A convex model for linguistic influence in group conversations

Conversational partners can influence each other's speaking patterns. In this paper, we aim to develop a computational model that infers influence levels directly from language samples. We propose a new approach to modeling linguistic influence in conversations based on a well-accepted model of social influence. Very generally, this approach assumes that an individual's language model can be expressed as a convex combination of language models from individuals with whom that person interacts. We propose an optimization criterion to estimate the pairwise influence between conversational partners directly from speech and language data. We evaluate the model on three different corpora: (1) a synthetic corpus where the language influence is experimentally set; (2) a corpus that tracks a child's interaction with her family during the early stages of language development; (3) a corpus of Supreme Court cases analyzing interactions between judges and attorneys.