Abstract
Multi-label intent detection aims to assign multiple labels to utterances, a critical task for modern task-oriented dialogue systems. As dialogue domains evolve and new intents emerge, the scarcity of annotated data necessitates multi-label few-shot intent detection (ML-FSID). Previous approaches often struggle with identical representations for utterances containing multiple labels and fail to account for intrinsic intra-class and inter-class interactions. To address these limitations, we propose the Dual Class Knowledge Propagation Network (DCKPN). Our framework introduces a label-semantic augmentation module that incorporates class name information to learn well-separated representations. Furthermore, we construct instance-level and class-level graph neural networks to propagate both label information and feature structures. Experimental results on two multi-label intent datasets demonstrate that DCKPN significantly outperforms strong baselines.
Introduction
In task-oriented dialogue systems, intent detection is the foundational step in understanding user requests. While single-label classification is well-studied, real-world interactions often involve multi-intent utterances (e.g., "What time is my meeting and where is it located?"). Detecting these intents in a few-shot setting—where only a handful of labeled examples are available—presents two major challenges:
- Representation Confusion: Standard models often produce a single, blended embedding for an utterance with multiple intents, making it difficult for classifiers to isolate individual components.
- Overlooked Relationships: Existing models frequently treat intents as independent, ignoring high-order correlations between classes (e.g., the likelihood that a request for a 'date' is accompanied by a request for a 'time').
Methodology
DCKPN addresses these hurdles through a novel architecture consisting of three primary components: Label-semantic Augmentation, a Dual-Graph Architecture, and Adaptive Intent Counting.
Label-semantic Augmentation (LA)
To resolve representation confusion, we append label names directly to the utterance (e.g., [CLS] utterance [SEP] label_name [SEP]). This process transforms a single multi-label example into multiple label-specific samples. By forcing the encoder to process the utterance in the context of a specific intent name, the model learns more distinct, intent-aware features.
Dual Class Knowledge Propagation Network
We utilize two interconnected graphs to model complex interactions:
- Instance-level Graph: This graph connects similar utterances to smooth their features, allowing the model to effectively generalize from the limited support set to query instances.
- Class-level Graph: This component models the relationships between the intents themselves. By leveraging the instance-level graph to generate pseudo-class representations for query samples, the model can infer label correlations even for previously unseen data.
Adaptive Intent Counting
Instead of relying on a fixed similarity threshold to determine intent presence, DCKPN employs a simple Multi-Layer Perceptron (MLP) to predict the exact number of intents within a given utterance. This adaptive approach is more robust to the varying complexity of user queries.
Experimental Evaluation
We evaluated DCKPN on two benchmark datasets: TourSG (tourism in Singapore) and StanfordLU (Schedule, Navigate, and Weather domains). We used Electra-small and BERT-base as the underlying text encoders.
Key Results
DCKPN achieved a new state-of-the-art across both datasets, particularly in the challenging 1-shot setting:
- StanfordLU (1-shot): DCKPN reached an F1 score of 56.58%, a dramatic 14% absolute improvement over the previous best model, CTLR (42.51%).
- TourSG (1-shot): DCKPN achieved 54.93% F1, outperforming CTLR's 51.07%.
Ablation studies confirmed that Label-semantic Augmentation is the most critical component, as removing it led to the most significant performance degradation. This highlights the value of using intent names as a "free" semantic signal.
Conclusion
DCKPN demonstrates that incorporating label semantics and modeling dual-level interactions (instance and class) provides a powerful framework for multi-label few-shot intent detection. By treating semantics as a signal and allowing bidirectional knowledge flow between graphs, the model achieves superior generalization. Future work will explore more sophisticated prompt engineering and the integration of intent detection with slot filling for holistic dialogue understanding.
Summary updated