Toward An Optimal Selection of Dialogue Strategies: A Target-Driven Approach for Intelligent Outbound Robots

Ruifeng Qian, Shijie Li, Mengjiao Bao, Huan Chen, Yu ChearXiv 2022 · DOI · arXiv

Abstract

Outbound robots are increasingly utilized in FinTech for tasks such as debt collection. However, traditional systems often rely on rigid flow-based logic or behavior-cloning models that fail to optimize for specific business outcomes. This paper introduces Policy2Target, a target-driven framework that treats dialogue strategy selection as a sequence labeling problem aimed at maximizing goal achievement. By integrating a dual-way learning model with an implicit context tracker, Policy2Target bridges the "Target-Strategy Gap." Real-world A/B tests demonstrate that this approach significantly outperforms existing systems, improving repayment rates by up to 1.60%.

Introduction: The Target-Strategy Gap

In the FinTech industry, outbound robots typically follow one of two paradigms: flow-based systems or behavior-cloning (BC) models. Flow-based systems utilize manually configured "if-then" logic, which is often too rigid to handle the nuances of human conversation. Conversely, BC models attempt to imitate human agents but often replicate suboptimal behaviors without understanding which strategies actually drive success.

Comparison of flow-based, behavior-cloning, and the proposed target-driven approach.

The authors identify a critical "Target-Strategy Gap": existing models do not explicitly link dialogue choices to business targets, such as debt repayment. Policy2Target addresses this by making the business target the primary driver of strategy selection, ensuring that the robot chooses the most effective "human-like" response.

Methodology: The Policy2Target Framework

Policy2Target decomposes conversations into a sequence of "atomic policies"—discrete actions like greetings or payment requests. The framework consists of several innovative components:

  1. Dual-way Joint Learning Model: This model simultaneously predicts the probability of achieving the final target (e.g., repayment) and the likelihood of a strategy being used by a human. This ensures the robot remains effective while sounding natural.
  2. Attention-based Policy Aggregation: To handle "mixed policies" where a single turn contains multiple strategies, an attention mechanism weights atomic policies based on user profiles and intentions.
  3. Implicit Context Tracker: A three-block memory module tracks policy history, GRU hidden states, and specific "slots" (e.g., promised amounts) to maintain conversational coherence and avoid repetitive or overly aggressive tactics.
The detailed architecture of the Policy2Target dual-way model and policy aggregation mechanism.

Data and Preprocessing

The researchers utilized a proprietary dataset from Meituan consisting of human-human dialogue corpora. To handle the noise of real-world speech, they employed BERT-Large to build a multi-label classification model. This model standardizes transcribed utterances into a finite set of atomic strategies, filtering out ASR errors and colloquialisms to focus the training on underlying strategic intent.

Key Results and Evaluation

To isolate the impact of dialogue strategy from user-specific biases (like credit scores), the authors introduced Weighted Bins AUC (WBAUC). This metric measures how well the model ranks strategies within groups of similar users.

  • Offline Performance: Policy2Target achieved a WBAUC of 0.6972, vastly exceeding the behavior-cloning baseline of 0.5116.
  • Online A/B Testing: In a live debt collection environment, the framework improved repayment rates by 1.60% over flow-based robots and 0.85% over BC models.
  • Dialogue Diversity: The system achieved a diversity score of 68.5%, more than double that of flow-based systems (26.4%), indicating a superior ability to navigate complex conversational paths.

Conclusion

Policy2Target demonstrates that prioritizing business effectiveness over simple anthropomorphism leads to superior performance in outbound robotics. While currently optimized for debt collection, the framework's reliance on target-driven strategy selection offers a promising path for other outbound applications like telemarketing and fraud prevention. Future work will focus on reducing dependency on ASR accuracy and exploring "cold start" scenarios with limited initial data.

Summary updated

BibTeX
@misc{qian2022toward,
  author = {Ruifeng Qian and Shijie Li and Mengjiao Bao and Huan Chen and Yu Che},
  title = {{Toward An Optimal Selection of Dialogue Strategies: A Target-Driven Approach for Intelligent Outbound Robots}},
  year = {2022},
  eprint = {2206.10953},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.10953},
}
Figure