Does ‘well-being’ translate on Twitter?

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2042–2047, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Does ‘well-being’ translate on Twitter? Laura K. Smith1 Salvatore Giorgi1 Rishi Solanki2 Johannes C. Eichstaedt1 H. Andrew Schwartz3 Muhammad Abdul-Mageed4 Anneke Buffone1 and Lyle H. Ungar5 1Department of Psychology, University of Pennsylvania 2Electrical and Systems Engineering, University of Pennsylvania 3Computer Science, Stony Brook University 4Library, Archival and Information Studies, University of British Columbia 5Computer and Information Science, University of Pennsylvania lasm@sas.upenn.edu, sgiorgi@sas.upenn.edu Abstract We investigate whether psychological well- being translates across English and Span- ish Twitter, by building and comparing source language and automatically translated weighted lexica in English and Spanish. We find that the source language models perform substantially better than the machine trans- lated versions. Moreover, manually correct- ing translation errors does not improve model performance, suggesting that meaningful cul- tural information is being lost in translation. Further work is needed to clarify when au- tomatic translation of well-being lexica is ef- fective and how it can be improved for cross- cultural analysis. 1 Introduction Interest in sentiment analysis spans academic and commercial domains, with wide-ranging applica- tions (Pang and Lee, 2008; Liu, 2012). While the majority of tools for sentiment analysis have been developed for English text, ideally sentiment and emotion could be analyzed across many languages. Does one need to build models for each language of interest, or can models be applied cross-culturally? More generally, how do cultures differ in the lan- guage they use to express sentiment and feeling? Sentiment in resource-poor languages has com- monly been assessed by first translating text into En- glish and then applying an English sentiment model (Mohammad et al., 2016). This approach is eco- nomical and efficient, as building each model of in- terest in every target language is resource-intensive. Yet it is not clear how much culturally specific information and accuracy are lost in the transla- tion process, and specifically how this varies across languages, cultures, linguistic content, and corpora (e.g., social media vs. news). While extensive work has demonstrated that automatic machine transla- tion (MT) methods are competitive when translat- ing opinion in news and blogs, less research has ex- amined the translation of sentiment on social me- dia, and specifically on Twitter, known for its re- striction of individual exchanges to short samples of text (140 characters) and informal language. More- over, research has not focused on translating subjec- tive well-being specifically. Beyond sentiment, this paper investigates how ex- pressions of personal well-being translate between English and Spanish on Twitter. We have English and Spanish speakers annotate Tweets in their na- tive language for five components of subjective well- being (positive emotion, engagement, positive re- lationships, meaning, and accomplishment) (Selig- man, 2011). We then compare how well models trained and tested in the same language compare to (a) models developed in one language, and then translated (using Google Translate) to the other lan- guage (e.g., how well English models translated to Spanish …

Content truncated to 500 words