A Personalized Utterance Style (PUS) based Dialogue Strategy for Efficient Service Requirement Elicitation

Demin Yu, Min Liu, Zhongjie WangarXiv 2023 · DOI · arXiv

Abstract

Capturing user requirements for complex, multi-domain services remains a significant challenge for dialogue systems. Traditional approaches often rely on static slots, which struggle with the massive search space of hierarchical and subjective requirements. This paper introduces the Personalized Utterance Style (PUS) module, a novel framework that models individual user habits—specifically act state transfers, requirement planning, and attribute preferences. By perceiving these personalized expression habits, the system can "shrink" the requirement search space, leading to more efficient elicitation. Evaluations on the CrossWOZ dataset demonstrate that the PUS-based strategy achieves higher accuracy with fewer dialogue rounds compared to traditional methods.

Introduction: The Challenge of Service Requirement Elicitation

Service Requirement Elicitation (SRE) involves capturing user needs for complex tasks, such as planning a multi-city trip. Unlike simple slot-filling tasks (e.g., booking a single flight), SRE requirements are often hierarchical, transboundary (spanning multiple domains), and highly subjective.

Traditional dialogue systems typically utilize "static slots" like [destination] or [price_range]. However, in complex SRE scenarios, the potential requirement space is vast. Without understanding a user's specific "style"—the sequence in which they prefer to provide information or their typical response patterns—conversations can become repetitive and inefficient. This research aims to bridge this gap by leveraging personalized expression habits to streamline the requirement gathering process.

Methodology: The PUS Framework

The proposed framework centers on two primary components: the Requirement Tree ($R\_Tree$) and the Personalized Utterance Style (PUS) module.

The Requirement Tree ($R\_Tree$)

Instead of a flat list of slots, the system models requirements as a tree structure. This hierarchical approach captures dependencies; for instance, selecting a "Hotel" might naturally trigger a sub-requirement for a "Pick-up Service."

The Personalized Utterance Style (PUS) Module

The PUS module is the core innovation, characterizing users across three distinct dimensions:

  1. Act State Transfer (AST) Preference: Utilizes weighted Finite State Transducers (wFST) to model how a specific user transitions between dialogue acts (e.g., moving from a statement to a question).
  2. Requirement Planning (RP) Preference: Employs Factorizing Personalized Markov Chains (FPMC) to predict the order in which a user prefers to discuss various topics.
  3. Requirement Attribute (RA) Preference: A key-value structure that tracks historical constraints, such as a preference for "luxury" over "budget" options.
The overall architecture of the dialogue system, highlighting the integration of the PUS module within the dialogue state and the requirement-driven strategy.

The dialogue strategy operates via an inner-outer loop process. The outer loop identifies the next requirement node to discuss based on RP preferences, while the inner loop manages specific dialogue acts using AST and RA preferences.

Experimental Setup

The framework was evaluated using CrossWOZ, a large-scale Chinese multi-domain task-oriented dialogue dataset. The researchers re-annotated the dataset to focus on requirement expression habits rather than simple entity search. An agenda-based user simulator was implemented to test the strategy across thousands of simulated interactions.

Key Insights: Habit as a Compass

A central insight of this work is that how a user communicates is as informative as what they communicate. By predicting a user's next likely move, the system can preemptively prepare for transitions, effectively ignoring irrelevant branches of the requirement tree. This significantly reduces the cognitive load on the user and speeds up the elicitation process.

An example dialogue demonstrating how the system captures intentions, recommends constraints based on patterns, and plans future steps using the PUS module.

Furthermore, the use of wFST for AST preference provides interpretability. Unlike "black-box" neural models, the system can explicitly track the conversation path, offering better symbolic control over the dialogue flow.

Results and Discussion

The PUS-based strategy demonstrated superior efficiency, eliciting accurate requirements in significantly fewer dialogue rounds than standard requirement-based policies. In medium-scale scenarios, the system maintained a higher success rate by selecting response actions tailored to the user's history.

Integration with deep learning models also showed promise. When PUS features were vectorized and added to the Dialogue Analysis and Management (DAMD) model, Joint Goal accuracy improved from 46.2% to 47.6%, alongside boosts in Slot F1 scores. This suggests that personalized style features are valuable even within end-to-end neural architectures.

Conclusion and Future Work

While the PUS module currently models three dimensions of style, future work will explore more complex factors such as sentiment and varying levels of user cooperativeness. Additionally, the researchers plan to test the scalability of the strategy in even broader, "transboundary" service environments where requirements might span medical, financial, and travel domains simultaneously.

Summary updated

BibTeX
@misc{yu2023personalized,
  author = {Demin Yu and Min Liu and Zhongjie Wang},
  title = {{A Personalized Utterance Style (PUS) based Dialogue Strategy for Efficient Service Requirement Elicitation}},
  year = {2023},
  eprint = {2301.04582},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.04582},
}
Figure