Dynamic Knowledge Routing Network For Target-Guided Open-Domain Conversation

This repository contains the source code for the following paper:

Dynamic Knowledge Routing Network For Target-Guided Open-Domain Conversation
Jinghui Qin, Zheng Ye, Jianheng Tang, Xiaodan Liang; AAAI 2020

Model Overview

Prerequisites

Create virtural environment (recommended):

conda create -n TGODC python=3.6
source activate TGODC

Install the required packages:

pip install -r requirements.txt

Install Texar locally:

cd texar-0.2.1
pip install .

Note: make sure that your environment has installed cuda 10.0.

Usage

Data Preparation

We conduct our experiments on the Target-Guided PersonaChat Dataset(TGPC) proposed by (Tang et al. 2019), and our proposed Chinese Weibo Conversation Dataset(CWC).
Download the processed data(TGPC, CWC) and unzip them into the root directory of this code repository.

For the data preprocessing details, you could see the code inside the preprocess and preprocess_weibo directories.

Turn-level Supervised Learning

In this project, we propose DKRN agent with more accurate next-topic prediction, compared with 5 different types of baseline agents (kernel/neural/matrix/retrieval/retrieval_stgy). You can modify the configration of each agent in the config directory for the TGPC dataset and config_weibo directory for the CWC dataset.

To train our DKRN dialogue agent, you need to first train the keyword prediction module, and then train the response retrieval module:

python train.py --mode train_kw --agent neural_dkr
python train.py --mode train --agent neural_dkr
python train.py --mode test --agent neural_dkr

To train the baseline agents, change the --agent argument with a specific agent name (kernel/neural/matrix/).

Specially, for the retrieval and retrieval_stgy agents, you only need to train the response retrieval module:

python train.py --mode train --agent retrieval
python train.py --mode test --agent retrieval

Note: the retrieval agent and the retrieval_stgy agent share the same retrieval module. You only need to train one of them.

In addition, we provide a checkpoint of our DKRN agent trained on TGPC dataset. If you wanna use it, download the DKRN checkpoint and unzip it into the save/ directory(if save directory doesn't exist in the root directory, just create a directory named save).

Target-guided Conversation

After turn-level training, you can start target-guided conversation with our DKRN agent.

python target_chat.py --agent neural_dkr --times 3

You can also watch the simulation of the target-guided conversation between the retrieval agent pretending the user and our DKRN agent. The success rate and average turns would be shown in the end.

python target_simulation.py --agent neural_dkr --times 500 --dataset TGPC --print_details=False --use_fixed_start_corpus=True

Note:

Baseline agents (kernel/neural/matrix/retrieval/retrieval_stgy) are also supported by changing the --agent argument.
Set --times argument to indicate the conversation times or the simulation times.