跳转至

文章背景与核心概要

商品目录对于电子商务的搜索、发现和推荐系统至关重要,但往往面临属性稀疏的问题,导致关键产品细节淹没在非结构化的文本和图像中,甚至完全缺失。由于手动进行目录增强在规模化应用中不可行,本文提出了 TRACE,这是一个利用代理式大语言模型(LLM)实现商品目录属性自动增强的框架。

该框架采用协作式多智能体设计:ScoutAgent 负责跨商家目录、联合数据源和身份匹配的网络搜索进行多模态证据三角测量,从而提出带有支持证据的候选属性值;JudgeAgent 则负责根据收集到的证据验证这些候选值,并决定是自动发布还是转交人工审核。

实验结果表明,TRACE 在离线人工评估数据集上达到了 98.2% 的准确率和 74.7% 的属性覆盖率。在工业级目录的生产部署中,TRACE 使四个业务垂直领域的加权展示增强覆盖率提升了 90.4%。在线实验进一步证明,在产品详情页展示这些新增强的属性,使结账转化率提高了 0.48%。


TRACE:基于多源证据溯源的代理式商品目录增强框架

摘要

Product catalogs are vital for e-commerce search, discovery, and recommendations, but they often suffer from sparse attributes—leaving essential product details buried in unstructured text and images or missing entirely. Manual enrichment is impractical at scale. This paper introduces TRACE, an automated framework for catalog attribute enrichment using agentic Large Language Models (LLMs).

该框架采用协作式多智能体设计:

The framework employs a cooperative multi-agent design:

  • ScoutAgent: 跨商家目录、联合数据源和身份匹配的网络搜索进行多模态证据三角测量,以提出带有支持证据的候选属性值。

    • ScoutAgent: Triangulates multimodal evidence across merchant catalogs, syndicated feeds, and identity-matched web searches to propose candidate attribute values with supporting evidence.
  • JudgeAgent: 根据收集到的证据验证提出的值,决定是自动发布还是转交人工审核。

    • JudgeAgent: Verifies the proposed values against the collected evidence, deciding whether to publish them automatically or route them for human review.

关键结果

Key Results

  • 准确性与覆盖率: 在离线人工评估数据集上实现了 98.2% 的准确率74.7% 的属性覆盖率

    • Accuracy & Coverage: Achieved 98.2% accuracy at 74.7% attribute coverage on an offline human evaluation dataset.
  • 生产影响: 在工业级目录中部署后,TRACE 使四个业务垂直领域的加权展示增强覆盖率提升了 90.4%

    • Production Impact: Deployed across an industry-scale catalog, TRACE increased impression-weighted enrichment coverage by 90.4% across four business verticals.
  • 商业价值: 在线实验表明,在产品详情页展示新增强的属性使结账转化率提高了 0.48%

    • Business Value: Online experiments demonstrated that surfacing the newly enriched attributes on product detail pages increased checkout conversion by 0.48%.

论文元数据

Paper Metadata

  • arXiv ID: arXiv:2608.20844 [cs.AI]

  • 主要学科: 计算机科学 > 人工智能 (cs.AI)

    • Primary Subject: Computer Science > Artificial Intelligence (cs.AI)
  • 提交日期: 2026年8月21日

    • Submission Date: August 21, 2026
  • 作者: Rohan Kumar, Steven Xu, Kyle MacDonald, Matthew Long, Bernice Chow, Mac VanRenterghem, Sudeep Das

    • Authors: Rohan Kumar, Steven Xu, Kyle MacDonald, Matthew Long, Bernice Chow, Mac VanRenterghem, Sudeep Das
  • 篇幅: 12 页,2 张图表

    • Length: 12 pages, 2 figures

链接与资源