Improving Sentiment Analysis in Twitter Using Multilingual Machine Translated Data

Proceedings of Recent Advances in Natural Language Processing, pages 49–55, Hissar, Bulgaria, 7-13 September 2013. Improving Sentiment Analysis in Twitter Using Multilingual Machine Translated Data Alexandra Balahur European Commission Joint Research Centre Via E. Fermi 2749 21027 Ispra (VA), Italy alexandra.balahur@jrc.ec.europa.eu Marco Turchi Fondazione Bruno Kessler-IRST Via Sommarive, 18 Povo, Trento, Italy turchi@fbk.eu Abstract Sentiment analysis is currently a very dy- namic field in Computational Linguistics. Research herein has concentrated on the development of methods and resources for different types of texts and various lan- guages. Nonetheless, the implementa- tion of a multilingual system that is able to classify sentiment expressed in various languages has not been approached so far. The main challenge this paper addresses is sentiment analysis from tweets in a multi- lingual setting. We first build a simple sen- timent analysis system for tweets in En- glish. Subsequently, we translate the data from English to four other languages - Ital- ian, Spanish, French and German - using a standard machine translation system. Fur- ther on, we manually correct the test data and create Gold Standards for each of the target languages. Finally, we test the per- formance of the sentiment analysis classi- fiers for the different languages concerned and show that the joint use of training data from multiple languages (especially those pertaining to the same family of lan- guages) significantly improves the results of the sentiment classification. 1 Introduction Sentiment analysis is a task in Natural Language Processing whose aim is to automatically detect and classify sentiments in texts. Generally, the “positive”, “negative” and “neutral” classes are considered, although other scales have also been used (e.g. from 1 to 5 “stars” - according to the reviewing systems put at the disposal of clients or users by amazon.com, booking.com, etc.; adding the “very positive” and “very negative” classes, scales from 1 to 10, etc.). In this article, we deal with the issue of sen- timent analysis in tweets, in a multilingual set- ting. We employ machine translation - which was shown to be at a sufficiently high level of perfor- mance (Balahur and Turchi, 2012) - to obtain data in four languages. Our goal is to test if the use of multilingual data can help to improve sentiment classification in tweets (as shown to be the case in formal texts - (Banea et al., 2010)) and if the joint use of data coming from similar languages or lan- guages that are different in structure can influence on the final result. The main problem when designing automatic methods for the treatment of tweets is that they are highly informal texts, i.e. they contain slang, emoticons, repetitions of letters or punctuation signs, misspellings (done on purpose or due to writing them from mobile devices), entire words in capital letters, etc. In order to test our hypotheses, we first design a simple tweet sentiment analysis system for En- glish, taking into account the specificity of ex- pressions employed, but without using language- specific text processing tools. The motivation is related …

Content truncated to 500 words