Optimizing Interpretation Generation in Natural Language Query Answering for Real Time End Users

Abstract

Natural Language Querying over Database is gaining popularity across different use cases. Most common of them is to democratize the process of data analysis and querying of backend data to naive end users especially business users, obviating the need of knowing back end query language. Natural Language Query answering systems have thus seen widespread usage in industry too where business users want to search their own data to make business decisions. However, a common challenge faced by any natural language query answering system is generation of precise interpretations. The research community although tries to handle the problem via asking clarification questions back to the user, in industry setup this remains an ineffective solution due to various practical usage limitations. For example, it is not fair to assume any end user will be aware of the correct option to answer these clarification questions. Moreover, involving clarification questions and user feedbacks makes the system unusable by one shot API calls, which is the most intuitive usage among common use cases in industry like automated report generation. In this paper, we investigate practical ways to address the problem of precise interpretation generation. We propose novel algorithms to make use of existing technologies like Functional Partitioning of Ontology and Lazy Inclusion to solve this problem. We take our previous state-of-the-art paper ATHENA and further extend it to include our proposed methods. We test with 3 benchmark ontologies to empirically demonstrate the huge improvement over state-of-the-art results by factors of at least 400% in number of interpretation generation and also in the computation time.