# WebCPM: Interactive Web Search for Chinese Long-form Question Answering

**Authors:** Yujia Qin, Zihan Cai, Dian Jin, Lan Yan, Shihao Liang, Kunlun Zhu, Yankai Lin, Xu Han, Ning Ding, Huadong Wang, Ruobing Xie, Fanchao Qi, Zhiyuan Liu, Maosong Sun, Jie Zhou  
**arXiv:** [2305.06849](https://arxiv.org/abs/2305.06849)  
**Date:** 2023-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.06849v2/

## Key points

- Introduction of WebCPM, the first Chinese Long-form Question Answering (LFQA) dataset
- LFQA aims to answer complex, open-ended questions with detailed responses
- LFQA involves two procedures: information retrieval and information synthesis
- WebCPM stands out because it uses interactive web search for information retrieval
- Development of a web search interface called WebGPT for annotators to search for relevant information and answer questions
- Dataset includes 5,500 question-answer pairs, 14,315 supporting facts, and 121,330 web search actions
- Pre-trained language models are fine-tuned to generate answers based on collected facts and imitate human behaviors in web search
- LFQA pipeline using fine-tuned models generates answers comparable to human-written ones in 32.5% and 47.5% of cases on their dataset and DuReader respectively
- WebCPM is a valuable resource for Chinese LFQA research by incorporating interactive web search and demonstrating the effectiveness of fine-tuned language models

## Summary

The paper introduces WebCPM, the first Chinese Long-form Question Answering (LFQA) dataset. LFQA aims to answer complex, open-ended questions with detailed responses. The process of LFQA involves two procedures: information retrieval and information synthesis. Information retrieval searches for relevant supporting facts, while information synthesis integrates these facts into a coherent answer. WebCPM stands out because its information retrieval is based on interactive web search, engaging with a search engine in real time. The authors develop a web search interface called WebGPT and recruit annotators to use this interface to search for relevant information and answer questions. The annotators' web search behaviors are recorded during the process. The dataset collected includes 5,500 high-quality question-answer pairs, along with 14,315 supporting facts and 121,330 web search actions. To generate answers based on the collected facts and imitate human behaviors for web search, pre-trained language models are fine-tuned. The authors build an LFQA pipeline using these fine-tuned models. The pipeline generates answers that are no worse than human-written ones in 32.5% and 47.5% of the cases on their dataset and DuReader respectively. Overall, this paper presents WebCPM as a valuable resource for Chinese LFQA research by providing a dataset that incorporates interactive web search for information retrieval and demonstrates the effectiveness of fine-tuned language models in generating accurate answers.