Language identification of multilingual posts from Twitter: a case study

Abstract

This paper describes a method for handling multi-class and multi-label classification problems based on the support vector machine formalism. This method has been applied to the language identification problem in Twitter. The system evaluation was performed mainly on a Twitter data set developed in the TweetLID workshop. This data set contains bilingual tweets written in the most commonly used Iberian languages (i.e., Spanish, Portuguese, Catalan, Basque, and Galician) as well as the English language. We address the following problems: (1) social media texts. We propose a suitable tokenization that processes the peculiarities of Twitter; (2) multilingual tweets. Since a tweet can belong to more than one language, we need to use a multi-class and multi-label classifier; (3) similar languages. We study the main confusions among similar languages; and (4) unbalanced classes. We propose threshold-based strategy to favor classes with less data. We have also studied the use of Wikipedia and the addition of new tweets in order to increase the training data set. Additionally, we have tested our system on Bergsma corpus, a collection of tweets in nine languages, focusing on confusable languages using the Cyrillic, Arabic, and Devanagari alphabets. To our knowledge, we obtained the best results published on the TweetLID data set and results that are in line with the best results published on Bergsma data set.