Computer Science Seminar
Mohit Bansal
Toyota Technological Institute at Chicago
Learning Weakly-Labeled and Cross-modal Semantics for Structured NLP
Abstract: For artificial intelligent agents to function autonomously in our homes
and workplaces, they must be able to effectively understand natural
language for instructions and conversational dialog. For this, it is
necessary to resolve the various deep semantic and pragmatic ambiguities
that exist in everyday natural language, a challenging goal that involves
two fundamental requirements. First, we need diverse, external world
knowledge which is simply not present in the standard training datasets
used for supervised NLP tasks. Second, we need to develop appropriate
machine learning models that can extract the precise disambiguation cues
lying latent in such diverse, large-scale knowledge datasets. My research
addresses both these requirements by learning novel weakly-labeled and
cross-modal semantic representations with accurate, well-formulated
disambiguation methods. We model world knowledge via unlabeled Web-scale
features, weakly-supervised language embeddings, and grounding cues from
vision and speech, and harness these in both structured and neural
network based learning methods to achieve the state-of-the-art on various
core NLP tasks and multimodal applications.
Monday February 15, 2016 at 3:00 PM in SEO 612