Knowledge-Graph based Question Answering by Graph Pattern Learning

Abstract

Knowledge Graph Question Answering (KGQA) aims to leverage knowledge graphs (KGs) to answer questions posed in natural langauge. It assists end users to more easily and more efficiently access the vast amounts of information available in KGs, without the need for knowledge about internal data structures or KG query languages. The aim of this master’s thesis is to evaluate a KGQA system which utilizes graph patterns pre-generated with a Graph Pattern Learner (GPL) to retrieve answer candidates. The GPL operates on a set of source-target entity pairs and a SPARQL endpoint to generate graph patterns using an evolutionary algorithm, which provides high versatility with regards to the underlying knowledge bases. The main contributions of this thesis are a GPL based KGQA system "GTFQ " (GPLand Target candidate FRED scoring-based Question answering) and a novel target candidate scoring function "FRED" (F1 pREDictor), which makes significant use of transfer learning: The FRED model utilizes a pre-trained BERT NLP model to retrieve dense question representations and operates on coverage based graph pattern embeddings. The FRED model matches the GPL graph patterns against the natural language question to score the answer entities retrieved with the patterns. The generated graph patterns are evaluated independently of FRED and also GTFQ is evaluated against a state-of-the-art approach and against simple scoring functions that do incorporate knowledge of the question. Promising experimental results on subsets of WebQuestionsSP and SimpleQuestions provide a reference point for future work in this area.