LiLT<T>

Models & Types · AiDotNet.Document.LayoutAware

LiLT (Language-Independent Layout Transformer) for document understanding.

Inheritance: NeuralNetworkBase<T> → DocumentNeuralNetworkBase<T> → LiLT<T>

Implements: IDocumentQA<T>, ILayoutDetector<T>

For Beginners

LiLT is designed for maximum flexibility:

  1. Layout understanding is learned separately from text
  2. Can plug in ANY language model (BERT, RoBERTa, XLM-R, etc.)
  3. Supports any language without retraining the layout part

Key features:

  • BiACM (Bi-directional Attention Complementation Mechanism)
  • Separate text and layout streams
  • Works with any pre-trained text encoder
  • Language-agnostic layout understanding

Example usage:

How It Works

LiLT separates the text and layout modalities during pre-training, enabling the layout model to be combined with ANY pre-trained text model at fine-tuning time, providing true language independence.

Reference: “LiLT: A Simple yet Effective Language-Independent Layout Transformer” (ACL 2022) https://arxiv.org/abs/2202.13669

Constructors

Constructor Summary
LiLT(NeuralNetworkArchitecture<T> architecture, LiLTOptions options, ITokenizer tokenizer, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a LiLT model using native layers for training and inference.
LiLT(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, ITokenizer tokenizer, LiLTOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Creates a LiLT model using a pre-trained ONNX model for inference.

Properties

Property Summary
DocumentType SupportedDocumentTypes { get; }
IReadOnlyList<LayoutElementType> SupportedElementTypes { get; }
bool RequiresOCR { get; }
int ExpectedImageSize { get; }
string TextBackbone { get; } Gets the text backbone model name.

Methods

Method Summary
Dictionary<string, DocumentQAResult<T>> ExtractFields(Tensor<T> documentImage, IEnumerable<string> fieldPrompts)
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) Sends an input through the BiACM dual stream, unpacking bounding boxes when they are present.
DocumentLayoutResult<T> DetectLayout(Tensor<T> documentImage)
DocumentLayoutResult<T> DetectLayout(Tensor<T> documentImage, double confidenceThreshold)
DocumentQAResult<T> AnswerQuestion(Tensor<T> documentImage, string question)
DocumentQAResult<T> AnswerQuestion(Tensor<T> documentImage, string question, int maxAnswerLength, double temperature)
IEnumerable<DocumentQAResult<T>> AnswerQuestions(Tensor<T> documentImage, IEnumerable<string> questions)
ModelMetadata<T> GetModelMetadata()
Tensor<T> EncodeDocument(Tensor<T> documentImage)
Tensor<T> EncodeDualStream(Tensor<T> textTokens, Tensor<T> layoutBoxes) Full dual-stream fusion entry (industry-standard LiLT): encodes text token IDs AND their layout bounding boxes through the coupled BiACM transformer.
Tensor<T> ForwardForTraining(Tensor<T> input)
string GetModelSummary()
void ValidateInputShape(Tensor<T> documentImage)