Doc2Bot

Doc2Bot，一种用于构建机器的新颖数据集，可帮助用户通过对话查找信息。这对于拥有大量手册或说明书的公司和组织特别感兴趣。尽管它具有潜力，但我们任务的性质却带来了一些挑战 :( 1) 文档包含各种结构，这些结构阻碍了机器的理解能力，并且 (2) 用户信息需求通常被低估了。与以前专注于单一结构类型或忽略质疑以发现用户需求的作用的数据集相比，Doc2Bot数据集的开发是为了系统地针对此类挑战。我们的数据集包含基于来自五个域的中文文档的100,000多个回合，比任何以前的基于文档的对话数据集都要大，以进行信息搜索。我们在Doc2Bot中提出了三个任务 :( 1) 对话状态跟踪以跟踪用户意图，(2) 对话策略学习以计划系统操作和内容，以及 (3) 响应生成，该响应生成基于对话策略的输出的响应。提出了基于最新深度学习模型的基线方法，表明我们提出的任务具有挑战性，值得进一步研究。Doc2Bot: a novel dataset for constructing machines that assist users in retrieving information via conversational interactions. This dataset is particularly relevant for companies and organizations that hold extensive collections of manuals and instruction documents. Despite its promising potential, the nature of our task introduces several challenges: (1) Documents exhibit diverse structural formats that hinder machine comprehension, and (2) User information needs are often underestimated. Compared with prior datasets that either focus on a single structural type or neglect the role of questioning in uncovering user needs, the Doc2Bot dataset was developed to systematically address these challenges. Our dataset contains over 100,000 dialogue turns based on Chinese documents across five domains, making it larger than any prior document-based dialogue dataset for information-seeking tasks. We propose three core tasks within Doc2Bot: (1) Dialogue State Tracking to track user intentions, (2) Dialogue Policy Learning to plan system operations and content, and (3) Response generation, which produces responses grounded in the outputs of the dialogue policy. Baseline methods based on state-of-the-art deep learning models are proposed, which confirm that the proposed tasks are challenging and worthy of further research.

搜集汇总

背景与挑战

背景概述

Doc2Bot是一个用于构建机器的新颖数据集，旨在通过对话帮助用户查找信息，特别适用于拥有大量文档的组织。该数据集包含100,000多个基于中文文档的对话回合，针对文档结构多样性和用户需求挑战，提出了对话状态跟踪、策略学习和响应生成三个任务，由南京大学和阿里巴巴于2022年发布。