Learning to Negotiate Answers in Multi-Pass Semantic Parsing
Summary: The true answer for a geographical query against a database such as OpenStreetMap (OSM) often lies in the eye of the beholder who asked the original question - answers can be open-ended lits (e.g., of restaurants) or fuzzily defined geo-positional objects (e.g., objects “near” or “in walking distance” of other objects). Machine learning for information access from real-word databases such as OSM has to go beyond standard supervised learning, and instead asks for an interactive setup where an end-user inter-operates with a semantic parsing system, for example, by providing feedback signals on erroneous parts of the semantic parse, through an engagement in a clarification dialogue, leading to improved parses, and thereby negotiating a correct answer in a dialogical process. Our project aims to identify techniques to learn semantic parsers from user natural language interactions, with possible wider ramifications on interactive learning for personalized assistants.