Graph Based Pattern Classification for NLU and Slot Filling: Approach and Analysis

Abstract

In Natural Language Understanding, semantic parsing refers to the task of extracting meaningful words from text, with slot filling being a special case. At the same time, syntactic parsing deals with the identification of syntactic structure in the text. In this paper, we analyze to which extent syntactic structure can be used for semantic parsing. For this purpose, we represent the syntactic structure of sentences from an annotated database as graphs, incrementally store a large number of graph prototypes in a knowledge base, and approach the slot filling problem as a graph matching problem. We analyze how this approach scales with the number of prototypes and show that it provides a lightweight framework that can be efficiently used to support semantic parsing tasks.