Slot Filling — NER-NLU Joint Extraction
Slot Filling (NER-NLU Joint Extraction) · Also known as: slot doldurma, Slot Doldurma (Slot Filling / NER-NLU), information slot extraction, dialogue slot filling
Slot filling is a natural-language-understanding task that extracts predefined template fields — such as date, location, or product name — from a user utterance. It emerged as a core component of dialogue systems and form-based information extraction, and became widely studied after Goo et al. (2018) introduced the Slot-Gated Model for joint slot filling and intent prediction, followed by Chen et al. (2019) who extended the paradigm with BERT-based joint modelling.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Slot filling applies whenever user utterances must be parsed into a fixed set of structured fields for a downstream process — dialogue systems, voice assistants, chatbot form completion, and template-based information extraction from documents. The slot schema must be defined in advance, and training data must be annotated in BIO or BIOES format. A minimum of roughly 30 annotated examples per slot is a practical floor, though hundreds of examples per slot are typical in production systems. If the set of required fields is not known in advance, open information extraction is more appropriate.
Strengths & limitations
- Produces structured, machine-readable output directly from free-form text, enabling downstream form filling, database population, and dialogue state tracking.
- Joint intent-plus-slot models share a single encoder, so intent context improves slot boundary detection and vice versa.
- Transformer-based models (BERT, RoBERTa) transfer well across domains, reducing the annotation burden when a related pretrained model exists.
- The slot schema must be fixed before training; changing or adding slots requires re-annotation and retraining.
- Performance degrades for rare slot types with few training examples; very low-frequency slots may not be reliably extracted.
- Overlapping or nested slot spans (e.g., a city name that is also part of a larger location phrase) are not handled by standard BIO labelling and require extended annotation schemes.
Frequently asked
What is the difference between slot filling and named-entity recognition?
Named-entity recognition (NER) identifies and classifies generic entity types — person, organisation, location — in open text. Slot filling identifies entities that correspond to predefined application-specific fields in a template, such as DEPARTURE_CITY or BOOKING_DATE. NER is domain-agnostic; slot filling is schema-driven and tied to a specific task.
Do I need to train a separate model for intent detection and slot filling?
Not necessarily. Joint NLU models such as JointBERT and the Slot-Gated Model predict intent and slot labels from a single shared encoder. Joint training generally outperforms two separate models because intent context helps resolve ambiguous slot boundaries, and the models share representations efficiently.
How much annotated data do I need?
A practical minimum is around 30 annotated utterances per slot, but production systems typically use hundreds to thousands of examples per slot for reliable extraction. Low-resource slots can benefit from data augmentation or transfer learning from related domains.
What if a user utterance contains no relevant slots?
Tokens not belonging to any slot receive the O (Outside) tag. The model will simply output all-O for utterances with no slot values, which is the correct behaviour — no slot values are extracted for that turn.
Sources
- Goo, C.W., Gao, G., Hsu, Y.K., Huo, C.L., Chen, T.C., Hsu, S.C., & Chen, Y.N. (2018). Slot-Gated Modeling for Joint Slot Filling and Intent Prediction. Proceedings of NAACL-HLT 2018. link ↗
- Chen, Q., Zhuo, Z., & Wang, W. (2019). BERT for Joint Intent Classification and Slot Filling. arXiv preprint arXiv:1902.10909. link ↗
How to cite this page
ScholarGate. (2026, June 1). Slot Filling (NER-NLU Joint Extraction). ScholarGate. https://scholargate.app/en/text-mining/slot-filling
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Entity LinkingText mining↔ compare
- Information ExtractionText mining↔ compare
- Intent DetectionText mining↔ compare
- Named Entity RecognitionText mining↔ compare
- Text ClassificationText mining↔ compare