Source-linked AI summary
Exploiting Persona Information for Diverse Generation of Conversational Responses
Haoyu Song, Wei-Nan Zhang, Yiming Cui, Dong Wang, Ting Liu
TL;DR
Persona-based dialogue models must handle multiple appropriate responses rather than predict a single target, while preserving coherent persona information. The paper introduces Persona-CVAE, combining memory-augmented persona modeling with a conditional variational autoencoder, and reports more diverse and engaging persona-based responses than baselines.
Problem
Existing persona dialogue models generally assume one target response despite human dialogue’s many valid responses, limiting response diversity.
Method
Persona-CVAE combines external persona memory, a latent variable in a conditional variational autoencoder, and decoding strategies for generating persona-based responses.
Results
Persona-CVAE outperforms state-of-the-art methods, especially in response diversity and persona integration, across automatic and human evaluations.
Takeaways & Limitations
The model delivers more diverse, engaging, and persona-informed responses than baseline approaches on the ConvAI2 persona-chat dataset.
Takeaways & Limitations
The approach models explicit chatbot persona and assumes persona differs from dialogue context by remaining unchanged and unilateral.
Abstract
from arXiv · showhide
In human conversations, due to their personalities in mind, people can easily carry out and maintain the conversations. Giving conversational context with persona information to a chatbot, how to exploit the information to generate diverse and sustainable conversations is still a non-trivial task. Previous work on persona-based conversational models successfully make use of predefined persona information and have shown great promise in delivering more realistic responses. And they all learn with the assumption that given a source input, there is only one target response. However, in human conversations, there are massive appropriate responses to a given input message. In this paper, we propose a memory-augmented architecture to exploit persona information from context and incorporate a conditional variational autoencoder model together to generate diverse and sustainable conversations. We evaluate the proposed model on a benchmark persona-chat dataset. Both automatic and human evaluations show that our model can deliver more diverse and more engaging persona-based responses than baseline approaches.
1 Introduction
Open-domain dialogue requires coherent persona modeling and multiple appropriate responses, but existing persona systems often inherit Seq2Seq’s one-response and low-diversity limitations.
- Persona helps chatbots deliver more realistic, trustworthy, and engaging conversations by maintaining coherent personal information.
- Human dialogue has multiple valid, non-repetitive responses for the same context, enabling varied dialogue strategies and greater diversity.
- Seq2Seq training predicts one target response with maximum likelihood, encouraging persona models to generate repetitive and non-specific replies.
- Persona-CVAE combines explicit persona text with a latent variable, memory augmentation, and decoding strategies that generate or copy persona-grounded words.
2 Related Work
Related work addresses persona modeling and response diversity through speaker or profile representations, objective changes, richer encoders, and conditional variational autoencoders.
- Persona dialogue models use speaker embeddings, context conditioning, or profile information, but some approaches require expensive and sparse speaker-tagged data.
- Dialogue-generation research targets dull responses by modifying the training objective or enriching the generative model’s encoder with additional information.
- Dialogue CVAE models latent discourse-level variation and generates diverse responses by sampling from the learned distribution.
- Persona-CVAE extends dialogue CVAE with external persona memory, new decoding strategies, and the ability to produce persona-based responses.
3 Persona-CVAE
Persona-CVAE combines persona memory with a conditional latent-variable model to generate multiple non-repetitive responses from dialogue context and persona texts. Its architecture selects relevant persona information and uses soft or force decoding strategies to express it during generation.
- Persona based CVAE: Persona-CVAE models diverse responses with latent variable z conditioned on dialogue context and persona texts, then decodes responses from the learned distribution.The task defines diversity as non-repetitive responses covering different aspects and differing in words.
- Persona Memory: Multi-hop attention converts persona texts into a persona memory, while a separate selector chooses a persona or None for response generation.The selected persona words are subsequently used by the response decoder.
- Persona Memory: The persona selector combines persona memory with sampled z to choose the persona with maximal probability, allowing the model to omit persona information when appropriate.The selected persona is determined by the maximal type probability and passed to the response decoder.
- Decoding Strategy: Soft decoding mixes persona-word and other-word distributions at each step, whereas force decoding feeds selected persona words after a matching prefix is detected.Force decoding is used once per decoding process and was reported to work well in some situations.
- Persona Memory: Three attention hops perform better than one or two, while increasing from four to five produces no significant performance change.This comparison reports the model’s observed performance across the tested hop counts.
- Training and Optimization: Training maximizes the variational lower bound while jointly optimizing persona-memory and type-distribution cross-entropy losses.The model combines these losses and optimizes them through backpropagation.
4 Experiment
Experiments on ConvAI2 compare Persona-CVAE with persona-free and persona-based baselines using automatic, human, preference, and ablation evaluations. Persona-CVAE achieves the strongest reported overall results for diverse, persona-integrated, engaging responses under the evaluated settings.
- Dataset and setup: The experiments use ConvAI2, an extended persona-chat benchmark containing 164,356 utterances across 10,981 dialogues and 1,155 personas.The final split contains 9,181 training, 800 validation, and 1,016 test dialogues, with 200 unseen test personas.
- Evaluation metrics: Automatic evaluation uses Distinct-1 and Distinct-2 for word-level diversity and Persona Coverage for expression of predefined persona information.BLEU and perplexity are omitted because the goal is to generate multiple diverse responses rather than predict one target response.
- Automatic evaluation: When generating N=5 or 10 responses, Persona-CVAE achieves the best performance across all automatic diversity and Persona Coverage metrics against every baseline.With N=1, it still exceeds all baselines in diversity but trails the Oracle model in Persona Coverage; generating too many responses slightly reduces coverage.
- Human evaluation: Persona-CVAE outperforms all baselines on engagingness, variety, and persona detection in human evaluation with N=5.Inter-rater agreement is moderate for engagingness, almost perfect for variety, and substantial for persona detection.
- Preference test: In pairwise preference tests, Persona-CVAE is significantly preferred over the other two persona-based models at p < 0.01.The comparison ranks top-response quality first, followed by response variety and persona coverage.
- Ablation tests: Removing the SDS decoding strategy produces the largest decrease in Persona Coverage but also improves Distinct-1.The reported trade-off is attributed to SDS encouraging profile words, which makes low-frequency words less likely to appear.
5 Conclusion and Future Work
The paper presents Persona-CVAE, a memory-augmented model for diverse persona-based response generation, and reports improvements over state-of-the-art methods in diversity and persona integration.
- Persona-CVAE outperforms state-of-the-art methods, especially in diversity and persona integration, on the ConvAI2 dataset.