Tower of Babel: a crowdsourcing game building sentiment lexicons for resource-scarce languages

Abstract

With the growths of the Web 2.0 and social media, the size of data on the Web has been continuously increasing. According a Cisco statistics, the Web traffic has already reached 5,000 PB per month worldwide as of 2012, and is continuously increasing. In the realm of social media, Twitter produces more than 400 million tweet messages everyday, and Facebook generates more than 2.5 billion items everyday in 2012. These Web data mainly take text format, and are written in multiple languages. Facebook currently serves 70 languages, and 75 percent of its users live outside of the U.S. More than 60 percents of tweet messages in 2012 were written in non-English languages, and the volume of multilingual messages is rapidly increasing.Sentiment analysis is one of the key enabling technologies in the field of natural language processing. Sentiment analysis finds many useful applications in various domains such as brand management, public opinion survey, predicting stock market, etc.For English, there have been numerous publicly available resources for sentiment analysis. Sentiment analysis has gone beyond the boundary of research and started to be used commercially. However, we still lack resources to perform sentiment analysis in non-English languages because most computational linguistic research focuses on English and building high quality resources for sentiment analysis is costly. We term non-English languages resource-scarce languages in this work.We propose Tower of Babel (ToB), a crowdsourcing game building sentiment lexicons for the resource-scarce languages. ToB aims to lower the costs for building the resources for sentiment analysis by crowdsourcing and gamifying manual annotation process which has been the best yet costly and inefficient practice for building the resources. ToB builds the sentiment lexicon over other types of sentiment resources since the sentiment lexicon is the most computationally convenient and generalizable. We conducted an experiment w...