hejunqing / multiturnresponseselection

This repo contains our ACL paper data and source code

Douban Conversation Corpus

Data set

We release Douban Conversation Corpus, comprising a training data set, a development set and a test set for retrieval based chatbot. The statistics of Douban Conversation Corpus are shown in the following table.

	Train	Val	Test
session-response pairs	1m	50k	10k
Avg. positive response per session	1	1	1.18
Fless Kappa	N\A	N\A	0.41
Min turn per session	3	3	3
Max ture per session	98	91	45
Average turn per session	6.69	6.75	5.95
Average Word per utterance	18.56	18.50	20.74

The test data contains 1000 dialogue context, and for each context we create 10 responses as candidates. We recruited three labelers to judge if a candidate is a proper response to the session. A proper response means the response can naturally reply to the message given the context. Each pair received three labels and the majority of the labels was taken as the final decision.

As far as we known, this is the first human-labeled test set for retrieval-based chatbots.

Data template

label \t conversation utterances (splited by \t) \t response

Source Code

We also release our source code to help others reproduce our result. The code has been tested under Ubuntu 14.04 with python 2.7.

Reference

Please cite our paper if you use the data or code in this repos.

Wu, Yu, et al. "Sequential Matching Network: A New Archtechture for Multi-turn Response Selection in Retrieval-based Chatbots." ACL. 2017.