Bilateral Personalized Dialogue Generation with Dynamic Persona-Aware Fusion

Bin Li, Bin Sun, Shutao LiarXiv 2021 · DOI · OpenReview

TL;DR

Most personalized chatbots focus solely on maintaining their own "persona," often ignoring the user's background and leading to jarring or inappropriate interactions. This paper introduces Bilateral Personalized Dialogue Generation (BPDG), a framework that integrates both the robot's and the user's personas into the conversation using a dynamic fusion module and a contrastive learning-based selection strategy to ensure responses are consistent with both parties.

What Problem Does This Paper Address?

In human-robot interaction, "personalization" has traditionally been a one-way street. Researchers have spent years making sure a robot doesn't claim to be a 20-year-old student in one sentence and a retired doctor in the next (unilateral consistency). However, humans naturally adapt their speech based on who they are talking to.

If a robot ignores the user's persona—their interests, location, or gender—it can come across as cold, irrelevant, or even offensive. The challenge lies in "bilateral" consistency: how can a model decide when to express its own traits and when to acknowledge the user's traits without sounding forced or repetitive?

Methods

The authors propose the BPDG method, which builds upon a pre-trained GPT-2 backbone (specifically the Chinese LCCC-base model). The architecture consists of three main components:

  1. Dialogue Context & Profile Modeling: The model takes structured persona attributes (e.g., <Gender, Female>, <Area, Beijing>) and embeds them alongside the dialogue history.
  2. Dynamic Persona-Aware Fusion: This is the "brain" of the system. Instead of simply concatenating persona data, the model uses a sub-network to predict the probability of three states: expressing the user's persona, expressing the robot's persona, or staying context-neutral. These probabilities ($\alpha, \beta, \gamma$) dynamically weight the different information streams during decoding.
  3. Contrastive Learning with CMIM: Standard training objectives (like cross-entropy) often favor safe, generic responses ("I don't know"). To fix this, the authors use the Conditional Mutual Information Maximum (CMIM) criterion. They generate a list of candidate responses and use a contrastive learning-trained classifier to select the one that is most relevant to the specific combination of user, robot, and history.
Fig. 2. The overview of the proposed BPDG method.

Datasets

The study utilizes the PersonalDialog dataset, a massive corpus of dialogues crawled from the Chinese social media site Weibo.

  • Filtering: Because many profiles in the original set were sparse, the authors used a heuristic script to select 410,000 dialogues where both parties had complete personalized information (Gender, Area, and Interests).
  • Evaluation Sets: They tested on two distinct sets: a Random Set (general conversations) and a Biased Set (conversations where speakers intentionally reveal persona information).

Key Insights and Ideas

  • Personalization as a Two-Way Street: The core "aha moment" is that a robot's response shouldn't just be about who it is, but who it is talking to.
  • The Perplexity Paradox: The authors found that models with higher personalization actually have higher (worse) perplexity scores. This is because personalized words (like specific hobbies or cities) are rarer than generic words. A "perfect" perplexity score often results in a boring, non-personalized bot.
  • Dynamic Weighting: You shouldn't always talk about your persona. The Persona Presence Prediction task (seen in Figure 4) allows the model to learn when it is appropriate to bring up personal details versus when to focus purely on the conversation flow.
Fig. 4. The structure of the dynamic persona-aware fusion module.

Key Results

The BPDG model was compared against several state-of-the-art baselines, including PWDWP and TTransfo.

  • Bi-Persona Accuracy (BPAcc): The authors designed a new metric to measure how well a response matches both personas. BPDG achieved 92.14% on the biased test set, significantly outperforming the strongest baseline (88.19%).
  • Human Evaluation: In blind tests, human judges rated BPDG higher in Sentence Fluency, Bilateral Persona Consistency, and Context Consistency compared to all other models.
  • Case Study: As shown in the paper's examples, when a user mentions they are from Shanghai, the BPDG model is more likely to generate a response acknowledging Shanghai specifically, rather than a generic "That's nice."

Limitations and Future Work

The authors acknowledge that open-domain dialogue is a "one-to-many" problem—there are infinite valid ways to respond to any prompt. While BPDG improves consistency, it doesn't yet handle complex reasoning. For example, if a user says they like "hiking," the model might not yet be able to reason that they might also enjoy "national parks" unless that specific keyword is in the persona.

Future work will focus on incorporating more comprehensive reasoning strategies to further improve the depth and quality of the generated responses.

Summary updated

Abstract

Generating personalized responses is one of the major challenges in natural human-robot interaction. Current researches in this field mainly focus on generating responses consistent with the robot's pre-assigned persona, while ignoring the user's persona. Such responses may be inappropriate or even offensive, which may lead to the bad user experience. Therefore, we propose a Bilateral Personalized Dialogue Generation (BPDG) method for dyadic conversation, which integrates user and robot personas into dialogue generation via designing a dynamic persona-aware fusion method. To bridge the gap between the learning objective function and evaluation metrics, the Conditional Mutual Information Maximum (CMIM) criterion is adopted with contrastive learning to select the proper response from the generated candidates. Moreover, a bilateral persona accuracy metric is designed to measure the degree of bilateral personalization. Experimental results demonstrate that, compared with several state-of-the-art methods, the final results of the proposed method are more personalized and consistent with bilateral personas in terms of both automatic and manual evaluations.
BibTeX
@misc{li2021bilateral,
  author = {Bin Li and Bin Sun and Shutao Li},
  title = {{Bilateral Personalized Dialogue Generation with Dynamic Persona-Aware Fusion}},
  year = {2021},
  eprint = {2106.07857},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.07857},
}

Loading related content...

Figure