Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

K-PERM: Personalized Response Generation Using Dynamic Knowledge Retrieval and Persona-Adaptive Queries

Kanak Raj, Kaushik Roy, Vamshi Bonagiri, Priyanshul Govil, Krishnaprasad Thirunarayanan, Manas Gaur

TL;DR

K-PERM addresses the challenge of personalized response generation by jointly leveraging dynamic knowledge retrieval and persona-adaptive queries. It selects a subset of user personas via a dedicated Persona Selector and grounds responses with retrieved knowledge through a Knowledge Retriever, then trains a reward-modulated generator to balance coherence and factual alignment. Evaluations on FoCus show state-of-the-art performance with notable gains for downstream LLMs like GPT-3.5 when augmented by K-PERM, underscoring the practical impact of retrieval-augmented personalization. The framework is model-agnostic and extensible to other domains, offering a scalable path to more contextually aware and persona-consistent conversational agents.

Abstract

Personalizing conversational agents can enhance the quality of conversations and increase user engagement. However, they often lack external knowledge to appropriately tend to a user's persona. This is particularly crucial for practical applications like mental health support, nutrition planning, culturally sensitive conversations, or reducing toxic behavior in conversational agents. To enhance the relevance and comprehensiveness of personalized responses, we propose using a two-step approach that involves (1) selectively integrating user personas and (2) contextualizing the response with supplementing information from a background knowledge source. We develop K-PERM (Knowledge-guided PErsonalization with Reward Modulation), a dynamic conversational agent that combines these elements. K-PERM achieves state-of-the-art performance on the popular FoCus dataset, containing real-world personalized conversations concerning global landmarks. We show that using responses from K-PERM can improve performance in state-of-the-art LLMs (GPT 3.5) by 10.5%, highlighting the impact of K-PERM for personalizing chatbots.

K-PERM: Personalized Response Generation Using Dynamic Knowledge Retrieval and Persona-Adaptive Queries

TL;DR

K-PERM addresses the challenge of personalized response generation by jointly leveraging dynamic knowledge retrieval and persona-adaptive queries. It selects a subset of user personas via a dedicated Persona Selector and grounds responses with retrieved knowledge through a Knowledge Retriever, then trains a reward-modulated generator to balance coherence and factual alignment. Evaluations on FoCus show state-of-the-art performance with notable gains for downstream LLMs like GPT-3.5 when augmented by K-PERM, underscoring the practical impact of retrieval-augmented personalization. The framework is model-agnostic and extensible to other domains, offering a scalable path to more contextually aware and persona-consistent conversational agents.

Abstract

Personalizing conversational agents can enhance the quality of conversations and increase user engagement. However, they often lack external knowledge to appropriately tend to a user's persona. This is particularly crucial for practical applications like mental health support, nutrition planning, culturally sensitive conversations, or reducing toxic behavior in conversational agents. To enhance the relevance and comprehensiveness of personalized responses, we propose using a two-step approach that involves (1) selectively integrating user personas and (2) contextualizing the response with supplementing information from a background knowledge source. We develop K-PERM (Knowledge-guided PErsonalization with Reward Modulation), a dynamic conversational agent that combines these elements. K-PERM achieves state-of-the-art performance on the popular FoCus dataset, containing real-world personalized conversations concerning global landmarks. We show that using responses from K-PERM can improve performance in state-of-the-art LLMs (GPT 3.5) by 10.5%, highlighting the impact of K-PERM for personalizing chatbots.
Paper Structure (19 sections, 4 equations, 3 figures, 4 tables)

This paper contains 19 sections, 4 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. $\mathcal{K}$-PERM
  4. Knowledge Retriever
  5. Persona Selector ($P_\text{select}$)
  6. Response Generation through Reward Modulation
  7. Reward Function
  8. Persona-tailored Reward Function
  9. Experiments
  10. Evaluation Criteria
  11. Baselines
  12. Results and Discussion
  13. Conclusion
  14. Limitations
  15. Dataset
  16. ...and 4 more sections

Figures (3)

  • Figure 1: $\mathcal{K}$-PERM Model Architecture Overview. The model architecture comprises a Persona Selector and Knowledge Extractor, which leverages the history and question prompt to identify pertinent persona and knowledge.
  • Figure 2: $\mathcal{K}$-PERM improves personalization in GPT 3.5 via zero-shot prompting. This experiment aimed to assess the performance improvement of GPT 3.5 when combined with $\mathcal{K}$-PERM. M1 is GPT 3.5 and M2, M3, and M4 represents zero-shot prompting of GPT 3.5 using responses from $\mathcal{K}$-PERM with (All P+$Z_k$), (GP+$Z_k$), and ($P_\text{select}$+$Z_k$) respectively.
  • Figure 3: $\mathcal{K}$-PERM was preferred 32% more than ChatGPT by the annotators based on a blind evaluation of 90 queries taken randomly from the FoCus dataset. GOLD is the ground truth in FoCus Dataset.