On bridging the semantic gap in knowledge-based question answering

Abstract

A knowledge-based question answering (KB-QA) system is one that answers natural language questions with information stored in a large-scale knowledge base (KB). The classical KB-QA approach involves two steps: semantic parsing and query execution. First, a given natural language question is semantically parsed into a structured meaning representation (e.g., Sparql query). Second, the representation is executed against a KB to retrieve the answer. On the one hand, natural language questions are flexible, diverse and expressive. On the other hand, the knowledge representation models (e.g., RDF data model) used by the back-end KBs are predefined and strictly follow a rigid schema. Such a semantic gap between the flexibility of natural languages and the fragility of knowledge representation models poses great challenge for a KB-QA system to interpret the semantics of an NL question and transform it into its correct meaning representation. In this thesis, we present two novel KB-QA systems that aim to bridge this semantic gap from two respective directions: revised knowledge representation and unified QA framework.

First, we propose TAQA, a KB-QA system powered by a novel knowledge representation model --- n-tuple open KB (nOKB). An nOKB uses semi-structured natural language (NL) assertions to represent knowledge extracted from massive web corpora. Compared with rigid knowledge representation schemas like the RDF data model, NL assertions in an nOKB are structurally flexible and semantically rich to answer questions with complex semantic constraints.

Next, we develop Neural Enquirer, a KB-QA system that jointly models both semantic parsing and query execution in distributional spaces using deep neural networks, instead of treating the two processes as separate and standalone steps. Neural Enquirer derives distributed representations of NL questions and KB tables, and performs query execution through a series of neural network computation. We evaluate Neural Enquirer using a synthetic question answering task, and demonstrate its effectiveness in learning to execute compositional NL questions on small-scale KB tables.