Microblog Language Identification: Overcoming the Limitations of Short, Unedited and Idiomatic Text

Multilingual posts can potentially affect the outcomes of content analysis on microblog platforms. To this end, language identification can provide a monolingual set of content for analysis. We find the unedited and idiomatic language of microblogs to be challenging for state-of-the-art language identification methods. To account for this, we identify five microblog characteristics that can help in language identification: the language profile of the blogger (blogger), the content of an attached hyperlink (link), the language profile of other users mentioned (mention) in the post, the language profile of a tag (tag), and the language of the original post (conversation), if the post we examine is a reply. Further, we present methods that combine these priors in a post-dependent and post-independent way. We present test results on 1,000 posts from five languages (Dutch, English, French, German, and Spanish), which show that our priors improve accuracy by 5 % over a domain specific baseline, and show that post-dependent combination of the priors achieves the best performance. When suitable training data does not exist, our methods still outperform a domain unspecific baseline. We conclude with an examination of the language distribution of a million tweets, along with temporal analysis, the usage of twitter features across languages, and a correlation study between classifications made and geo-location and language metadata fields.