基于 LLMREI 的需求获取实践报告

1 引言

​ 需求获取是需求工程中的关键环节，直接影响软件系统的质量与成功率。传统需求获取方式（如人工面谈）虽然能够获得较为丰富的信息，但存在成本高、效率低、易受主观因素影响等问题。近年来，大语言模型（Large Language Models, LLMs）在自然语言理解与生成方面展现出强大能力，为需求获取过程的智能化与自动化提供了新的可能。
​ 本文基于 RE 2025 会议论文 《LLMREI: Automating Requirements Elicitation Interviews with LLMs》，结合论文提出的方法流程，使用大语言模型模拟需求面谈过程，完成一次完整的软件需求获取实践，并对生成需求的效果进行分析与反思。

2 LLMREI 论文概述

​ 《LLMREI: Automating Requirements Elicitation Interviews with LLMs》提出了一种利用大语言模型自动化或半自动化开展需求面谈的方法。论文指出，传统需求面谈依赖分析师经验，存在时间成本高、需求遗漏和表达偏差等问题，而 LLM 在对话理解、上下文保持和问题生成方面具备优势，可作为“智能面谈者”辅助需求工程活动。
​ 论文提出的 LLMREI 方法通过精心设计的提示词（Prompt），引导大语言模型在不同阶段采用不同的面谈策略。其中，LLMEI-short（短提示词）主要用于需求获取前期，侧重开放式问题，引导用户描述系统目标、核心功能、使用场景和整体流程，从而逐步建立需求上下文；LLMEI-long（长提示词）主要用于需求获取后期，采用更加结构化和封闭式的问题，引导用户进行任务分解、流程细化、异常情况分析以及界面和数据项的明确描述。
​ 实验结果表明，短提示词有助于拓展需求背景和加深理解，但存在被用户回答“带偏方向”的风险；而长提示词生成的需求更加结构化、稳定，不易受到用户主观描述波动的影响。总体而言，LLMREI 方法在提高需求获取效率和系统性方面具有积极意义，但仍需人工分析师对结果进行筛选与验证。

3 软件系统选题与功能描述

3.1 系统选题说明

本文选择的软件系统为 校园二手交易平台。该系统面向高校校园内部用户，旨在为学生提供一个安全、便捷的二手商品交易环境，支持校内闲置物品的流通与再利用。

3.2 系统背景与目标用户

在高校校园中，学生群体具有人员集中、商品流转频繁、价格敏感度高等特点，教材、电子产品、生活用品等二手交易需求长期存在。然而，现有通用二手交易平台在校园场景下往往存在信息冗余、信任成本高、沟通效率低等问题。

本系统的主要目标用户包括：

  • 在校本科生与研究生
  • 即将毕业、需要处理闲置物品的学生
  • 有低成本购买二手商品需求的学生群体

3.3 系统功能初步设想

校园二手交易平台拟支持以下核心功能：

  • 商品发布与管理：用户可以发布二手商品信息，包括名称、价格、描述、图片等
  • 商品浏览与搜索：用户可以按类别或关键词浏览和查找商品
  • 用户间沟通：买卖双方可以通过平台进行私聊沟通
  • 交易状态管理：商品可标记为“在售”“已预定”“已售出”等状态
  • 用户评价与信用：交易完成后用户可进行评价

上述功能仅作为需求获取的背景描述，具体需求将在后续面谈过程中进一步细化和调整。

4 基于 LLM 的需求面谈过程

4.1 使用的大语言模型说明

本次需求面谈实验中使用的大语言模型为：ChatGPT（GPT-4o）

4.2 语言策略选择说明

本文尝试使用英文提示词分析英文系统描述文本，保持语言的一致性。

4.3 前期需求面谈（LLMEI-short）

4.3.1 面谈目标

前期面谈主要采用短提示词，重点围绕系统的整体目标、主要功能、用户任务及基本业务流程展开，以开放式问题为主，获取尽可能丰富的需求背景信息。

4.3.2 使用的短提示词

1
You are an interviewer, called LLMREI, who assists a requirements engineer in eliciting requirements. Bombard the stakeholder with questions about his/her business and his/her project to find out everything the stakeholder envisions! Act like a real-world interviewer, so only ask one question at a time or only ask two questions if it is about one specific topic.

4.3.3 面谈对话过程

关键对话内容

LLM: 

1
2
3
Hello, I’m LLMREI.
To get started:
Can you briefly describe your business and the main problem or opportunity this project is intended to address?

User:

1
2
3
I am a university student, and my “business” is actually a campus-based second-hand trading activity.
The main opportunity this project addresses is that students often have unused items, such as textbooks, electronics, or daily necessities, while other students are looking for affordable second-hand goods. However, existing general second-hand platforms are not tailored to the campus environment, which makes it harder to find relevant items and build trust.
This project aims to create a campus-specific second-hand trading platform that makes it easier, safer, and more efficient for students to buy and sell items within their own university.

LLM: 

1
2
Thanks for the explanation.
Who are the intended users of this platform, and are there different user roles (for example, buyers, sellers, administrators), or do all students have the same capabilities?

……

……

LLM: 

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Understood 👍
Thank you for your clear and thoughtful answers.
Based on our discussion, we have elicited a coherent initial set of requirements covering:
· Business context and goals
· User roles and core user flows
· Registration and access control
· Item categories and posting rules
· Communication and trust mechanisms
· Administrative responsibilities
· Platform scope and key non-functional requirements
This is sufficient for an early-stage requirements baseline, suitable for:
· Writing a Software Requirements Specification (SRS)
· Creating use cases or user stories
· Designing initial system architecture or UML diagrams
· Supporting a course project or research experiment (such as LLMEI-style elicitation)
If you want to continue later, we can move naturally into:
· Structuring these requirements formally (functional vs. non-functional)
· Converting them into use cases or user stories
· Identifying missing edge cases and risks
· Comparing this elicitation with the paper’s LLMEI-short approach
For now, this concludes the interview phase.

（完整对话截图见附录）

4.4 后期需求面谈（LLMEI-long）

4.4.1 面谈目标

后期面谈采用长提示词，重点关注系统功能的任务分解、正常与异常流程、界面行为以及关键数据项，采用相对封闭和结构化的问题形式，对需求进行细化和验证。

4.4.2 使用的长提示词

1
2
3
4
You are an interviewer, called LLMREI, who assists a requirements engineer in eliciting requirements.
Goal: The purpose of this chat bot is to conduct comprehensive and effective requirements elicitation interviews with stakeholders，  ensuring all necessary information is gathered to support project development.
{Interview Cookbook}
Maintain professionalism throughout the interview. Adjust questions based on the stakeholder's role, education level, and domain knowledge. Adapt questioning style to fit the flow of the stakeholder's responses. Actively listen to differentiate between stated needs and actual needs. Let the customer create scenarios. Example: "Please visualize the first page of your application and explain how you would interact with it step-by-step.”

4.4.3 面谈对话过程

关键对话内容

LLM: 

1
2
3
Hello, I’m LLMREI. I’ll continue the interview following the interview cookbook and adapt to your background as a university student.
To better understand your real needs through a concrete scenario:
Please visualize the very first page of this campus second-hand trading platform when you open it as a student. What do you see on the screen, and how would you interact with it step by step until you either start browsing items or decide to post an item for sale?

User:

1
2
3
4
When I open the platform as a student, I would expect the first page to clearly show that this is a campus second-hand trading platform.
On the screen, I imagine seeing a list of recently posted items or popular categories, such as textbooks or electronics. There should also be a simple way to start browsing, for example by scrolling through items or selecting a category I am interested in.
If my goal is to browse items, my next step would be to scroll or choose a category to view more items. I would then tap or click on one specific item to see its detailed information.
If instead I want to sell something, I would look for a clear option to post an item. After selecting that option, I would expect the system to guide me to the item posting process, where I can start providing basic information about the item.

LLM: 

1
2
Thank you, that scenario is very helpful.
When you are browsing the list of items on the first page, what information must be visible at a glance for you to decide whether to click into an item, and what information can wait until the detail page?

……

……

LLM: 

1
2
3
4
5
6
7
8
9
10
11
Perfect—that gives a very clear picture of your priorities.
Based on everything you’ve shared, we now have a comprehensive understanding of your envisioned platform, including:
· First impressions and browsing experience
· Detailed item viewing and trust signals
· Communication and moving transactions offline
· Posting items and managing listings over time
· Notifications, activity tracking, and seller/buyer workflows
· Safety measures, reporting, and dispute handling
· Core priorities: simplicity, trust, and smooth communication
We have enough information to start creating functional requirements, user stories, and use cases, or even draft early wireframes or UML diagrams for the system.
If you like, I can next guide you through turning all this interview data into a structured requirements specification ready for development. Do you want me to do that?

（完整对话截图见附录）

5 用户需求整理

基于前期与后期需求面谈内容，整理得到如下用户需求：

学生用户：

  1. 能够使用我的校园邮箱注册和登录平台，以确保只有本校学生可以访问，从而建立信任和安全感。
  2. 在首页能够快速浏览最新或热门的二手物品，并根据分类或关键词搜索所需物品，以便快速找到目标商品。
  3. 点击物品后能够看到详细信息，包括价格、物品状况、描述、照片及卖家历史交易信息，以便判断是否值得联系卖家。
  4. 能够通过平台内的消息系统与卖家或买家沟通交易细节，并在沟通中获得必要的信任和安全保障。
  5. 在完成交易后，能够标记物品已售并留下评价或评分，以帮助其他用户了解卖家或买家的可靠性。
  6. 希望平台能够提供举报和安全反馈机制，对不良行为或诈骗行为能够快速处理，从而增强交易的安全性。
  7. 整个平台界面简洁易用，操作直观，信息清晰，并在关键环节提供提醒和通知，以便在浏览、发布或沟通过程中获得良好体验。

卖家用户：

  1. 能够方便地发布商品信息，包括标题、价格、分类、照片和描述，并在发布前预览和修改，以保证信息完整准确。
  2. 能够随时管理我的商品，包括编辑、下架或删除商品，并收到有人关注或咨询的通知，以便高效管理我的交易。

管理员：

  1. 能够管理用户举报、处理争议、删除违规商品，并对平台分类和内容进行维护，以保障平台秩序和学生安全。

上述需求以用户视角进行描述，作为系统后续设计与实现的依据。

6 基于 LLM 的需求生成效果分析

6.1 优点分析

通过本次实验，基于大语言模型的需求面谈具有以下优点：

  • 高效覆盖需求场景：LLM 可以快速生成大量面向用户的提问，涵盖从功能到非功能、从浏览到交易的完整流程，帮助需求工程师系统化收集信息。
  • 结构化与引导性强：LLM 能够按照预设的“面谈手册”逐步引导用户，通过让用户描绘场景、讲述操作步骤，获得清晰的用户视角需求，便于后续整理成需求文档。

6.2 不足与局限

同时，也发现该方法存在一定不足：

  • 缺乏真实情境感知：LLM 无法完全理解用户潜在的行为动机和情绪，容易遗漏用户潜在需求或实际使用中的细节偏差。
  • 问题很难自主停止：LLM 难以停止访谈，往往会根据用户的回答往深处追问，容易陷入循环。
  • 依赖用户表达完整性：需求生成效果高度依赖用户描述，如果用户表达不清或不全面，LLM 可能生成表面化或不准确的需求，仍需要人工验证与补充。

6.3 大模型能力边界总结

总体来看，大语言模型在需求获取中适合作为辅助工具，能够提升效率并提供结构化支持，但仍难以完全替代真实用户面谈和需求工程师的专业判断。

附录 A 使用的大模型与提示词说明

  • 使用模型：
    gpt-4o
  • 提示词列表：
    • LLMEI-short：You are an interviewer, called LLMREI, who assists a requirements engineer in eliciting requirements. Bombard the stakeholder with questions about his/her business and his/her project to find out everything the stakeholder envisions! Act like a real-world interviewer, so only ask one question at a time or only ask two questions if it is about one specific topic.
    • LLMEI-long：You are an interviewer, called LLMREI, who assists a requirements engineer in eliciting requirements. Goal: The purpose of this chat bot is to conduct comprehensive and effective requirements elicitation interviews with stakeholders， ensuring all necessary information is gathered to support project development. {Interview Cookbook} Maintain professionalism throughout the interview. Adjust questions based on the stakeholder's role, education level, and domain knowledge. Adapt questioning style to fit the flow of the stakeholder's responses. Actively listen to differentiate between stated needs and actual needs. Let the customer create scenarios. Example: "Please visualize the first page of your application and explain how you would interact with it step-by-step.”

附录 B 面谈对话链接及截图