Improve the range and quality of answers to challenging event-related questions

Background: Personally, I am most attracted to interdisciplinary, applied research: the cross-pollination possible in the spaces between traditional research areas increases the potential for significant achievements, and the applied nature makes the time to impact relatively short. When I entered the Berkeley Computer Science department, I joined the Neural Theory of Language (NTL) group, specifically because of its combination of disciplines: artificial intelligence, cognitive linguistics, cognitive science, and neurobiology. Within the group, I chose to focus my doctoral studies on the structure of events and their ties to language in order to better answer heretofore unanswerable-but-common question types.

To give a gist of my Question Answering research, consider the following example: Let's say you wanted to know which car some guy, Joe, bought. You might ask Google, 'what car did Joe buy?' Google, a keyword-based search engine, would return to you a list of documents which each contain the words "car," "Joe," and "buy" - probably a lot of documents, most irrelevant. The chance that a document that says, "Joe is going to buy a great car, a Ferrari" exists and is one of the top results, is very slim. Really, you'd like to be able to find documents that say, "Joe owns a Ferrari," or even "Joe crashed his Ferrari." To do so, the system you use has to understand the relationship between buying and ownership, and, in the context of cars, buying and crashing. The system we created uses information about event structure and relationships to find relevant data and infer information of interest.

My advisors for this work were Srini Narayanan, head of the ICSI AI group, and Jerry Feldman, PI of the NTL research group. I am grateful for their support, as well as that of our funders at IARPA.

Status: Doctorate completed. Plenty of future work for others.

Dissertation Abstract

Reasoning about event structure is a fundamental research problem in Artificial Intelligence. Event scenarios and procedures are inherently about change of state. To understand them and answer questions about them requires a means of describing, simulating and analyzing the underlying processes, taking into account preconditions and effects, the resources they produce and consume, and their interactions with each other. We propose a novel, comprehensive event schema that covers many of the parameters required and has explicit links to language through FrameNet. Based on the event schema, we have implemented a dynamic model of events capable of simulation and causal inference. We describe the results of applying this event reasoning platform to question answering and system diagnosis, providing responses to questions on justification, temporal projection, ability and 'what-if' hypotheticals, as well as complex problems in diagnosis of systems with incomplete knowledge.

Full Text: Answering Questions about Complex Events

Related Projects:
Artificial Intelligence - background knowledge