Question answering and query processing for extended knowledge graphs

Abstract

Knowledge graphs have seen wide adoption, in large part owing to their schemaless nature that enables them to grow seamlessly, allowing for new relationships and entities as needed. With this rapid growth, several issues arise: (i) how to allow users to query knowledge graphs in an expressive and user-friendly manner, which shields them from all the underlying complexity, (ii) how, given a structured query, can we return satisfactory answers to the user despite possible mismatches between the query vocabulary and structure and the knowledge graph, and (iii) how to automatically acquire new knowledge, which can be fed into a knowledge graph. In this dissertation, we make the following contributions to address the above issues: – We present DEANNA, a framework for question answering over knowledge graphs, allowing users to easily express complex information needs using natural language and obtain tuples of entities as answers thereby taking advantage of the structure in the knowledge graph. – We introduce TriniT, a framework that compensates for unsatisfactory results of structured queries over knowledge graphs, either due to mismatches with the knowledge graph or the knowledge graph’s inevitable incompleteness. TriniT tackles the two issues by extending the knowledge graph using information extraction over textual corpora, and supporting query relaxation where a user’s query is rewritten in a manner transparent to the user to compensate for any mismatches with the data. – We present ReNoun, an open information extraction framework for extracting binary relations mediated by noun phrases and their instances from text. Our scheme extends the state-of-the-art in open information extraction which has thus far focused on relations mediated by verbs. Our experimental evaluations of each of the above contributions demonstrate the effectiveness of our methods in comparison to state-of-the-art approaches.