Iterated Attentive Convolution Matching Network (IACMN)

This is an implementation of our CIKM 2019 paper: [Multi-Turn Response Selection in Retrieval-Based Chatbots with Iterated Attentive Convolution Matching Network].

Network

IACMN is a neural deep matching network proposed for multi-turn response selection in the retrieval-based chatbot.

IACMN iteratively constructs multi-grained representations of the response candidate and its multi-turn history context entirely based on hierarchical stacking of the proposed AGDR block, which is a refined combination of gated dilated-convolution and self-attention.

IACMN calculates and integrates the interactive matrices between each utterance-response pair from different views, then accumulates the sequencial matching vectors into a fused vector to obtain the final score.

• Model Overview

• AGDR Block

Results

We test IACMN on two large-scale multi-turn response selection tasks, i.e., the Ubuntu Corpus v1 and Douban Conversation Corpus, experimental results are bellow:

Usage

First, please download data according to data/ReadMe.txt and unzip it:

cd data
unzip data.zip

Train and test the model by:

python main.py

Dependencies

• Python >= 3.5
• Tensorflow >= 1.4

Citation

If you use this code, please cite the following paper:

@inproceedings{wang2019multi,
  title={Multi-Turn Response Selection in Retrieval-Based Chatbots with Iterated Attentive Convolution Matching Network},
  author={Wang, Heyuan and Wu, Ziyi and Chen, Junyu},
  booktitle={Proceedings of the 28th ACM International Conference on Information and Knowledge Management},
  pages={1081--1090},
  year={2019},
  organization={ACM}
}