Multi-hop Question Answering with Knowledge Graph Embedding in a Similar Semantic Space

Multi-hop Question Answering using the knowledge graph (KG) as a data source requires subject entities and relations that are obtained from natural language questions; the answers are then obtained by reasoning through multiple triples in the KG. However, even large KGs are incomplete, and the reasoning often fails to obtain the correct answer due to the lack of relations in the KG. Recently, researchers have proposed introducing KG embedding into the multi-hop knowledge graph question answering (KGQA) field to solve the incompleteness of the KG. However, these methods embed the question and the KG into different semantic spaces, and it is difficult to obtain the correct answers. Furthermore, due to the limitation of the question embedding sequence, the contribution of each word to the question semantics cannot be distinguished. To overcome the above problems, this paper proposes an effective multi-hop KGQA model, TIPNet, using relation embeddings in knowledge graph triples, which uses the idea of translation models to narrow the semantic spatial distance between question embeddings and KG embeddings. At the same time, TIP weighting technology is proposed to distinguish the semantic contribution of words to the question. To validate the performance of TIPNet, experiments were conducted on the WebQSP and CWQ datasets, and the model reached advanced levels under both KG-full and KG-half settings.