Fluency detection on communication networks

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1025–1029, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Fluency detection on communication networks Tom Lippincott and Benjamin Van Durme Human Language Technology Center of Excellence Johns Hopkins University tom@cs.jhu.edu, vandurme@cs.jhu.edu Abstract When considering a social media corpus, we often have access to structural information about how messages are flowing between peo- ple or organizations. This information is par- ticularly useful when the linguistic evidence is sparse, incomplete, or of dubious quality. In this paper we construct a simple model to leverage the structure of Twitter data to help determine the set of languages each user is flu- ent in. Our results demonstrate that imposing several intuitive constraints leads to improve- ments in performance and stability. We re- lease the first annotated data set for exploring this task, and discuss how our approach may be extended to other applications. 1 Introduction Language identification (LID) is an important first step in many NLP pipelines since most downstream tasks need to employ language-specific resources. In many situations, LID is a trivial task that can be addressed e.g. by a simple Naive Bayes classifier trained on word and character n-gram data (Lui and Baldwin, 2012): a document of significant length will be quickly disambiguated based on its vocab- ulary (King et al., 2014). However, social media platforms like Twitter produce data sets in which individual documents are extremely short, and lan- guage use is idiosyncratic: LID performance on such data is dramatically lower than on traditional corpora (Bergsma et al., 2012; Carter et al., 2013). The widespread adoption of social media throughout the world amplifies the problem as less-studied lan- guages lack the annotated resources needed to train the most effective NLP models (e.g. treebanks for statistical parsing, tagged corpora for part-of-speech tagging, etc). All of this motivates the research com- munity’s continued interest in LID (Zampieri et al., 2014). Tweet #1 Êîç ýâåðèñèíã þ äó èñ ìýäæèê Tweet #2 omg favourite day of the week! Table 1: Multilingual social media users often communicate in different languages depending on their intended audience, such as with these Russian and English tweets posted by the same Twitter account In this paper, we consider the closely-related task of determining an actor’s fluencies, the set of lan- guages they are capable of speaking and understand- ing. The observed language data will be the same as for LID, but is now considered to indicate a latent property of the actor. This information has a num- ber of downstream uses, such as providing a strong prior on the language of the actor’s future communi- cations, constructing monolingual data sets, and rec- ommending appropriate content for display or fur- ther processing. This paper also focuses on the situation where a very small amount of content has been observed from the particular user. While this may seem strange considering the volume of data generated by social media, this is dominated by particularly active users: for example, …

Content truncated to 500 words