/CEM

EMNLP'22, CEM improves MHCH performance by correcting prediction bias and training an auxiliary cost simulator based on user state and labor cost causal graph, without requiring complex model crafting.

Primary LanguagePythonMIT LicenseMIT

CEM: Machine-Human Chatting Handoff via Causal-Enhance Module

GitHub GitHub

This repository is the implementation of "CEM: Machine-Human Chatting Handoff via Causal-Enhance Module" [paper] on Clothes and Makeup2 datasets. Our paper has been accepted for presentation at EMNLP 2022.

Introduction

CEM solves the problem of Machine-Human Chatting Handoff, establishing the causal graph of MHCH, which is a simple-yet-effective module and can be easy to plug into the existing MHCH methods.

Requirement

Activate an enviroment of Python 3.7, then sh env.sh.

Data Format

Our experiments are conducted based on two publicly available Chinese customer service dialogue datasets, namely Clothes and Makeup2, collected by Song et al. (2019) from Taobao.

  • Each pkl file is a data list that includes dialogue samples. The content lists of the dataset can be seen in data_loader.py.

  • The vocab.pkl contains a vocabulary class which contains the pre-trained glove word embeddings of token ids.

Usage

  • Train the model (including training, validation, and testing)
python -u -W ignore main.py --task train --model cem --data makeup2
  • Test the model
python -u -W ignore main.py --task test --model cem --data makeup2 --model_path pretrained_model_dir

Acknowledgments

Many thanks to LauJames for his Tensorflow framework for MHCH task.