Balancing Translation Quality and Sentiment Preservation (Non-archival Extended Abstract)

Balancing Translation Quality and Sentiment Preservation Pintu Lohar1 pintu.lohar@adaptcentre.ie Haithem Afli1-2 haithem.afli@cit.ie Andy Way1 andy.way@adaptcentre.ie 1ADAPT Centre, School of Computing, Dublin City University 2Cork Institute of Technology Abstract Social media platforms such as Twitter and Facebook are hugely popular websites through which Internet users can communicate and spread information worldwide. On Twitter, mes- sages (tweets) are generated by users from all over the world in many different languages. Tweets about different events almost always encode some degree of sentiment. As is often the case in the field of language processing, sentiment analysis tools exist primarily in English, so if we want to understand the sentiment of the original tweets, we are forced to translate them from the source language into English and pushing the English translations through a sentiment analysis tool. However, Lohar et al. (2017) demonstrated that using freely available translation tools often caused the sentiment encoded in the original tweet to be altered. As a consequence, they built a series of sentiment-specific translation engines and pushed tweets containing either positive, neutral or negative sentiment through the appropriate engine to improve sentiment preservation in the target language. For certain tasks, maintaining sentiment polarity in the target language during the translation process is arguably more important than the absolute translation quality obtained. In the work of Lohar et al. (2017), a small drop off in translation quality per se was deemed tolerable. In this work, we focus on maintaining the level of sentiment preservation while trying to improve translation quality still further. We propose a nearest sentiment class- combination method to extend the existing sentiment-specific translation systems by adding training data from the nearest-sentiment class. Our experimental results on German-to-English reveal that our approach is capable of achieving a proper balance between translation quality and sentiment preservation. 1 Introduction The rapid development of internet technologies has given rise to a significant growth in gener- ating and sharing of user-generated content (UGC). Internet users from all over the world stay connected via widely used social networking websites such as Twitter and Facebook by shar- ing information in different languages. On Twitter, tweets on different events related to sports, festivals, conferences and political events almost always encode some degree of sentiment. Ac- cordingly, the task of sentiment analysis is important on datasets such as these. However, given the lack of such tools for most languages, this can only be achieved via an MT-based sentiment analysis approach, where the tweets are first translated from the original language to English and then sentiment analysis is performed on th English translations (Araujo et al. (2016)). Proceedings of AMTA 2018, vol. 1: MT Research Track Boston, March 17 - 21, 2018 | Page 81 However, the 140-character limitation – recently expanded to 280 – in Twitter encourages users to use short forms at word and phrase levels. Moreover, tweets often contain (deliber- ate) spelling errors, hashtags, user names and URLs which pose challenges in the translation process. In order to build corpus-based MT systems, a parallel …

Content truncated to 500 words