We consider the problem of constructing personalized symbolic knowledge base (KB) through natural language instructions. This problem presents several challenges, including (1) integrating symbolic knowledge from the evolving KB with user utterances to produce the appropriate KB modification commands, and (2) handling open domain utterances that may, e.g., introduce new entities at test time. We design alternative neural network encoder-decoder models that combine the unstructured context from the utterance with the structured context from the KB. Empirical results and analysis show that our models are able to construct the knowledge bases from user utterances with high accuracy. We also contribute an evaluation dataset, and perform detailed analysis that reveals interesting properties when applying neural models on this task.

38 pages

Thesis Committee:
Matthew Gormley (Chair)
Tom Mitchell

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu