WDMN

This is the original implementation for wide-and-deep matching network (WDMN). [Response Ranking with Multi-Types of Deep Interactive Representations in Retrieval-based Dialogues.]

Requirements

Option 1: using container image

tensorflow/tensorflow:1.15.5-gpu-py3

Option 2: building Python environment on your own

• python==3.7.6
• numpy==1.18.1
• tensorflow-gpu==1.15.0

Usage

1. Build directory structure

   ...$ mkdir WDMN && cd WDMN
   .../WDMN$ mkdir code data

2. Download data and unzip it

   .../WDMN$ unrar x dataset.rar data

   Thank chunyuanY for providing the data.

   • In case you cannot download that data, use this link instead (need not to unrar it)
3. Set up the environment (see Requiments above)
4. Clone this reposity and train WDMN

   .../WDMN$ cd code
   .../WDMN/code$ git clone https://github.com/RayXu14/WDMN.git
   .../WDMN/code$ cd WDMN
   # (optional) modify run.sh to change the configuration 
   .../WDMN/code/WDMN$ bash run.sh # please make sure the environment is set up properly

Results

Dataset	R_2@1	R_10@1	R_10@2	R_10@5	MAP	MRR	P@1
Ubuntu (Lowe et al., 2015)	0.957	0.821	0.911	0.981	-	-	-
Douban (Wu et al., 2017)	-	0.301	0.460	0.799	0.594	0.644	0.490
E-commerce (Zhang et al., 2018)	-	0.669	0.831	0.956	-	-	-

Citation

The paper is published at TOIS, 2021 and will be presented at SIGIR, 2022.