Entity extraction with encoder decoder machine learning model

1. A method comprising:

executing an optical character recognition (OCR) engine on a document image of a document to obtain OCR output comprising a plurality of token values of the document image;

inputting the plurality of token values from the document image into an encoder machine learning model;

creating an encoder hidden state vector, the creating comprising executing the encoder machine learning model on the plurality of token values contained in the document;

generating, by a decoder machine learning model executing on the encoder hidden state vector, raw text comprising an entity value and an entity label for each entity of a plurality of entities, wherein the plurality of entities is in the document;

suppressing the plurality of entities in the raw text that are determined to have a less than threshold probability of being in the OCR output;

generating a structural representation of the plurality of entities directly from the raw text; and

outputting the structural representation of the plurality of entities.

2. The method of claim 1 , further comprising:

inputting formatting information generated from executing the OCR engine into the encoder machine learning model.

3. The method of claim 1 , further comprising:

inputting the document image into the encoder machine learning model, wherein the encoder machine learning model processes image features obtained from the document image.

4. The method of claim 1 , wherein the encoder machine learning model is executed directly on the OCR output without additional processing, the OCR output comprising the plurality of token values.

5. The method of claim 1 , further comprising:

comparing the plurality of entities with ground truth information to obtain a comparison result;

generating, by a loss function, a loss using the comparison result; and

backpropagating the loss through the decoder machine learning model and the encoder machine learning model.

6. The method of claim 1 , wherein executing the encoder machine learning model comprises:

executing a plurality of encoder blocks on the plurality of token values.

7. A method comprising:

creating an encoder hidden state vector, the creating comprising executing an encoder machine learning model on a plurality of token values contained in a document;

generating, by a decoder machine learning model executing on the encoder hidden state vector, raw text comprising an entity value and an entity label for each entity of a plurality of entities, wherein the plurality of entities is in the document;

generating a structural representation of the plurality of entities directly from the raw text;

executing a document classification model on the document to obtain a document type;

identifying a set of expected entity labels corresponding to the document type;

filtering, from the structural representation, an additional entity in the raw text generated by the decoder machine learning model based on the additional entity having the entity label failing to be in the set of expected entity labels; and

outputting the structural representation of the plurality of entities.

8. A system comprising:

an encoder machine learning model, executing on at least one computer processor, configured to process a plurality of token values extracted from a document to create an encoder hidden state vector;

a decoder machine learning model, executing on the at least one computer processor, configured to:

generate, by processing the encoder hidden state vector, raw text comprising an entity value and an entity label for each entity of a plurality of entities, wherein the plurality of entities is in the document,

a document classification model configured to classify the document to determine a document type of the document; and

a filter configured to:

identify a set of expected entity labels corresponding to the document type, and

filter, from a structural representation generated from the raw text, an additional entity in the raw text generated by the decoder machine learning model based on the additional entity having the entity label failing to be in the set of expected entity labels,

wherein the system is configured to output the plurality of entities of the document.

9. The system of claim 8 , further comprising:

an optical character recognition (OCR) engine processing a document image to generate a sequence of characters with bounding boxes representing a plurality of words of the document image,

wherein the plurality of words from the document image are used as the plurality of token values that are input into the encoder machine learning model.

10. The system of claim 9 , wherein the OCR engine is further configured to generate layout information that is input into the encoder machine learning model.

11. The system of claim 9 , wherein the encoder machine learning model receives, as input, the document image and is configured to generate the encoder hidden state vector using the document image.

12. The system of claim 8 , wherein the encoder machine learning model is executed directly on optical character recognition (OCR) output without additional processing, the OCR output comprising the plurality of token values.

13. The system of claim 8 , further comprising:

a loss function executing on the at least one computer processor and configured to:

compare the plurality of entities with ground truth information to obtain a comparison result, and

generate a loss using the comparison result,

wherein the loss is back-propagated through the decoder machine learning model and the encoder machine learning model.

14. The system of claim 8 , wherein the encoder machine learning model comprises:

a plurality of encoder blocks to execute on the plurality of token values.

15. A method comprising:

creating an encoder hidden state vector, the creating comprising executing an encoder machine learning model on a document image of a document;

generating, by a decoder machine learning model executing on the encoder hidden state vector, raw text comprising an entity value and an entity label for each entity of a plurality of entities, wherein the plurality of entities is in the document;

generating a structural representation of the plurality of entities directly from the raw text;

executing a document classification model on the document to obtain a document type;

identifying a set of expected entity labels corresponding to the document type;

filtering, from the structural representation, an additional entity in the raw text generated by the decoder machine learning model based on the additional entity having the entity label failing to be in the set of expected entity labels; and

outputting the plurality of entities of the document.