CNNBiLSTMCRFOptions

Options & Configuration · AiDotNet.NER.Options

Configuration options for the CNN-BiLSTM-CRF Named Entity Recognition model.

Inheritance: ModelOptions → NeuralNetworkOptions → CNNBiLSTMCRFOptions

For Beginners

CNN-BiLSTM-CRF is like BiLSTM-CRF but with an extra component that looks at the letters within each word. A CNN (Convolutional Neural Network) slides a small window across the characters of each word to detect patterns like capitalization (“John” vs “john”), suffixes (“-tion”, “-ing”), and word shapes. These character features are combined with word embeddings before being fed into the BiLSTM.

Compared to BiLSTM-CRF with character BiLSTM:

  • CNN is faster to compute (parallel vs sequential processing)
  • CNN is better at capturing local character n-grams (suffixes, prefixes)
  • BiLSTM is better at capturing long-range character dependencies (rare)
  • In practice, both achieve similar NER accuracy (~91% F1 on CoNLL-2003)

How It Works

CNN-BiLSTM-CRF (Ma and Hovy, ACL 2016 - “End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF”) extends the BiLSTM-CRF architecture with character-level Convolutional Neural Network (CNN) embeddings. The architecture has three main components:

  1. Character-level CNN: A 1D CNN processes each word’s character sequence to capture

morphological features (capitalization, prefixes, suffixes, word shape). Unlike the character BiLSTM in Lample et al. (2016), a CNN is faster and better at capturing local n-gram patterns. The CNN output is max-pooled to produce a fixed-size vector per word.

  1. BiLSTM encoder: Processes the concatenation of word embeddings and character CNN

features in both forward and backward directions. Each token gets a representation that captures its full sentence context plus its morphological features.

  1. CRF decoder: Models label transition dependencies to produce globally optimal

label sequences, enforcing valid BIO constraints via the Viterbi algorithm.

Default values follow the original Ma and Hovy (2016) paper:

  • 100-dimensional GloVe word embeddings
  • Character CNN with 30-dimensional embeddings, 30 filters of width 3
  • Single BiLSTM layer with 200 hidden units per direction
  • 50% dropout rate
  • 9 CoNLL-2003 BIO labels

Constructors

Constructor Summary
CNNBiLSTMCRFOptions() Initializes a new instance with default values matching the original Ma and Hovy (2016) paper.
CNNBiLSTMCRFOptions(CNNBiLSTMCRFOptions other) Initializes a new instance by deep-copying all settings from another instance.

Properties

Property Summary
NERModelVariant Variant { get; set; } Gets or sets the model size variant, which controls the overall capacity of the model.
OnnxModelOptions OnnxOptions { get; set; } Gets or sets the ONNX Runtime configuration options.
bool UseCRF { get; set; } Gets or sets whether to use CRF (Conditional Random Field) decoding.
double DropoutRate { get; set; } Gets or sets the dropout rate for regularization during training.
double LearningRate { get; set; } Gets or sets the learning rate for the optimizer during training.
int CharCNNFilters { get; set; } Gets or sets the number of CNN filters for character-level feature extraction.
int CharCNNKernelSize { get; set; } Gets or sets the kernel (window) size for the character CNN.
int CharEmbeddingDimension { get; set; } Gets or sets the character embedding vector dimension.
int EmbeddingDimension { get; set; } Gets or sets the input token embedding dimension.
int HiddenDimension { get; set; } Gets or sets the LSTM hidden state dimension per direction.
int MaxSequenceLength { get; set; } Gets or sets the maximum input sequence length in tokens.
int NumLSTMLayers { get; set; } Gets or sets the number of stacked BiLSTM layers.
int NumLabels { get; set; } Gets or sets the number of entity label classes in the BIO tagging scheme.
string ModelPath { get; set; } Gets or sets the path to a pre-trained ONNX model file for inference mode.
string[] LabelNames { get; set; } Gets or sets the BIO label names for the tagging scheme.