Introduction
Modern virtual assistants are increasingly expected to handle complex, multi-layered user queries. While single-intent Spoken Language Understanding (SLU) is a mature field, multi-intent SLU—where a single utterance contains multiple requests—remains a significant challenge. For instance, a user might say, "List the flight number between LA and Seattle and show me the cheapest ticket." This requires identifying two distinct intents ("Atis_Flight_No" and "Atis_Airfare") and correctly mapping slots like locations and price preferences to the appropriate intent.
The Multi-View Approach
The core innovation of the Uni-MIS framework, presented at AAAI-24, is treating multi-intent SLU as a "multi-view" problem. Traditional models often rely solely on token-level interactions, which can lead to error propagation if local word-level noise confuses the model. Uni-MIS addresses this by capturing intent-slot interactions at three distinct levels:
- Utterance-level: Captures the global semantic context and the overall "vibe" of the entire sentence.
- Chunk-level: Identifies transition points between different intents using a sliding window mechanism to find where one request ends and another begins.
- Token-level: Provides fine-grained, word-by-word guidance for precise slot filling.
Methodology
The Uni-MIS architecture utilizes a RoBERTa encoder to generate contextualized word representations. These representations are then processed by task-specific BiLSTMs for intent detection and slot filling. The authors argue that intent information is "anisotropic," meaning its relevance varies across different scales; global context prevents the model from being distracted by local noise, while chunk-level information is critical for identifying boundaries between multiple requests.
Following the multi-view interaction layer, the model employs an Adaptive Intent-Slot Graph Interaction module. This module explicitly models the dependencies between predicted intents and slots using a graph neural network, ensuring that the final output is logically consistent and reducing the likelihood of misaligned predictions.
Experimental Results
Uni-MIS was evaluated on two primary benchmarks: MixATIS (airline-related queries) and MixSNIPS (multi-domain queries). The model achieved state-of-the-art (SoTA) performance on both:
- MixATIS: Achieved an Overall Accuracy of 52.5%, a 3.1% improvement over previous baselines.
- MixSNIPS: Achieved an Overall Accuracy of 83.4%.
Notably, the performance advantage of Uni-MIS becomes more pronounced as the complexity of the utterance (i.e., the number of intents) increases, proving its robustness for real-world applications.
The LLM Performance Gap
A striking finding of this research is the significant performance gap between specialized SLU models and general-purpose Large Language Models (LLMs) like ChatGPT (gpt-3.5-turbo). Despite its broad reasoning capabilities, ChatGPT struggled with the structured sequence labeling required for slot filling. On the MixATIS dataset, ChatGPT achieved an overall accuracy of only 0.36%, highlighting the continued necessity for specialized architectures like Uni-MIS in production-grade SLU systems.
Conclusion
Uni-MIS demonstrates that integrating global, local, and transitional perspectives through multi-view interaction significantly enhances the robustness of multi-intent SLU. By mitigating error propagation and leveraging structured graph interactions, Uni-MIS sets a new standard for processing complex human speech in conversational AI.
Summary updated