slotfilling

Slot filling is a natural language understanding task that extracts predefined semantic attributes, or slots, from a user utterance to populate a structured representation such as a dialogue state or task query. The resulting slot values enable downstream actions, like booking a flight or retrieving information. It is typically formulated as a sequence labeling problem, with tokens assigned slot tags (for example BIO tagging) and the labeled spans mapped to slot values. Slot filling is commonly paired with intent detection to capture the utterance's overall purpose.

Approaches range from rule-based systems to statistical and neural models. Classical methods use conditional random fields or hidden Markov models, while modern systems rely on neural networks such as BiLSTM or transformer encoders (for example BERT) with slot tagging layers. Many systems employ joint models that predict slots and intents together. Pretrained language models are fine-tuned on labeled data, and annotation usually follows BIO schemes. Standard benchmarks include ATIS and SNIPS, though production data are domain-specific.

Evaluation uses slot-level F1 scores and semantic frame accuracy. Challenges include domain adaptation, data sparsity for rare slots, multi-turn dialogues, overlapping or nested slots, and robustness to speech recognition errors. Solutions emphasize data augmentation, transfer learning, few-shot learning, and active learning to improve generalization.

Applications include voice assistants, chatbots, and information extraction pipelines in travel, hospitality, customer support, and e-commerce.