Abstract
Understanding user intention in natural dialogue remains a significant challenge for Conversational AI. Traditional methods often rely on either rigid, brittle schemas or overly vague unstructured text. This paper introduces IntentionFrame, a semi-structured framework that decomposes user intentions into four critical dimensions: situation, emotion, action, and knowledge. To scale this framework effectively, we present WeRG (Weakly-supervised Reinforced Generation), a training methodology that utilizes reinforcement learning to bridge the gap between high-quality human annotations and large-scale noisy data. Experimental results on benchmarks like DuRecDial and ESConv demonstrate that IntentionFrame significantly improves model understanding and response quality.
Introduction
Effective Conversational Understanding (CU) requires a balance between machine-readability and the expressive nuance of human speech. Existing approaches generally fall into two categories: rigid schemas (e.g., intent classes and slot-filling) which lack flexibility, and unstructured free-text summaries which lack the consistency needed for reliable model grounding.
As Large Language Models (LLMs) evolve, there is a growing need for a representation that captures emotional subtext and situational context while providing a structured foundation for response generation. IntentionFrame addresses this by providing a "middle ground" representation that is rich enough for human-like nuance yet structured enough for machine learning.
The IntentionFrame Framework
Inspired by cognitive psychology, IntentionFrame breaks down every user turn into four interrelated aspects:
- Situation: The physical or situational context (e.g., "Pressure at work").
- Emotion: The user's psychological state (e.g., "Expresses admiration").
- Action: The communicative goal (e.g., "Shares knowledge about Jacky Cheung").
- Knowledge: Specific entities or facts mentioned (e.g., achievements in music and film).
By explicitly identifying these components, the framework acts as a "semantic pointer," focusing the LLM's attention on the most relevant parts of the dialogue history. Unlike rigid ontologies, IntentionFrame allows for free-form text within its categories, preserving natural language nuance while maintaining a consistent structure for the model to follow.
Weakly-supervised Reinforced Generation (WeRG)
A primary hurdle for complex frameworks like IntentionFrame is the high cost of human annotation. To overcome this, the authors developed the WeRG methodology, which employs a "coarse-to-fine" supervision hierarchy to train a policy model (such as Llama-2-7b).
The pipeline utilizes three levels of data quality:
- Coarse labels: Automatically mapped from existing datasets.
- Mid-level labels: Generated by cost-effective LLMs (e.g., GPT-3.5).
- Fine labels: High-quality human annotations, typically comprising only 10% of the training set.
The model is fine-tuned using Reinforcement Learning (RL) with a tiered reward system ($r_{coarse} < r_{mid} < r_{fine}$). This approach incentivizes the model to prioritize the high-fidelity nuances of human labels while still benefiting from the broad patterns present in larger, noisier datasets.
Experimental Results
The framework was evaluated on DuRecDial (recommendation) and ESConv (emotional support) datasets. Key findings include:
- Improved Performance: IntentionFrame + WeRG consistently outperformed Direct Prompting and Chain-of-Thought (CoT) baselines in human evaluations for Informativeness and Understanding.
- Downstream Impact: Guiding response generation with IntentionFrame led to a 4-6% increase in Success Rate (SR) on the DuRecDial dataset.
- Ablation Studies: Removing any of the four aspects resulted in performance degradation. Notably, the "Emotion" aspect proved most critical in emotional support scenarios (ESConv), where capturing psychological state is paramount.
Conclusion
IntentionFrame and the WeRG training method provide a robust path forward for fine-grained conversational understanding. By combining structured flexibility with a tiered reinforcement learning approach, the system achieves high performance without requiring exhaustive human labeling. Future work will focus on dynamic reward calibration and expanding the framework to safety-critical domains such as toxic behavior detection.
Summary updated