Improving Question Answering over Knowledge Graphs with a Chunked Learning Network

Abstract

The objective of knowledge graph question answering is to assist users in answering questions by utilizing the information stored within the graph. Users are not required to comprehend the underlying data structure. This is a difficult task because, on the one hand, correctly understanding the semantics of a problem is difficult for machines. On the other hand, the growing knowledge graph will inevitably lead to information retrieval errors. Specifically, the question-answering task has three difficulties: word abbreviation, object complement, and entity ambiguity. An object complement means that different entities share the same predicate, and entity ambiguity means that words have different meanings in different contexts. To solve these problems, we propose a novel method named the Chunked Learning Network. It uses different models according to different scenarios to obtain a vector representation of the topic entity and relation in the question. The answer entity representation that yields the closest fact triplet, according to a joint distance metric, is returned as the answer. For sentences with an object complement, we use dependency parsing to construct dependency relationships between words to obtain more accurate vector representations. Experiments demonstrate the effectiveness of our method.