BMQA: Leveraging Large Language Models for Semantic Guidance in Temporal Knowledge Graph Q&A

Temporal Knowledge Graph Q&A (TKGQA) faces significant challenges in handling queries across temporal granularities due to the complex relationships among four elements (entity, relation, time, and query). Although existing models introduce multi-granularity temporal aggregation mechanisms, they suffer from pervasive semantic space mismatch problems between upstream, knowledge graph embedding (KGE) and downstream, questions and answers (QA) tasks, limiting their effectiveness in TKGQA scenarios. To address these issues, we propose the BMQA model, which leverages large language models (LLMs) as semantic guidance for initialization during the KGE phase and introduces semantic supervision during the QA phase. LLM-based guidance is applied in the problem preprocessing stage to better align entity and relation embeddings, while the semantic loss function supervises feature deviations during training, thereby reducing semantic inconsistencies. Experimental results on the public dataset demonstrate that BMQA outperforms state-of-the-art baseline models, effectively mitigating semantic space mismatch problems while preserving strong temporal reasoning capabilities.