Enhancing Question Embedding with Relation Chain for Multi-hop KGQA

Knowledge Graph (KG) is a semantic network that describes entities or concepts in reality and the relationships among them. The Knowledge Graph Question Answering (KGQA) task aims to provide answers to natural language (NL) questions posed over the KG. Multi-hop KGQA requires reasoning across multiple relational edges to obtain correct answers. However, KGs are consistently incomplete, and the absence of certain relational edges poses significant challenges for multi-hop KGQA. Inspired by the successful application of Prompt Tuning in the field of natural language processing (NLP), we propose to utilize relevant relation chain information to guide NL question embedding. We introduce QERChain (Question Embedding enhanced by Relation Chain) into multi-hop KGQA. Initially, the question is embedded using RoBERTa, then aligned with the knowledge graph embedding space via a GRU network. Subsequently, this aligned question embedding, along with the embeddings of relevant candidate relations, is fed into a transformer encoder. Through an attention mechanism, the semantic information of the question is further integrated with the embedding information of the relation chain. This results in a more appropriate representation of the question in the knowledge graph embedding space. Our extensive experimentation across multiple benchmark datasets has shown that QERChain-KGQA achieves state-of-the-art performance in both MetaQA (Wikimovies-KG) and WebQSP (Freebase), irrespective of the completeness of the knowledge graph (KG). Most importantly, QERChain-KGQA significantly outperforms the original EmbedKGQA model, indicating that relation chain-enhanced question embedding holds profound significance in embedded Q &A methods.