[Paper Review] Generating Personalized Dialogue via Multi-Task Meta-Learning

Jing Yang Lee, Kong Aik Lee+1 more|DR-NTU (Nanyang Technological University)|Aug 7, 2021

Topic Modeling34 references4 citations

TL;DR

This paper proposes Multi-Task Meta-Learning (MTML) and Alternating Multi-Task Meta-Learning (AMTML), frameworks that enhance personalized dialogue generation by training models to infer persona traits from dialogue context alone, using a persona reconstruction task during meta-learning. The key contribution is improved persona consistency in generated responses without requiring predefined persona statements at inference time, outperforming prior methods on the PersonaChat benchmark.

ABSTRACT

Conventional approaches to personalized dialogue generation typically require a large corpus, as well as predefined persona information. However, in a real-world setting, neither a large corpus of training data nor persona information are readily available. To address these practical limitations, we propose a novel multi-task meta-learning approach which involves training a model to adapt to new personas without relying on a large corpus, or on any predefined persona information. Instead, the model is tasked with generating personalized responses based on only the dialogue context. Unlike prior work, our approach leverages on the provided persona information only during training via the introduction of an auxiliary persona reconstruction task. In this paper, we introduce 2 frameworks that adopt the proposed multi-task meta-learning approach: the Multi-Task Meta-Learning (MTML) framework, and the Alternating Multi-Task Meta-Learning (AMTML) framework. Experimental results show that utilizing MTML and AMTML results in dialogue responses with greater persona consistency.

Motivation & Objective

• To address the practical limitation of requiring large persona-specific datasets and predefined persona statements in real-world personalized dialogue systems.
• To enable rapid adaptation to unseen personas without relying on explicit persona annotations during inference.
• To improve persona consistency in generated responses by leveraging a persona reconstruction task during meta-learning.
• To explore whether multi-task meta-learning with persona reconstruction enhances the model's ability to infer and reflect personality from dialogue context.

Proposed method

• The MTML framework combines response generation loss and persona reconstruction loss into a single weighted objective during meta-learning.
• The AMTML framework alternates between optimizing response generation and persona reconstruction losses across training iterations.
• Persona reconstruction is performed as an auxiliary task during meta-learning, where the model generates full persona statements from dialogue context.
• The model is trained using a meta-learning paradigm inspired by MAML, enabling fast adaptation to new personas with few examples.
• Only the dialogue context is used during inference; persona statements are not provided or used at test time.
• The approach leverages pretrained language models fine-tuned via meta-learning to improve generalization to unseen personas.

Experimental results

Research questions

• RQ1 Can a dialogue model learn to generate persona-consistent responses without access to predefined persona statements during inference?
• RQ2 Does introducing a persona reconstruction task during meta-learning improve the model's ability to infer and reflect personality from dialogue context?
• RQ3 How does multi-task meta-learning with persona reconstruction compare to standard meta-learning baselines in terms of persona consistency?
• RQ4 What is the impact of alternating vs. joint optimization of response generation and persona reconstruction on model performance?
• RQ5 Can the model reconstruct accurate persona statements from short dialogue contexts, and how does context length affect reconstruction quality?

Key findings

• In the 10-shot setting, both MTML and AMTML outperformed P²Bot in persona consistency, demonstrating improved ability to reflect persona traits in generated responses.
• The MTML and AMTML frameworks generated responses with significantly higher persona consistency compared to PAML and P²Bot, even without using persona statements at inference.
• The persona reconstruction task during meta-learning helped the model learn to infer persona information more effectively from dialogue context.
• Longer dialogue contexts led to better persona reconstruction performance, indicating that sufficient contextual cues are crucial for accurate inference.
• Despite improved persona consistency, fluency and coherence were still outperformed by P²Bot, suggesting a trade-off between persona fidelity and response quality.
• The results indicate that multi-task meta-learning with persona reconstruction is a viable and effective approach for zero-shot personalized dialogue generation.