Personalized Language Models for Computer-mediated Communication

In this paper, we investigate the performance of statistical language models on Instant Messaging (IM) data. Language Models (LM) are quite useful for modeling text data, and hence they are helpful in different contexts like spelling correction, speech recognition, part-ofspeech tagging etc. Construction of LM on a users past messaging data would be a strategy to model her writing style, and that LM can then be used to predict the next word in her future communications. However, we hypothesize that a user follows a specific pattern of communication with each of her virtual acquaintances. As a consequence, LM built on her entire messaging history would degrade the performance of the next word predictor, while communicating with a specific person. In this paper, we deploy a special method that excludes some specific message contents from the entire history in order to build LM. Our method suggests that, at the time of communicating with a specific user, a special LM should be invoked from a set of models for increasing accuracy. We analyze the IM data of a set of users, and show that our method performs well in terms of perplexity.