Multimodal multitask machine learning system for document intelligence tasks
Multimodal multitask machine learning system for document intelligence tasks includes a feature extractor processing token values obtained from a document to obtain features, and a token extraction head classifying, using the features, the token values to obtain classified tokens. The classified tokens are aggregated into entities. A document classification model is executed on the features to classify the document and obtain a document label prediction. Further a confidence head model applying the document label prediction processes the entities to obtain a result.
1 . A method comprising:
processing, by a feature extractor of an arrangement of machine-learning models, a plurality of token values obtained from a document, to obtain a plurality of features;
classifying, using the plurality of features, the plurality of token values by a token extraction head of the arrangement of machine-learning models, to obtain a plurality of classified tokens;
aggregating the plurality of classified tokens into a plurality of entities;
generating a document label prediction based on executing a document classification model on the plurality of features;
processing, by a confidence head of the arrangement of machine-learning models, based on applying the document label prediction, the plurality of entities to obtain a task consistent confidence loss wherein the task consistent confidence loss is a function of parameters comprising a binary cross entropy loss, a document label probability and an indicator function;
jointly optimizing, during training, the arrangement of machine-learning models comprising the feature extractor, the confidence head, and the token extraction head by processing the task consistent confidence loss through the feature extractor, the confidence head, and the token extraction head; and
processing a plurality of documents by the trained arrangement of machine-learning models to extract content comprising entity identifier, entity value pairs from the plurality of documents.
2 . The method of claim 1 , further comprising:
calculating a cross entropy loss by comparing a training document label with the document label prediction; and
backpropagating the cross entropy loss through the document classification model.
3 . The method of claim 1 , further comprising:
processing a document image through an optical character recognition (OCR) engine to obtain the plurality of token values, wherein the document comprises the document image.
4 . The method of claim 1 , further comprising:
obtaining the plurality of token values and a layout of the document; and
processing, by the feature extractor, the layout of the document, and the plurality of token values to obtain a plurality of multimodal features,
wherein the plurality of features comprises the plurality of multimodal features.
5 . The method of claim 1 , further comprising:
processing a document image through an image embedding model to obtain an image feature vector for the document image, wherein the document is the document image; and
processing, by the feature extractor, the plurality of token values, a layout of the document image, and the image feature vector to obtain a plurality of multimodal features,
wherein the plurality of features comprises the plurality of multimodal features.
6 . A system comprising:
an arrangement of machine-learning models, comprising:
a feature extractor executing on a computer processor for processing a plurality of token values obtained from a document to obtain a plurality of features,
a token extraction head executing on the computer processor for classifying, using the plurality of features, the plurality of token values to obtain a plurality of classified tokens,
a token aggregator executing on the computer processor for aggregating the plurality of classified tokens into a plurality of entities,
a document classification model executing on the computer processor for classifying, using the plurality of features, the document to obtain a document label prediction,
a confidence head model executing on the computer processor for processing, by applying the document label prediction, the plurality of entities to obtain a task consistent confidence loss, and
a task consistent loss function executing on the computer processor to calculate the task consistent confidence loss using parameters comprising a binary cross entropy loss, a document label probability, and an indicator function;
wherein the system is configured for:
jointly optimizing, during training, the arrangement of machine-learning models by processing the task consistent confidence loss through the feature extractor, the confidence head model, and the token extraction head; and
processing a plurality of documents by the trained arrangement of machine-learning models to extract content comprising entity identifier, entity value pairs from the plurality of documents.
7 . The system of claim 6 , further comprising:
a document classification loss function configured to
calculate a cross entropy loss by comparing a training document label with the document label prediction,
wherein the cross entropy loss is backpropagated through the document classification model.
8 . The system of claim 6 , further comprising:
an optical character recognition (OCR) engine for obtaining the plurality of token values from a document image, the document comprising the document image.
9 . The system of claim 6 , further comprising:
an optical character recognition (OCR) engine for obtaining the plurality of token values and a layout from a document image, the document comprising the document image,
wherein the feature extractor processes a layout of the document image, and the plurality of token values to obtain a plurality of multimodal features,
wherein the plurality of features comprises the plurality of multimodal features.
10 . The system of claim 6 , further comprising:
an image embedding model for processing a document image to obtain an image feature vector for the document image, wherein the document comprises the document image,
wherein the feature extractor processes the image feature vector, a layout of the document image, and the plurality of token values to obtain a plurality of multimodal features,
wherein the plurality of features comprises the plurality of multimodal features.