Mitigating Negative Style Transfer in Hybrid Dialogue System

Authors: Shimin Li, Qinyuan Cheng, Linyang Li, Xipeng Qiu

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We performed extensive experiments on three dialogue datasets, including a hybrid dialogue dataset and two task-oriented dialogue datasets. The experimental results demonstrate that our method can mitigate the negative style transfer issue and achieves state-of-the-art performance on multiple dialogue datasets.
Researcher Affiliation Academia Shimin Li1, Qinyuan Cheng1, Linyang Li1, Xipeng Qiu1,2 *, 1 School of Computer Science, Fudan University 2 Shanghai Key Laboratory of Intelligent Information Processing, Fudan University {smli20, linyangli19, xpqiu}@fudan.edu.cn, chengqy21@m.fudan.edu.cn
Pseudocode No No clearly labeled pseudocode or algorithm blocks are present in the paper.
Open Source Code Yes Code: https://github.com/whatissimondoing/Hi S-Dialog
Open Datasets Yes Fused Chat (Young et al. 2021) This dataset expands or rewrites each conversation based on the task-oriented task. Multi WOZ (Budzianowski et al. 2018; Eric et al. 2020) This dataset is one of the most prevalent datasets in task-oriented dialogue systems, collected via Wizard-of-Oz, and contains a total of 8438/1000/1000 multiturn dialogues.
Dataset Splits Yes Multi WOZ (Budzianowski et al. 2018; Eric et al. 2020) This dataset is one of the most prevalent datasets in task-oriented dialogue systems, collected via Wizard-of-Oz, and contains a total of 8438/1000/1000 multiturn dialogues.
Hardware Specification Yes All experiments were performed on a Ge Force RTX 3090 GPU (24G)
Software Dependencies No Hi S-Dialog and other baselines based on pre-trained models are implemented with Hugging Face s Transformers. No specific version numbers are provided for this or any other software.
Experiment Setup Yes We employ Adam W as the optimizer and configure the warmup rate to 0.1. For Fused Chat, the learning rate is 6e-4 and the batch size is set to 12 for 12 epochs. For Multi WOZ, we set the learning rate 5e-4, epoch 10, and batch size 12.