Introduction
Spoken Language Understanding (SLU) aims to map user utterances into structured semantic representations, typically formatted as (Act, Slot, Value) triples. For instance, "I want a cheap Italian restaurant" maps to inform(price=cheap, food=Italian). Traditional SLU systems often rely on word-level alignment, which is costly to annotate and fragile in the face of Automatic Speech Recognition (ASR) errors. This paper addresses the challenge of unaligned SLU, where only sentence-level labels are available, by proposing a hierarchical approach that mirrors the natural structure of semantic triples.
Methodology
The proposed Hierarchical Decoding (HD) model decomposes the SLU task into a recursive process: identifying acts, then slots for those acts, and finally the corresponding values.
The architecture consists of four primary components:
- Shared Utterance Encoder: A bidirectional LSTM (BLSTM) that generates a contextual representation of the input utterance.
- Act Type Classifier (ATC): A multi-label classifier that identifies the dialogue acts present in the utterance (e.g.,
inform,request). - Slot Type Classifier (STC): Conditioned on the predicted act and the utterance vector, this module identifies relevant slots (e.g.,
price,food). - Value Decoder with Pointer Network: A sequence-to-sequence LSTM that generates values. By incorporating a Pointer Network, the model can either generate values from a fixed vocabulary or extract them directly from the utterance, effectively handling Out-of-Vocabulary (OOV) entities.
Experimental Setup
The model was evaluated using the DSTC2 (Dialog State Tracking Challenge 2) dataset, comprising 11,677 training and 9,890 testing pairs. Experiments were conducted on both manual transcriptions and noisy "1-best" ASR hypotheses to test robustness against speech recognition errors.
Key Results and Discussion
The HD model achieved a state-of-the-art F1-score of 86.9% on the DSTC2 benchmark, outperforming existing neural architectures and statistical baselines.
A significant advantage of this hierarchical approach is its ability to generalize to unseen semantic combinations. Because the model learns representations for acts and slots independently, it can predict triples that were never encountered during training.
As shown in the results, the HD model maintains a high F1-score (72.6% on manual transcriptions) for unseen labels, whereas traditional semantic tuple classifiers (ST) fail completely (0.0%) in these scenarios. Furthermore, the model demonstrates high data efficiency, maintaining performance even with significantly reduced training sets.
Conclusion
By leveraging the inherent hierarchy of semantic triples, the HD model provides a robust solution for unaligned SLU. Its ability to handle OOV values via pointer networks and generalize to unseen act-slot combinations marks a significant step forward in building flexible dialogue systems. Future work will explore domain adaptation and more sophisticated joint training strategies to further enhance performance across diverse conversational contexts.
Summary updated