Semantic Simplification for Sentiment Classification

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11022 - 11032 December 7-11, 2022 ©2022 Association for Computational Linguistics Semantic Simplification for Sentiment Classification Xiaotong Jiang, Zhongqing Wang∗, and Guodong Zhou Natural Language Processing Lab, Soochow University, Suzhou, China devjiang@outlook.com {wangzq, gdzhou}@suda.edu.cn Abstract Recent work on document-level sentiment clas- sification has shown that the sentiment in the original text is often hard to capture, since the sentiment is usually either expressed implicitly or shifted due to the occurrences of negation and rhetorical words. To this end, we enhance the original text with a sentiment-driven sim- plified clause to intensify its sentiment. The simplified clause shares the same opinion with the original text but expresses the opinion much more simply. Meanwhile, we employ Abstract Meaning Representation (AMR) for generating simplified clauses, since AMR explicitly pro- vides core semantic knowledge, and potentially offers core concepts and explicit structures of original texts. Empirical studies show the ef- fectiveness of our proposed model over several strong baselines. The results also indicate the importance of simplified clauses for sentiment classification. 1 Introduction As a critical application of natural language pro- cessing, document-level sentiment classification has received considerable attention during the last two decades with the underlying assumption that the entire text has an overall polarity. In the literature, previous studies focus on pre- dicting the overall sentiment from original text using either statistical (Pang et al., 2002; Wilson et al., 2005; Xia et al., 2011) or neural models (Kim, 2014; Tang et al., 2016b; Chen et al., 2020). How- ever, as shown in Figure 1(a), the overall sentiment of the original text is often hard to capture, since the overall sentiment is usually either expressed implic- itly or shifted due to the occurrences of negation and rhetorical words. To address above challenges in the original text, both attention-based (Chen et al., 2017; Amplayo et al., 2018; Tay et al., 2018) and rhetorical structure-based approaches (Li et al., ∗ Corresponding author Our puppy loves to dump and chew up all water bowls . Puppy don't have to lick bowls. If it broke I would go out the same day and buy a new one. (a) Original Text Puppy loves to chew up bowls. (b) AMR-based Simplified Semantic Graph (c) Simplified Clause Linearization: (love-01 :ARG0(puppy) :ARG1(chew up :ARG1(bowl))) puppy love chew up bowl :ARG0 :ARG1 :ARG1 Figure 1: Example of simplified clause with AMR- based representation. 2010; Xia et al., 2015; Pröllochs et al., 2019; Ya- dav et al., 2021) have been proposed. Although the above pioneer studies have achieved certain suc- cess, they either heavily rely on human knowledge or suffer from the complex structure of the original text. To tackle the above limitations, we simplify the original text to a simplified clause and employ the simplified clause for sentiment classification. As shown in Figure 1(c), a qualified simplified clause shares the same opinion with the original text and expresses the opinion much more simply. There- fore, it is much easier …

Content truncated to 500 words