Doctoral Dissertation Research: Disambiguating Information in Speech and Context

University Of California-Irvine, Irvine CA

Investigators

Abstract

Language serves as a remarkable tool for human communication, but in everyday conversations, humans often need to navigate ambiguity when multiple potential meanings are possible. How do listeners successfully understand what a speaker meant? One intuition is that listeners use clues from context (the preceding conversation) and prosody (the melodic and rhythmic elements of speech) to decipher the intended meaning. This study investigates the relationship between these two information sources, asking whether (i) context and prosody provide two different streams of information, or if (ii) context is the main source of information while prosody reflects the context. To better understand how context and prosody help speakers resolve ambiguity, this doctoral dissertation research takes advantage of the large-scale amounts of data on everyday language use that have only recently become available. Specifically, this project builds a unique corpus of naturalistic ambiguity that captures the attested and diverse sentences people really say. First, researchers search through archives of conversation recordings from public radio and TV to find instances of potential ambiguity, which is then used to create an Ambiguity Corpus, preserving each case with its original context and the audio recording that contains the prosody. For each case, human annotators judge what they thought the speaker meant. Judgments are provided for just the transcript, just the recording, the transcript in its context, and the recording in its context. By comparing people's interpretations, the researchers compare the roles of context and prosody on these interpretations. To further investigate hypotheses about how context and prosody work together to resolve ambiguity in everyday conversations, the research team builds a computational cognitive model that simulates how someone interprets ambiguous utterances like the ones identified in the Ambiguity Corpus, given these information sources. To evaluate the hypotheses about context and prosody that this model encodes, follow-up experiments are conducted that test specific model predictions about how context, prosody, and their interaction matter for interpretations. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.