Slot-Gated Modeling for Joint Slot Filling and Intent Prediction

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/N18-2118.pdf

Summary

Spoken language understanding (SLU) usually involves two linked subtasks: classifying the intent of an utterance and labeling each word with a semantic slot. Earlier joint attention-based RNN models trained both subtasks under a shared loss, but each subtask had its own attention weights. Any interaction between slots and intent was therefore learned only implicitly. Goo et al. add a "slot gate" to a BLSTM attention model. The gate combines the intent context vector with each word's slot context vector and produces a scalar weight that scales the slot context before the slot label is predicted. Slot tagging is thus explicitly conditioned on intent information.

They test two variants, one with full (slot and intent) attention and a lighter one with intent attention only. Both are evaluated on ATIS and Snips against the joint sequence model of Hakkani-Tür et al. and the attention model of Liu and Lane. The gated models improve slot F1, intent accuracy and especially whole-frame accuracy. This shows that explicitly modeling slot–intent dependency benefits joint SLU.

Key points

• Datasets: ATIS has 4,478 / 500 / 893 utterances for train / dev / test, with 120 slot labels and 21 intents. It is single-domain and heavily skewed, with atis_flight making up about 74% of training data. Snips (custom-intent-engines) has 13,084 / 700 / 700 utterances, 72 slots and 7 balanced intents across diverse topics.
• ATIS results: The intent-attention-only variant scored best, with 95.2 slot F1, 94.1% intent accuracy and 82.6% sentence accuracy. The attention baseline scored 94.2, 91.1 and 78.9. This variant also uses fewer parameters (251K vs. 284K).
• Snips results: The full-attention gated model scored best, with 88.8 F1, 97.0% intent accuracy and 75.5% sentence accuracy, compared with 74.1% sentence accuracy for the attention baseline. The authors suggest that the more complex Snips data still needs separate slot attention.
• Relative gains: The paper reports sentence-level gains of roughly 4.2% on ATIS and 1.9% on Snips over the attention model. The paper's own text gives about 4.1% for ATIS in one place. Significance was checked with a one-tailed t-test (p < 0.05).
• Setup: Hidden size 64, the Adam optimizer, early stopping, and scores averaged over 20 runs. Code is available at github.com/MiuLab/SlotGated-SLU.