BiaffineNEROptions
Options & Configuration · AiDotNet.NER.Options
Options for Biaffine-NER, defaulting to the values published in Yu et al., ACL 2020, “Named Entity Recognition as Dependency Parsing” (arXiv:2005.07150).
Inheritance: SpanBasedNEROptions → BiaffineNEROptions
For Beginners
These are the settings the paper’s authors actually used to get their published results. You can change any of them, but the defaults are the paper.
How It Works
SpanBasedNEROptions is shared by Biaffine-NER, SpERT and PURE, which are
three different papers with genuinely conflicting hyperparameters — SpERT samples negative
spans, whereas Biaffine-NER classifies every span with a dedicated non-entity class. A single
set of shared defaults therefore cannot be faithful to all three, so each model carries its
own subclass whose defaults are literally its own paper’s.
Values taken from the paper’s hyperparameter table: BiLSTM size 200 with 3 layers and dropout 0.4; FFNN size 150 with dropout 0.2; embeddings dropout 0.5; Adam at learning rate 1e-3; a BERT-Large encoder using the last 4 layers (1024-dimensional).
Constructors
| Constructor | Summary |
|---|---|
BiaffineNEROptions() |
Initializes a new instance with the paper’s published defaults. |
BiaffineNEROptions(BiaffineNEROptions other) |
Initializes a new instance by copying another Biaffine-NER options instance. |
BiaffineNEROptions(SpanBasedNEROptions other) |
Initializes a new instance by copying another Biaffine-NER options instance. |
Properties
| Property | Summary |
|---|---|
double BiLstmDropout { get; set; } |
Gets or sets the dropout applied inside the BiLSTM stack. |
double EmbeddingsDropout { get; set; } |
Gets or sets the dropout applied to the input embeddings. |
int BiLstmHiddenSize { get; set; } |
Gets or sets the per-direction hidden size of the BiLSTM stacked on the encoder. |
int BiLstmLayers { get; set; } |
Gets or sets the number of stacked BiLSTM layers. |