How Universal are Universal Dependencies? Exploiting Syntax for Multilingual Clause-level Sentiment Detection

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4063–4073 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4063 How Universal are Universal Dependencies? Exploiting Syntax for Multilingual Clause-level Sentiment Detection Hiroshi Kanayama∗, Ran Iwamoto† ∗IBM Research - Tokyo 19-21 Nihonbashi Hakozaki-cho, Chuo-ku, Tokyo 103-8510 Japan hkana@jp.ibm.com †Keio University 3-14-1 Hiyoshi, Kohoku-ku, Yokohama-shi, Kanagawa 223-8522 Japan r.iwamoto@keio.jp Abstract This paper investigates clause-level sentiment detection in a multilingual scenario. Aiming at a high-precision, fine-grained, configurable, and non-biased system for practical use cases, we have designed a pipeline method that makes the most of syntactic structures based on Universal Dependencies, avoiding machine-learning approaches that may cause obstacles to our purposes. We achieved high precision in sentiment detection for 17 languages and identified the advantages of common syntactic structures as well as issues stemming from structural differences on Universal Dependencies. In addition to reusable tips for handling multilingual syntax, we provide a parallel benchmarking data set for further research. Keywords: Universal Dependencies, sentiment analysis, multilinguality, parsing 1. Introduction Sentiment analysis and opinion mining (Pang and Lee, 2008) along with their multilingualization (Korayem et al., 2016) have been studied for many years. In a typical use case for enterprises that continually seek information to im- prove their products or services while estimating future de- mands, it is very important to detect the individual utter- ances that specify positive or negative properties, beyond estimating the overall preference (typically, a number of stars) written in a review. In this work, we pursue clause-level sentiment detection, which shares a similar motivation to aspect-based sentiment analysis (ABSA) (Pontiki et al., 2016). Our goal is to ex- plore a general purpose system that doesn’t require data- specific features such as user and time information in Twit- ter data or Amazon reviews. Most of the prior work has exploited machine learning for sentiment classification (Pang et al., 2002; Wang et al., 2012) or dictionary induction (Hamilton et al., 2016). How- ever, it has been pointed out that statistical approaches to user-generated data (e.g. Twitter) may extract “iPod“ as a positive keyword (Saif et al., 2012), and while this will im- prove the score in benchmarking datasets in a world where many people like iPods, we need to design a system free from such prior biases. Systems that can accurately detect positive and negative opinions are essential when it comes to solving other tasks, such as the recognition of sarcasm (Tungthamthiti et al., 2014) and social analysis to investi- gate race bias (Merullo et al., 2019). To achieve a multilingual system that meets these require- ments, we utilize a pipeline approach for clause-level sen- timent detection. Our approach applies sentiment lexicon and syntactic rules to the output of dependency parser based on Universal Dependencies (UD) (Nivre et al., 2016). By capturing syntactic phenomena such as coordination and negation, we can accurately extract positive and negative polarities along with their corresponding predicate and tar- get by means of lexicon that can be manually …

Content truncated to 500 words