Dual Knowledge Graph Boosts User Intent Inference

This paper introduces DKG-MTI, a novel dual knowledge graph framework designed to improve user intent inference from online reviews, specifically focusing on the travel domain. Current methods often suffer from issues like error propagation in hierarchical pipelines or a failure to leverage structural relationships within domain knowledge when using retrieval-based approaches. DKG-MTI aims to address these limitations by offering a more unified and robust solution. The framework operates by dynamically building a User-Specific Intent Knowledge Graph from each individual review. This personalized graph is then aligned with a pre-existing Global Hotel Knowledge Graph through a process called structure-aware semantic smoothing. This alignment integrates the specific user's expressed intent with broader domain knowledge. The combined and augmented knowledge, along with the original review, is then fed into a large language model. This integrated approach allows DKG-MTI to simultaneously predict aspect ratings (e.g., cleanliness, service quality) and generate reverse user intent statements (e.g., "user wanted a quiet room"). Experiments conducted on TripAdvisor reviews demonstrate that DKG-MTI consistently outperforms both strong LLM and retrieval-based baselines in both classification and generation tasks, highlighting the effectiveness of its structure-aware knowledge alignment for scalable and explainable intent inference.