TemplateNER<T>
Models & Types · AiDotNet.NER.TransformerBased
Template-NER: Template-based prompt approach for few-shot and zero-shot NER.
Inheritance: NeuralNetworkBase<T> → NERNeuralNetworkBase<T> → SequenceLabelingNERBase<T> → TransformerNERBase<T> → TemplateNER<T>
For Beginners
Template-NER treats entity recognition like a fill-in-the-blank exercise. Instead of labeling each word, it asks the model: “What person is mentioned in this sentence?” and the model answers: “Barack Obama.” This approach works even with very few training examples because the model already understands language from pre-training.
How It Works
Template-NER (Cui et al., ACL 2021 - “Template-Based Named Entity Recognition Using BART”) reformulates NER as a sequence-to-sequence generation problem using natural language templates. Instead of predicting labels, the model generates text that fills in template slots.
Key Innovation - Template-Based Generation: Given an input sentence, Template-NER constructs a template that describes the NER task in natural language, then uses a seq2seq model (BART) to generate the filled template.
Example: Input: “Barack Obama visited New York City” Template: “[Person] visited [Location]” Model generates: “Barack Obama visited New York City” Extracted: Person=“Barack Obama”, Location=“New York City”
How Templates Work:
- Define templates for each entity type: “[Person] is a person”, “[Location] is a location”
- For each entity type, prompt the model: “In the sentence X, [Person] is a person”
- The model generates the entity that fills [Person] based on the sentence context
- Repeat for all entity types to extract all entities
Few-Shot Capability: Because the templates are expressed in natural language, the model can leverage its pre-trained knowledge to perform NER with very few labeled examples:
- Zero-shot: No labeled data, just template definitions
- Few-shot (10 examples): ~70-75% F1 on CoNLL-2003
- Few-shot (100 examples): ~80-85% F1 on CoNLL-2003
- Full training: ~93.0% F1 on CoNLL-2003
Template Design Patterns:
- Entity extraction: “In the sentence, [TYPE] refers to ___”
- Type classification: “___ is a [PERSON/LOCATION/ORGANIZATION]”
- Cloze-style: “Obama is a [MASK] entity” -> “person”
Example
var architecture = new NeuralNetworkArchitecture<float>(inputFeatures: 512, numClasses: 9);
var templateNER = new TemplateNER<float>(architecture);
Constructors
| Constructor | Summary |
|---|---|
TemplateNER(NeuralNetworkArchitecture<T> architecture, TransformerNEROptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
Creates a Template-NER model in native training mode. |
TemplateNER(NeuralNetworkArchitecture<T> architecture, string modelPath, TransformerNEROptions options) |
Creates a Template-NER model in native training mode. |