Statistical Models for Linguistic Variation in Social Media

Language on the Internet and social media varies due to time, geography and social factors. For example, consider an online chat forum where people from different regions across the world interact. In such scenarios, it is important to track and detect regional variation in language. A person from the UK, who is in conversation with someone from the USA could say “he is stuck in the lift” to mean ”he is stuck in an elevator”, since the word lift means an elevator in the UK. Note that in the US, lift does not refer to an elevator. Modeling such variation can allow for applications to prompt or suggest the intended meaning to the other participants of the conversation. In this thesis I conduct two related lines of inquiry focusing on (a) language itself and the variation it manifests and (b) the user and what we can infer about them based on their language use on social media. First I develop computational methods to track and detect changes in word usage, including semantic and syntactic variation. I examine two modalities: time and geography. Specifically I outline methods to use distributional word representations (word embeddings) to detect semantic variation in word usage. Our methods are scalable to large datasets, making them particularly suited for social media. Second, I turn my attention towards users. In particular, I seek to model latent personality traits of users based on their language use on social media. I propose to develop generative latent factor models, that explicitly seek to build representations of each user based on their inferred latent personality traits. These models seek to capture latent personality traits that serve as useful co-variates for a wide variety of tasks like predicting what topics users like on social media and the number of friends in their social circle. This work has broad applications in several fields like information retrieval, semantic web applications, socio-variational linguistics, computational social science including digital health care, psycho-linguistics and ad-targeting. April 18, 2016 DRAFT