Douban Conversation Corpus

我们发布豆瓣会话语料库，包括一个训练数据集、一个开发集和一个基于检索的聊天机器人的测试集。豆瓣会话语料统计如下表所示。
火车
瓦尔
测试
会话响应对
1m
50k
10k
平均每个会话的积极响应
1
1
1.18
无情河童
不适用
不适用
0.41
每次会话的最小转数
3
3
3
每个会话的最大值
98
91
45
每次会话的平均转数
6.69
6.75
5.95
每个话语的平均字数
18.56
18.50
20.74
测试数据包含 1000 个对话上下文，对于每个上下文，我们创建 10 个响应作为候选。我们招募了三名标注员来判断候选人是否对会议做出了适当的回应。正确的响应意味着响应可以自然地回复给定上下文的消息。每对收到三个标签，大部分标签被视为最终决定。
据我们所知，这是第一个用于基于检索的聊天机器人的人工标记测试集。整个语料链接 https://www.dropbox.com/s/90t0qtji9ow20ca/DoubanConversaionCorpus.zip?dl=0
数据模板
标签 \t 对话话语（由 \t 分割）\t 响应

We hereby release the Douban Conversation Corpus, which includes a training dataset, a development set, and a retrieval-based chatbot test set. The statistics of the Douban Conversation Corpus are shown in the following table.

| Statistic | Training Set | Development Set | Test Set |
|---|---|---|---|
| Number of conversation-response pairs | 1M | 50K | 10K |
| Average positive responses per session | 1 | 1 | 1.18 |
| Average negative responses per session | N/A | N/A | 0.41 |
| Minimum number of turns per session | 3 | 3 | 3 |
| Maximum number of turns per session | 98 | 91 | 45 |
| Average number of turns per session | 6.69 | 6.75 | 5.95 |
| Average number of words per utterance | 18.56 | 18.50 | 20.74 |

The test set contains 1000 dialogue contexts. For each context, we created 10 candidate responses. We recruited three annotators to judge whether the candidates made appropriate responses to the conversation. An appropriate response means that the response can naturally reply to the message of the given context. Each pair receives three labels, and the majority label is taken as the final decision.

To the best of our knowledge, this is the first manually annotated test set for retrieval-based chatbots. The entire corpus is available at: https://www.dropbox.com/s/90t0qtji9ow20ca/DoubanConversaionCorpus.zip?dl=0

Data Template: Label Dialogue Utterances (split by ) Response

搜集汇总

背景与挑战

背景概述

Douban Conversation Corpus是一个用于检索式聊天机器人研究的会话语料库，包含训练集、开发集和人工标记的测试集，提供了丰富的会话统计指标。该数据集由北京航空航天大学、南开大学和Microsoft Research联合发布，特别适合多轮对话响应选择任务的研究。