跳转至

填补鸿沟：用于统一多任务用户意图推理的双知识图谱框架

文章背景与核心概要

在在线旅游评论分析中，准确且统一地推断用户意图和预测细粒度评分长期以来面临着巨大挑战。传统方法通常依赖分层流水线，容易引发严重的错误传播；或者采用标准检索技术，却忽视了领域知识中的结构化关系。

为了克服这些局限性，本文提出了一种创新的双知识图谱框架——DKG-MTI。该框架通过纯推理的知识增强方法，动态构建“用户特定意图知识图谱”与“全局酒店知识图谱”，并利用结构-感知语义平滑技术将二者对齐。结合大语言模型，DKG-MTI能够同时且高精度地预测方面评分（aspect ratings）并生成反向用户意图语句，在可解释性和预测性能上均显著优于现有基线。

执行摘要 (Executive Summary)

本论文介绍了 DKG-MTI，一个旨在解决基于在线旅游评论的统一多任务用户意图推理挑战的创新双知识图谱框架。传统方法常常在分层流水线中遭受错误传播的困扰，或者在使用标准检索技术时忽略了结构化关系。DKG-MTI通过一种仅推理的知识 augmentation（增强）方法解决了这些问题，该方法通过结构感知的语义平滑，动态构建并对齐了用户特定意图知识图谱与全局酒店知识图谱。结合大语言模型，该方法有效地预测了方面评分并以高准确性和可解释性生成了反向用户意图语句。

This paper introduces DKG-MTI, an innovative dual knowledge graph framework designed to solve the challenges of unified multi-task user intent inference using online travel reviews. Traditional methods often suffer from error propagation in hierarchical pipelines or ignore structural relationships when using standard retrieval techniques. DKG-MTI resolves these issues through an inference-only knowledge augmentation approach that dynamically builds and aligns a User-Specific Intent Knowledge Graph with a Global Hotel Knowledge Graph via structure-aware semantic smoothing. Combined with large language models, this approach effectively predicts aspect ratings and generates reverse user intent statements with high accuracy and explainability.

元数据 (Metadata)

  • arXiv 标识符: [arXiv:2608.06752 [cs.AI]] (https://arxiv.org/abs/2608.06752)
  • 主学科: 人工智能 (cs.AI)
  • 次学科: 计算与语言 (cs.CL)
  • 提交日期: 2026年8月7日
  • 期刊参考: 太平洋亚洲信息系统会议论文集 (PACIS 2026), 论文 12, 17页, 5张图表。
  • AIS 电子图书馆链接: PACIS 2026 Paper 12
  • arXiv Identifier: [arXiv:2608.06752 [cs.AI]] (https://arxiv.org/abs/2608.06752)
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Secondary Subject: Computation and Language (cs.CL)
  • Submission Date: August 7, 2026
  • Journal Reference: Proceedings of the Pacific Asia Conference on Information Systems (PACIS 2026), Paper 12, 17 pages, 5 figures.
  • AIS eLibrary Link: PACIS 2026 Paper 12

作者 (Authors)

  1. Tzu-Cheng Peng — 国立台湾大学 (National Taiwan University)
  2. Chien Chin Chen — 国立台湾大学 (National Taiwan University)
  3. Chih-Hao Ku — 北德克萨斯大学 (University of North Texas)
  4. Yung-Chun Chang — 台北医学大学 (Taipei Medical University)
  1. Tzu-Cheng Peng — National Taiwan University
  2. Chien Chin Chen — National Taiwan University
  3. Chih-Hao Ku — University of North Texas
  4. Yung-Chun Chang — Taipei Medical University

摘要 (Abstract)

本文提出了 DKG-MTI，这是一个用于从在线旅游评论中进行统一多任务用户意图推理的双知识图谱框架。现有方法通常依赖于分层流水线（容易导致错误传播），或者依赖于忽略领域知识中结构化关系的检索方法。

为了解决这些局限性，作者引入了一个纯推理的知识增强框架，该框架从每条评论中动态构建一个用户特定意图知识图谱，并通过结构感知的语义平滑将其与全局酒店知识图谱进行对齐。对齐后的知识与原始评论相结合，由大语言模型进行处理，从而同时预测方面评分并生成反向用户意图语句。在 TripAdvisor 评论上的实验表明，DKG-MTI 在分类和意图生成任务上均持续优于强大的 LLM 和基于检索的基线，展示了结构感知知识对齐在可扩展且可解释的意图推理中的有效性。

This paper proposes DKG-MTI, a dual knowledge graph framework for unified multi-task user intent inference from online travel reviews. Existing approaches often rely on hierarchical pipelines that suffer from error propagation or retrieval methods that ignore structural relationships in domain knowledge.

To address these limitations, the authors introduce an inference-only knowledge augmentation framework that dynamically constructs a User-Specific Intent Knowledge Graph from each review and aligns it with a Global Hotel Knowledge Graph through structure-aware semantic smoothing. The aligned knowledge is combined with the original review and processed by a large language model to simultaneously predict aspect ratings and generate reverse user intent statements. Experiments on TripAdvisor reviews show that DKG-MTI consistently outperforms strong LLM and retrieval-based baselines in both classification and intent generation tasks, demonstrating the effectiveness of structure-aware knowledge alignment for scalable and explainable intent inference.

核心框架组件 (Key Framework Components)

  • 动态用户特定意图知识图谱： 从每条在线评论中单独提取，以捕捉微妙的用户情感和意图。
  • 全局酒店知识图谱： 封装了关于酒店和服务的全领域结构化关系。
  • 结构感知语义平滑： 将用户特定图谱与全局图谱对齐，从而弥合知识鸿沟。
  • 统一多任务推理： 利用大语言模型（LLMs）联合预测细粒度方面评分并生成反向用户意图语句。
  • Dynamic User-Specific Intent Knowledge Graph: Extracted individually from each online review to capture nuanced user sentiment and intent.
  • Global Hotel Knowledge Graph: Encapsulates domain-wide structural relationships regarding hotels and services.
  • Structure-Aware Semantic Smoothing: Aligns the user-specific graph with the global graph, mitigating knowledge gaps.
  • Unified Multi-Task Inference: Leverages large language models (LLMs) to jointly predict granular aspect ratings and generate reverse user intent statements.

全文与外部资源 (Full-Text and External Resources)