An Intelligent Assistant for Problem Behavior Management

Penghe Chen, Yu Lu, Jiefei Liu, Qi XuAAAI 2021 · DOI

Introduction

Managing student problem behaviors—actions that deviate from social norms and cause concern—is a cornerstone of moral education. However, effective management requires deep psychological expertise that many young teachers and parents lack. While psychological literature offers guidelines, they are often scattered and difficult to apply in real-world scenarios.

PB-Advisor is an AI-driven assistant designed to bridge the gap between theoretical psychology and practical application. By combining a task-oriented dialogue system for diagnosis with a BERT-powered community question-answering (CQA) module, the system identifies underlying "need deficiencies" (based on Maslow’s hierarchy) and provides evidence-based management strategies. It aims to help adults understand the why behind a behavior, offering professional suggestions rather than generic advice.

Methodology

The PB-Advisor architecture is built on four primary pillars designed to support both specific diagnosis and general pedagogical inquiry.

1. Dialogue Diagnosis Module

This task-oriented system identifies "need deficiencies" (unfulfilled needs like safety or esteem) through structured interaction. It utilizes a Long Short-Term Memory (LSTM) network for Natural Language Understanding (NLU) to interpret user intent and extract semantic slots. The dialogue policy employs a Deep Q-Network (DQN) trained on real-life cases to decide whether to request more information or inform the user of a diagnosis. This entire process is grounded in a Psychological Knowledge Graph that defines relationships between behaviors, needs, and environmental factors.

The architecture of the Dialogue Diagnosis Module, showing the flow from NLU to Policy Learning and the Knowledge Graph.

2. Community Question Answering (CQA) Module

For general pedagogical questions (e.g., "How do I apply appreciation education?"), the system uses a dual-layer approach to match user queries with expert discussions. A BERT model analyzes the semantics of the user query against a database of historical expert discussions, while a Multi-Layer Perceptron (MLP) ranks these questions based on a semantic relevance vector to find the best match.

The BERT-based CQA module architecture, illustrating the semantic matching and ranking process.

3. Learning Center and Data Analyzer

The system also includes a Learning Center, which serves as a resource hub for theoretical frameworks and case studies, and a Data Analyzer. The analyzer uses descriptive statistics and Random Forest models to discover patterns in behavior data, such as how parenting styles correlate with specific student needs.

Key Insights and Results

PB-Advisor shifts the focus from punishment to support by treating behavior as a symptom of an underlying unmet need. By offering both structured dialogue and unstructured CQA, the system accommodates different user information-seeking states.

Key performance metrics include:

  • Dialogue Performance: The task-oriented dialogue system achieved a success rate of approximately 44%, with an average of 11 turns per dialogue.
  • CQA Accuracy: The BERT-based question matching system achieved a classification accuracy of approximately 80%.
  • Analytical Validation: The Data Analyzer quantitatively demonstrated that parenting styles significantly influence student need deficiencies, validating the system's ability to find meaningful correlations in behavioral data.

Conclusion

While PB-Advisor is an assistant tool meant to provide suggestions rather than replace professional intervention, its current performance demonstrates the feasibility of using AI to deliver psychological expertise. Future work will focus on improving dialogue policy and NLU to handle more complex user inputs and deploying the system more broadly to gather more data and serve a larger population of educators and parents.

Summary updated

BibTeX
@inproceedings{chen2021intelligent,
  author = {Penghe Chen and Yu Lu and Jiefei Liu and Qi Xu},
  title = {{An Intelligent Assistant for Problem Behavior Management}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17995},
}
Figure