Invited talk from Julian Brooke

The UKP Lab will host a talk by Julian Brooke of the University of Toronto. The talk, titled “Approaches to lexical induction of multiple correlated styles”, will be held on 30 July at 12:20 in S1|03 223 (Hochschulstraße 1) at Technische Universität Darmstadt. The abstract and speaker biography are as follows:

Abstract: Lexical induction of individual stylistic aspects such as concreteness (Turney et al., 2011) or formality (Brooke et al., 2010) from corpora ignore the strong positive and negative correlations among styles, oriented around the oral/written spectrum, which has been identified by work in the study of register (Leckie-Tarry, 1995). Using Bayesian topic models, i.e. latent Dirichlet allocation (LDA) and correlated topic models (CTM), to induce 6 styles, we demonstrate that this is indeed an issue, but that explicitly modelling these correlations when extracting information from the corpus does not appear to resolve it. Instead, we construct a hybrid model that efficiently extracts initial expectations for each style from the corpus using any of various possible methods (LSA, LDA, or PMI), and then refines this information based on the relationships between styles, using label propagation in a graph. We seed our model and evaluate using a new annotation of 900 English words, tagged for all the styles of interest.

Biography: Julian Brooke is a PhD student in the computational linguistics group at the University of Toronto, working with Graeme Hirst. He has a Master's in Linguistics from Simon Fraser University, and a Bachelor's in Symbolic Systems from Stanford University. His research interests include lexical resources, stylistics, sentiment analysis, native language detection, discourse, semantics, and social media. He is particularly interested in multidisciplinary applications of CL, for instance in the fields of education, sociolinguistics, and literary analysis.