On the Compositional Generalization in Versatile Open-domain Dialogue

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13585–13605 July 9-14, 2023 ©2023 Association for Computational Linguistics On the Compositional Generalization in Versatile Open-domain Dialogue Tingchen Fu1†, Xueliang Zhao2†, Lemao Liu3, Rui Yan1,4∗ 1Gaoling School of Artificial Intelligence, Renmin University of China 2The University of Hong Kong 3Tencent AI Lab 4Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education lucas.futingchen@gmail.com redmondliu@tencent.comxlzhao22@connect.hku.hk ruiyan@ruc.edu.cn Abstract Previous research has demonstrated the po- tential of multi-task learning to foster a con- versational agent’s ability to acquire a vari- ety of skills. However, these approaches ei- ther suffer from interference among different datasets (also known as negative transfer), or fail to effectively reuse knowledge and skills learned from other datasets. In contrast to previous works, we develop a sparsely acti- vated modular network: (1) We propose a well- rounded set of operators and instantiate each operator with an independent module; (2) We formulate dialogue generation as the execution of a generated programme which recursively composes and assembles modules. Extensive experiments on 9 datasets verify the efficacy of our methods through automatic evaluation and human evaluation. Notably, our model outper- forms state-of-the-art supervised approaches on 4 datasets with only 10% training data thanks to the modular architecture and multi-task learn- ing. 1 1 Introduction Building an open-domain dialogue system is an in- triguing and challenging task. A good open-domain chatbot should be equipped with a well-rounded set of skills (Roller et al., 2021) including but not limited to providing an informative response, show- ing different emotions, keeping a consistent per- sona and conducting commonsense inference. Up to now, with more and more datasets proposed to train multiple conversation skills (e.g., Wizard of Wikipedia (Dinan et al., 2019), Personachat (Zhang et al., 2018)), multi-task learning is an efficient way to grasp all the versatile skills and quickly transfer to newly emerging datasets (Roller et al., 2021). †Tingchen Fu and Xueliang Zhao contribute equally to this work. This work was done during their internship at Tencent AI Lab. *Corresponding author: Rui Yan(ruiyan@ruc.edu.cn) 1The code is available at https://github.com/ TingchenFu/ACL23-ModularDialogue However, as a core problem in multi-task learn- ing, it is not easy to strike a balance between trans- fer and interference (negative transfer) among mul- tiple datasets (Rosenbaum et al., 2019). For this, recent researchers mainly follow two lines of re- search. On one line, Roller et al. (2021) and Shuster et al. (2022a) simply mix all the datasets together to embody the blended skill required in dialogue. They update all the model parameters to minimize the loss of all of the data, which is also dubbed dense training (Gururangan et al., 2022). In spite of its simplicity, it easily incurs interference among different datasets (Aribandi et al., 2022). On the other line, Li and Liang (2021) learn multiple skills and store the knowledge from different datasets with different sets of parameter-efficient architec- tures. This approach eliminates underlying nega- tive transfer among different corpora, but hinders …

Content truncated to 500 words