Title: Crowdsourcing Semantic Structure
Advisors: Luke Zettlemoyer and Yejin Choi
Abstract: End-to-end neural models have pushed the state-of-the-art on many natural language understanding tasks, but they are generally data-hungry and prone to overfitting. Efforts to fix this often involve using task-general semantic structures as intermediate representations, but gains from this are rare because the models to produce the intermediates often suffer from the same problems of generalization and efficiency.
Building effective models for semantic intermediates is especially difficult because annotating the training data is expensive and time-consuming. As a result, data annotated with semantics is often much smaller than the available data for end tasks, limiting its usefulness. However, recent work has explored creative methods to scale up semantic annotation with non-experts to cheaply and rapidly annotate linguistic structure.
In this talk, I describe crowdsourcing projects on three such approaches: Human-in-the-loop Parsing, Question-Answer Meaning Representation, and Question-Answer Driven Semantic Role Labeling. I outline the key lessons that have emerged from this research and a trajectory for future work on crowdsourcing semantic structure.
Place:
CSE 303
When:
Tuesday, February 27, 2018 - 14:30 to Thursday, April 25, 2024 - 06:27