Use Image Sound Quality General Aspect Sentence Figure 1: Example of product review with aspect an- notations: each individual sentence of the review dis- cusses a different aspect (e.g., price) of the TV. TV such as price, ease of use, and sound quality. Given the vast number of online reviews, both sell- ers and customers would benefit from automatic methods for detecting fine-grained segments that discuss particular aspects of interest. Fine-grained aspect detection is also a key task in downstream applications such as aspect-based sentiment anal- ysis and multi-document summarization (Hu and Liu, 2004; Liu, 2012; Pontiki et al., 2016; Ange- lidis and Lapata, 2018). In this work, we …

Content truncated to 500 words