Learning to Customize Model Structures for Few-shot Dialogue Generation Tasks

Abstract

Training the generative models with minimal corpus is one of the critical challenges for building open-domain dialogue systems. Existing methods tend to use the meta-learning framework which pre-trains the parameters on all non-target tasks then fine-tunes on the target task. However, fine-tuning distinguishes tasks from the parameter perspective but ignores the model-structure perspective, resulting in similar dialogue models for different tasks. In this paper, we propose an algorithm that can customize a unique dialogue model for each task in the few-shot setting. In our approach, each dialogue model consists of a shared module, a gating module, and a private module. The first two modules are shared among all the tasks, while the third one will differentiate into different network structures to better capture the characteristics of the corresponding task. The extensive experiments on two datasets show that our method outperforms all the baselines in terms of task consistency, response quality, and diversity.

Coffee With a Hint of Data: Towards Using Data-Driven Approaches in Personalised Long-Term Interactions

“…Data-driven models can also be combined with graphical models ( Zhou et al, 2020 ; Song et al, 2019 ; Moon et al, 2019 ; Shi et al, 2020 ; Wu B. et al, 2020 ; Xu et al, 2020 ), rule-based or slot-filling systems ( Tammewar et al, 2018 ; Zhang Z. et al, 2019 ), a knowledge-base ( Ganhotra and Polymenakos, 2018 ; Ghazvininejad et al, 2018 ; Luo et al, 2019 ; Yavuz et al, 2019 ; Moon et al, 2019 ; Wu et al, 2019 ; Lian et al, 2019 ; Zhang B. et al, 2020 ; Majumder et al, 2020 ; Tuan et al, 2021 ) or with automatic extraction of attributes from dialogue ( Tigunova et al, 2019 , 2020 ; Wu C.-S. et al, 2020 , 2021 ; Ma et al, 2021 ) to improve the personalised entity selection in responses. Methods that adopt transfer learning ( Genevay and Laroche, 2016 ; Lopez-Paz and Ranzato, 2017 ; Mo et al, 2017 , 2018 ; Yang et al, 2017 , 2018 ; Wolf et al, 2019 ; Golovanov et al, 2020 ), meta-learning ( Finn et al, 2017 ; Santoro et al, 2016 ; Vinyals et al, 2016 ; Munkhdalai and Yu, 2017 ; Madotto et al, 2019 ; Zhang W.-N. et al, 2019 ; Song et al, 2020 ; Tian et al, 2021 ) and key-value memory structures ( Xu et al, 2017 ; Kaiser et al, 2017 ; Zhu and Yang, 2018 , 2020 ; de Masson d’Autume et al, 2019 ) could provide effective insights to alleviate data scarcity and enable quick adaption to various users through improving few-shot and lifelong learning capabilities of the dialogue models ( Wang et al, 2020b ).…”

“…Another solution could be to combine a data-driven model with another approach to compensate for the deficiencies in the models, such as combining a generative model (e.g., Sequence-to-Sequence) with a Memory Network (Madotto et al, 2018;Zhang B. et al, 2020) or with transformers (Vaswani et al, 2017), such as in the work of Roller et al (2020), Generative Pre-trained Transformer (GPT) (Radford et al, 2018(Radford et al, , 2019Brown et al, 2020;Zhang Y. et al, 2020), Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al, 2019;Song et al, 2021), and Poly-encoders (Humeau et al, 2020;Li et al, 2020). Data-driven models can also be combined with graphical models (Zhou et al, 2020;Song et al, 2019;Moon et al, 2019;Shi et al, 2020;Wu B. et al, 2020;Xu et al, 2020), rule-based or slot-filling systems (Tammewar et al, 2018;Zhang Z. et al, 2019), a knowledge-base (Ganhotra and Polymenakos, 2018;Ghazvininejad et al, 2018;Luo et al, 2019;Yavuz et al, 2019;Moon et al, 2019;Wu et al, 2019;Lian et al, 2019;Zhang B. et al, 2020;Majumder et al, 2020;Tuan et al, 2021) or with automatic extraction of attributes from dialogue (Tigunova et al, 2019(Tigunova et al, , 2020Wu C.-S. et al, 2020Wu C.-S. et al, , 2021Ma et al, 2021) to improve the personalised entity selection in responses.…”

“…The results show that on average, Split Memory is the best performing model, however, End-to-End Memory Network (MemN2N) is the best model for task 8 (containing all tasks). Santoro et al, 2016;Vinyals et al, 2016;Munkhdalai and Yu, 2017;Madotto et al, 2019;Zhang W.-N. et al, 2019;Song et al, 2020;Tian et al, 2021) and key-value memory structures (Xu et al, 2017;Kaiser et al, 2017;Yang, 2018, 2020;de Masson d'Autume et al, 2019) could provide effective insights to alleviate data scarcity and enable quick adaption to various users through improving few-shot and lifelong learning capabilities of the dialogue models (Wang et al, 2020b). Another line of attack could be to learn from users during deployment (similar to Zhao and Eskenazi, 2016;Hancock et al, 2019;Sreedhar et al, 2020;Liu, 2020;Irfan et al, 2020b), where the feedback could be the user response or emotions of the user to evaluate the user satisfaction with the agent's responses.…”

BoB: BERT Over BERT for Training Persona-based Dialogue Models from Limited Personalized Data

“…Meta-learning has recently been explored in addressing the limited personalized data issue. CMAML (Song et al, 2020c) is a meta-learning based method that learns from few shot personas by customizing the model structures. Besides the meta-learning methods, GDR (Song et al, 2020a) introduces inference ability on the PersonaChat with a generate-refine framework.…”

Graphine: A Dataset for Graph-aware Terminology Definition Generation

“…This problem becomes even more severe in emerging research topics (Baig, 2020;Baines et al, 2020), such as COVID-19, where curated definitions could be imprecise and do not scale to rapidly proposed terminologies. Neural text generation (Bowman et al, 2016;Vaswani et al, 2017;Sutskever et al, 2014;Song et al, 2020b) could be a plausible solution to this problem by generating definition text based on the terminology text. Encouraging results by neural text generation have been observed on related tasks, such as paraphrase generation (Li et al, 2020), description generation (Cheng et al, 2020), synonym generation (Gupta et al, 2015) and data augmentation (Malandrakis et al, 2019).…”