Ontology-Based Natural Language Query Interfaces for Data Exploration

Enterprises are creating domain-specific knowledge bases by curating and integrating all their business data, structured, unstructured and semi-structured, and using them in enterprise applications to derive better business decisions. One distinct characteristic of these enterprise knowledge bases, compared to the open-domain general purpose knowledge bases like DBpedia [16] and Freebase [6], is their deep domain specialization. This deep domain understanding empowers many applications in various domains, such as health care and finance. Exploring such knowledge bases, and operational data stores requires different querying capabilities. In addition to search, these databases also require very precise structured queries, including aggregations, as well as complex graph queries to understand the various relationships between various entities of the domain. For example, in a financial knowledge base, users may want to find out “which startups raised the most VC funding in the first quarter of 2017”; a very precise query that is best expressed in SQL. The users may also want to find all possible relationships between two specific board members of these startups, a query which is naturally expressed as an all-paths graph query. It is important to note that general purpose knowledge bases could also benefit from different query capabilities, but in this paper we focus on domain-specific knowledge graphs and their query needs. Instead of learning and using many complex query languages, one natural way to query the data in these cases is using natural language interfaces to explore the data. In fact, human interaction with technology through conversational services is making big strides in many application domains in recent years [13]. Such interfaces are very desirable because they do not require the users to learn a complex query language, such as SQL, and the users do not need to know the exact schema of the data, or how it is stored. There are several challenges in building a natural language interface to query data sets. The most difficult task is understanding the semantics of the query, hence the user intent. Early systems [3, 30] allowed only a set of keywords, which had very limited expressive power. There have been works to interpret the semantics of a full-blown English language query. These works in general try to disambiguate among the potentially multiple meanings of the words and their relationships. Some of these are machine-learning based [5, 24, 29] that require good training sets, which are hard to obtain. Others require user feedback [14, 17, 18]. However, excessive user interaction to resolve ambiguities can be detrimental to user experience. In this paper, we describe a unique end-to-end ontology-based system for natural language querying over complex data sets. The system uses domain ontologies, which describe the semantic entities and their relationships, to reason about and capture user intent. To support multiple query types, the system provides a poly store