We will hear from Elias Stengel-Eskin who is a new postdoc in the CS department (working with Mohit Bansal), focusing on how AI models represent meaning.
Abstract: The mapping between language and meaning is not always clear. This is especially true of ambiguous utterances, which can have multiple meanings. While ambiguity has been widely studied in linguistics, it has often been ignored or deliberately removed in NLP tasks. I will present two recent papers attempting to bridge the gap between linguistics and NLP in modeling ambiguity. The first focuses on ambiguity in questions about images. Here, we examine questions which drive annotators to disagree in their answers and develop an ontology of ambiguity types. I’ll then describe a model for rewriting questions to resolve ambiguities in this context. The second paper examines ambiguity in semantic parsing, where I will describe a new dataset pairing ambiguous utterances with their (multiple) logical forms. I’ll benchmark the performance of modern semantic parsing models on this dataset and contrast the results with human evaluations and findings from psycholinguistics.