A Semantic Distance of Natural Language Queries Based on Question-Answer Pairs

Abstract

Many Natural Language Processing (NLP) techniques have been applied in the field of Question Answering (QA) for understanding natural language queries. Practical QA systems classify a natural language query into vertical domains, and determine whether it is similar to a question with known or latent answers. Current mobile personal assistant applications process queries, recognized from voice input or translated from cross-lingual queries. Theoretically speaking, all these problems rely on an intuitive notion of semantic distance. However, it is neither definable nor computable. Many studies attempt to approximate such a semantic distance in heuristic ways, for instance, distances based on synonym dictionaries. In this paper, we propose a unified algorithm to approximate the semantic distance by a well-defined information distance theory. The algorithm depends on a pre-constructed data structure - semantic clusters, which is built from 35 million question-answer pairs automatically. From the semantic measurement of questions, we implement two practical NLP systems, including a question classifier and a translation corrector. Then a series of comparison experiments have been conducted on both implementations. Experimental results demonstrate that our distance based approach produces fewer errors in classification, compared with other academic works. Also, our translation correction system achieves significant improvements on the Google translation results.