Better Document-level Sentiment Analysis from RST Discourse Parsing

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2212–2218, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Better Document-level Sentiment Analysis from RST Discourse Parsing∗ Parminder Bhatia and Yangfeng Ji and Jacob Eisenstein School of Interactive Computing Georgia Institute of Technology Atlanta, GA 30308 parminder.bhatia243@gmail.com, {jiyfeng,jacobe}@gatech.edu Abstract Discourse structure is the hidden link be- tween surface features and document-level properties, such as sentiment polarity. We show that the discourse analyses produced by Rhetorical Structure Theory (RST) parsers can improve document-level senti- ment analysis, via composition of local in- formation up the discourse tree. First, we show that reweighting discourse units ac- cording to their position in a dependency representation of the rhetorical structure can yield substantial improvements on lexicon-based sentiment analysis. Next, we present a recursive neural network over the RST structure, which offers sig- nificant improvements over classification- based methods. 1 Introduction Sentiment analysis and opinion mining are among the most widely-used applications of language technology, impacting both industry and a vari- ety of other academic disciplines (Feldman, 2013; Liu, 2012; Pang and Lee, 2008). Yet senti- ment analysis is still dominated by bag-of-words approaches, and attempts to include additional linguistic context typically stop at the sentence level (Socher et al., 2013). Since document-level opinion mining inherently involves multi-sentence texts, it seems that analysis of document-level structure should have a role to play. A classic example of the potential relevance of discourse to sentiment analysis is shown in Fig- ure 1. In this review of the film The Last Samu- rai, the positive sentiment words far outnumber the :::::::: negative ::::::::: sentiment words. But the discourse structure — indicated here with Rhetorical Struc- ture Theory (RST; Mann and Thompson, 1988) — ∗Code is available at https://github.com/ parry2403/R2N2 R CONCESSION � JUSTIFY 1A CONJUNCTION � ELABORATION 1B 1C 1D � JUSTIFY 1E CONJUNCTION 1F 1G 1H [It could have been a great movie]1A [It does have beautiful scenery,]1B [some of the best since Lord of the Rings.]1C [The acting is well done,]1D [and I really liked the son of the leader of the Samurai.]1E [He was a likable chap,]1F [and I :::: hated to see him die.]1G [But, other than all that, this movie is :::::: nothing more than hid- den ::::: rip-offs.]1H Figure 1: Example adapted from Voll and Taboada (2007). clearly favors the final sentence, whose polarity is negative. This example is illustrative in more than one way: it was originally identified by Voll and Taboada (2007), who found that manually- annotated RST parse trees improved lexicon- based sentiment analysis, but that automatically- generated parses from the SPADE parser (Soricut and Marcu, 2003), which was then state-of-the-art, did not. Since this time, RST discourse parsing has im- proved considerably, with the best systems now yielding 5-10% greater raw accuracy than SPADE, depending on the metric. The time is therefore right to reconsider the effectiveness of RST for document-level sentiment analysis. In this pa- per, we present two different ways of combin- …

Content truncated to 500 words