Bilingual Sentiment Consistency for Statistical Machine Translation

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 607–615, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Bilingual Sentiment Consistency for Statistical Machine Translation Boxing Chen and Xiaodan Zhu National Research Council Canada 1200 Montreal Road, Ottawa, Canada, K1A 0R6 {Boxing.Chen, Xiaodan.Zhu}@nrc-cnrc.gc.ca Abstract In this paper, we explore bilingual sentiment knowledge for statistical machine translation (SMT). We propose to explicitly model the consistency of sentiment between the source and target side with a lexicon-based approach. The experiments show that the proposed mod- el significantly improves Chinese-to-English NIST translation over a competitive baseline. 1 Introduction The expression of sentiment is an interesting and integral part of human languages. In written text sentiment is conveyed by senses and in speech also via prosody. Sentiment is associated with both evaluative (positive or negative) and potency (de- gree of sentiment) ― involving two of the three major semantic differential categories identified by Osgood et al. (1957). Automatically analyzing the sentiment of mono- lingual text has attracted a large bulk of research, which includes, but is not limited to, the early ex- ploration of (Turney, 2002; Pang et al., 2002; Hat- zivassiloglou & McKeown, 1997). Since then, research has involved a variety of approaches and been conducted on various type of data, e.g., prod- uct reviews, news, blogs, and the more recent so- cial media text. As sentiment has been an important concern in monolingual settings, better translation of such information between languages could be of interest to help better cross language barriers, particularly for sentiment-abundant data. Even when we ran- domly sampled a subset of sentence pairs from the NIST Open MT1 training data, we found that about 48.2% pairs contain at least one sentiment word on both sides, and 22.4% pairs contain at least one 1 http://www.nist.gov/speech/tests/mt intensifier word on both sides, which suggests a non-trivial percent of sentences may potentially involve sentiment in some degree2. # snt. pairs % snt. with sentiment words % snt. with intensifiers 103,369 48.2% 22.4% Table 1. Percentages of sentence pairs that contain sen- timent words on both sides or intensifiers3 on both sides. One expects that sentiment has been implicitly captured in SMT through the statistics learned from parallel corpus, e.g., the phrase tables in a phrase-based system. In this paper, we are interest- ed in explicitly modeling sentiment knowledge for translation. We propose a lexicon-based approach that examines the consistency of bilingual subjec- tivity, sentiment polarity, intensity, and negation. The experiments show that the proposed approach improves the NIST Chinese-to-English translation over a strong baseline. In general, we hope this line of work will help achieve better MT quality, especially for data with more abundant sentiment, such as social media text. 2 Related Work Sentiment analysis and lexicon-based approach- es Research on monolingual sentiment analysis can be found under different names such as opinion, stance, appraisal, and semantic orientation, among others. The overall goal is to label a span of text either as positive, negative, …

Content truncated to 500 words