Multilingual Entity-Centered Sentiment Analysis Evaluated by Parallel Corpora

Proceedings of Recent Advances in Natural Language Processing, pages 770–775, Hissar, Bulgaria, 12-14 September 2011. Multilingual Entity-Centered Sentiment Analysis Evaluated by Parallel Corpora Josef Steinberger and Polina Lenkova and Mijail Kabadjov Ralf Steinberger and Erik van der Goot EC Joint Research Centre 21027, Ispra (VA), Italy { Josef.Steinberger, Polina.Lenkova, Ralf.Steinberger Mijail.Kabadjov, Erik.van-der-Goot} @jrc.ec.europa.eu Abstract We propose the creation and use of a mul- tilingual parallel news corpus annotated with opinion towards entities, produced by projecting sentiment annotation from one language to several others. The objective is to save annotation time for development and evaluation purposes, and to guarantee comparability of opinion mining evalua- tion results across languages. By creating this resource, we answered the question whether sentiment is consistently trans- lated across languages so that projection can actually be an option. We describe our approach to multilingual sentiment anal- ysis and show its performance in 7 lan- guages of the parallel corpus. 1 Introduction In sentiment analysis the goal is to detect and clas- sify subjective content of a text. The text can be classified as a whole such as in product reviews, in which an overall judgment is assigned to the prod- uct. If we move to the news domain, the overall sentiment score of an article can be used for de- tecting bad or good news. It can be used also for detecting the changes in sentiment in a particular topic. However, if the goal is to detect sentiment expressed towards entities, the aggregated senti- ment of the articles, in which the entity appears, need not to correspond to opinions expressed to- wards the entity. The entity can be mentioned pos- itively in a very negative article. We have to go down and analyze each entity mention based on the surrounding context. Solving the problem in multilingual environ- ment and gathering large amounts of articles from many sources give advantage to detect news opin- ions expressed in different countries towards same persons. Also, it eliminates the biased news. How- ever, multilinguality brings another challenge. For instance, it is not easy to develop NLP tools like parsers or taggers in many languages, also using them can cause computational problems when ap- plied on large amounts of articles every day. An- other difficulty comes with resources. Sentiment- annotated data are not usually available for other types of texts then reviews, or they are almost ex- clusively available for English. Sentiment dictio- naries are also mostly available for English only or, if they exist for other languages, they are not comparable, in the sense that they have been developed for different purposes, have different sizes, are based on different definitions of what sentiment or opinion means. We addressed the resource bottleneck for senti- ment dictionaries, by developing highly multilin- gual and comparable sentiment dictionaries hav- ing similar sizes and based on a common specifi- cation (Steinberger et al., 2011). Our sentiment system is simply based on counting subjective terms around entity mentions (mainly persons and organizations). Evaluating its performance in more …

Content truncated to 500 words