Abstract
Profile-based intent detection and slot filling are important tasks aimed at reducing the ambiguity in user utterances by leveraging user-specific supporting profile information. However, research in these two tasks has not been extensively explored. To fill this gap, we propose a joint model, namely JPIS, designed to enhance profile-based intent detection and slot filling. JPIS incorporates the supporting profile information into its encoder and introduces a slot-to-intent attention mechanism to transfer slot information representations to intent detection. Experimental results show that our JPIS substantially outperforms previous profile-based models, establishing a new state-of-the-art performance in overall accuracy on the Chinese benchmark dataset ProSLU.
Introduction
Most virtual assistants rely primarily on the text of a user's utterance to determine intent. However, human language is inherently ambiguous. For example, a command like "Book a ticket to Hanoi" could refer to a flight, train, or bus depending on the context. Profile-based Spoken Language Understanding (SLU) attempts to resolve this by incorporating external knowledge: long-term user preferences (User Profiles) and real-time situational data (Context Awareness).
While joint models for Intent Detection (ID) and Slot Filling (SF) are common, they typically treat intent as the primary task that guides slot filling. JPIS challenges this paradigm by arguing that the relationship should be a two-way street: recognizing specific entities (slots) can significantly help clarify the user's overall goal (intent).
Methodology
The JPIS model consists of four main components designed to integrate profile information and facilitate information transfer between tasks:
- Utterance Encoder: Processes input text using a BiLSTM and self-attention. It employs a multiplicative attention mechanism to inject user profile and context vectors directly into word representations.
- Slot-to-Intent Attention: The core innovation of this work. It creates label-specific representations for both intents and slots and uses a bilinear attention matrix to calculate similarity, effectively transferring knowledge from the slot-filling task back to intent detection.
- Intent Decoder: Predicts the overall intent using the weighted representations derived from the slot-to-intent mechanism.
- Slot Decoder: Uses a Linear-chain Conditional Random Field (CRF) to predict slot labels, incorporating the predicted intent label to maintain consistency across the joint task.
Experimental Setup
The model was evaluated on ProSLU, the primary Chinese benchmark dataset for profile-based SLU. The dataset contains 4,196 training, 522 validation, and 531 test utterances. Each utterance is paired with four user-associated feature vectors and four context awareness vectors to provide the necessary background for disambiguation.
Results and Discussion
JPIS established a new state-of-the-art (SOTA) on the ProSLU dataset, outperforming previous models across all key metrics.
Key Insights
- Information Flow: The success of JPIS proves that slot information is a powerful disambiguator for intent. For instance, identifying a "song name" slot makes the "Play Music" intent significantly more probable than others.
- The Necessity of Context: Ablation studies highlighted the critical role of profile information. Removing user and context data caused intent accuracy to plummet from approximately 88% to 42%, demonstrating that for specific domains, the "who" and "where" are as vital as the "what."
- Pre-trained Model Integration: When combined with the Chinese ELECTRA model, JPIS's overall accuracy reached 86.67%, further widening the gap over non-profile-based approaches.
Conclusion and Future Work
JPIS demonstrates that a bidirectional information flow between intent detection and slot filling, combined with robust profile integration, significantly improves SLU performance. While currently validated on a Chinese dataset, the architecture is generalizable. Future research will focus on handling more complex, dynamic user profiles—such as those stored in Knowledge Graphs—and addressing potential computational overhead as label sets expand.
Summary updated