QUEST: A Natural Language Interface to Relational Databases

QUEST: A Natural Language Interface to Relational Databases Vadim Sheinin, Elahe Khorashani, Hangu Yeo, Kun Xu, Ngoc Phuoc An Vo, Octavian Popescu IBM Research, Yorktown Heights, US firstname.lastname@us.ibm.com Abstract Natural language interfaces to databases systems allow the user to use natural language to interrogate a database. Current systems mainly focus on simple queries but neglect nested queries, which are predominant in real cases. We present a NLIDB system, QUEST, which is able to cope with nested logic queries, without imposing any restrictions on the input query. QUEST outperforms a strong baseline system by 11% accuracy. Keywords: natural language interface to relational database, machine learning. 1. Introduction Natural language interfaces to database systems (NLIDB) allow the user to use natural language to interrogate a database. This is an old desideratum, but there is no easy way to attain it. To begin with, even for simple queries that can be solved into a single SQL select query, namely Simple Logic Queries (SLQs), may be challenging to get the SQL query because of different joins needed to be ex- ecuted. Moreover, the queries that need a succession of select in SQL, namely Nested Logic Queries (NLQs), are very difficult to be handled automatically. NLQs usually display a wide range of linguistic phenomena like ellipsis and anaphora, garden path parsing, etc. Due to this, many systems tend to impose strict limitations on the input query language, in a way that makes the extraction of logical re- lationships controllable. But this is not actually a very real- istic scenario. There is a large literature on this research field describing systems that are purely rule based, semi-supervised, and unsupervised. We point to excellent overviews of such sys- tems in (Copestake and Jones, 1990; Androutsopoulos et al., 1995; Li and Jagadish, 2014). The recent developments in deep learning algorithms seem to bring back into the foreground the NLIDB. Working on related topics, such as learning inferences over constrained outputs, global inference in natural languages, open do- main relation extraction (Schwartz et al., 2014) , or more recently, resolving algebraic problems expressed in natural language (Koncel-Kedziorski et al., 2015), provide a strong ground for approaching the NLIDB by taking advantage of the power of semi-supervised and unsupervised tech- niques. However, most systems did not deal with nested logic queries. In this paper we present a semi-supervised methodology to deal with NLQs in the most general setting: (i) we do not presuppose any restriction on the input query, and (ii) we do not presuppose any restriction on the structure of the database. From a practical point of view, we do not rely on large training data, but rather we build on different tech- nologies that carry out linguistic processing of a query up to a point where a specialized engine is able to produce a series of SQL queries that return the answer in a recur- sive bottom-up process of computation. We developed a system, QUEST, on top of IBM Watson UIMA pipeline (McCord et al., 2012) and Cognos1, …

Content truncated to 500 words