PersonaPKT: Building Personalized Dialogue Agents via Parameter-efficient Knowledge Transfer

“Personalized dialogue agents (DAs) powered by large pre-trained language models (PLMs) often rely on explicit persona descriptions to maintain personality consistency. However, such descriptions may not always be available or may pose privacy concerns. To tackle this bottleneck, we introduce PersonaPKT, a lightweight transfer learning approach that can build persona-consistent dialogue models without explicit persona descriptions. By representing each persona as a continuous vector, PersonaPKT learns implicit persona-specific features directly from a small number of dialogue samples produced by the same persona, adding less than 0.1% trainable parameters for each persona on top of the PLM backbone. Empirical results demonstrate that PersonaPKT effectively builds personalized DAs with high storage efficiency, outperforming various baselines in terms of persona consistency while maintaining good response generation quality. In addition, it enhances privacy protection by avoiding explicit persona descriptions. Overall, PersonaPKT is an effective solution for creating personalized DAs that respect user privacy.”

To the best of our knowledge, our work is the first on building personalized dialogue models via parameter-efficient knowledge transfer.

Previous studies have shown that when explicit persona descriptions are available, they can be encoded into memory networks (e.g., Zhang et al., 2018) or appended to dialogue histories to generate persona-consistent responses (e.g., Wolf et al., 2019). However, the utilization of explicit persona descriptions raises privacy concerns as it may require access to personal information. To address this issue, Madotto et al. (2019) introduced a persona-agnostic meta-learning (PAML) algorithm that enables the learning of persona-specific dialogue generation models without the need for explicit persona descriptions. Subsequently, several studies have explored this direction using various methodologies (Lee et al., 2021b; Yao et al., 2021; Wu et al., 2021). For example, Lee et al. (2021b) trained persona-specific models via multitask meta-learning without any explicit persona descriptions. While the PAML algorithm and its follow-up work demonstrate the feasibility of training persona-consistent dialogue agents without explicit persona description, they still require modifying the entire language model parameters and storing a full copy of the pre-trained model for each persona. To address this limitation, our approach here provides a more storage-efficient solution for creating effective personalized DAs without the need for explicit persona descriptions.

In PersonaPKT, a personalized prefix is trained on each persona’s dialogue data only. The limited dialogue data per persona will result in a low-data resource scenario, potentially leading to a significant performance drop in terms of dialogue generation quality. In light of this, PersonaPKT is novel in introducing source prefix tuning, an extra training stage before personalized prefix tuning (Fig 1). It first trains one source prefix over multiple personas’ data agnostically and then uses the source prefix to initialize the training of the personalized prefix for a target persona. Via such a two-stage transfer learning process, PersonaPKT is able to transfer the knowledge learned from various personas to the target prefix training, preventing the generation quality from dramatically dropping due to limited training data per persona.