System and method for parsing visual information to extract data elements from randomly formatted digital documents
An information handling system is provided for extracting data associated with a data field of a database from a randomly formatted source document. A processor of the information handling system receives the document, embeds a text-based representation of text of the document into vector data associated with the text, and encodes the vector data through a first neural network into reconstructed text activations. The processor further concatenates the vector data with an image-based representation of the document to provide first concatenated data, and encodes the first concatenated data through a second neural network into extracted visual feature activations. The processor further concatenates the reconstructed text activations with the extracted visual feature activations to provide second concatenated data, decodes the second concatenated data to identify particular data associated with the data field, and updates an entry of the database with the particular data in the data field.
1. An information handling system for extracting data associated with a data field of a database from a randomly formatted source document, the information handling system comprising:
a memory; and
a processor in communication with the memory and configured to:
receive the document;
embed a text-based representation of text of the document into vector data associated with the text;
in parallel operations:
encode the vector data through a first neural network into reconstructed text activations; and
concatenate the vector data with an image-based representation of the document to provide first concatenated data, and encode the first concatenated data through a second neural network into extracted visual feature activations, wherein the second neural network comprises a dilated convolution neural network;
concatenate the reconstructed text activations with the extracted visual feature activations to provide second concatenated data;
decode the second concatenated data to identify particular data associated with the data field; and
update an entry of the database with the particular data in the data field.
2. The information handling system of claim 1 , wherein:
the document is received in a first file format; and
the processor is further configured to convert the document from the first file format to a second file format, wherein the first file format is different from the second file format, and wherein the second file format includes the text-based representation.
3. The information handling system of claim 2 , wherein the processor is further configured to convert the document from the first file format to a third file format, wherein the first and second file formats are different from the third file format, and wherein the third file format includes the image-based representation.
4. The information handling system of claim 1 , wherein in embedding the text-based representation, the processor utilizes a word-to-vector embedding model.
5. The information handling system of claim 4 , wherein in embedding the text-based representation, the processor further utilizes a character-to-vector embedding model.
6. The information handling system of claim 1 , wherein in decoding the second concatenated data, the processor utilizes two dense layers of rectified linear units (ReLUs), followed by a single layer of bi-directional Long Short-Term Memory (LSTM) units, and followed by two additional dense layers of ReLUs.
7. The information handling system of claim 6 , wherein each dense layer of ReLUs includes at least five hundred layers.
8. The information handling system of claim 7 , wherein the dense layer of LSTM units includes at least three hundred layers.
9. The information handling system of claim 1 , wherein:
the information handling system represents an invoice processing system; and
the particular data includes at least one of invoicing entity, an invoicing entity address, an invoice number, a line item number, a line item description, and a price.
10. A method for extracting data associated with a data field of a database from a randomly formatted source document, method comprising:
receiving, by a hardware processor, the document;
embedding a text-based representation of text of the document into vector data associated with the text;
in parallel operations:
encoding the vector data through a first neural network into reconstructed text activations; and
concatenating the vector data with an image-based representation of the document to provide first concatenated data, and encoding the first concatenated data through a second neural network into extracted visual feature activations, wherein the second neural network comprises a dilated convolution neural network;
concatenating the reconstructed text activations with the extracted visual feature activations to provide second concatenated data;
decoding the second concatenated data to identify particular data associated with the data field; and
updating an entry of the database with the particular data in the data field.
11. The method of claim 10 , wherein the document is received in a first file format, the method further comprising:
converting the document from the first file format to a second file format, wherein the first file format is different from the second file format, and wherein the second file format includes the text-based representation.
12. The method of claim 11 , further comprising:
converting the document from the first file format to a third file format, wherein the first and second file formats are different from the third file format, and wherein the third file format includes the image-based representation.
13. The method of claim 10 , wherein in embedding the text-based representation, the processor utilizes a word-to-vector embedding model.
14. The method of claim 13 , wherein in embedding the text-based representation, the processor further utilizes a character-to-vector embedding model.
15. The method of claim 10 , wherein in decoding the second concatenated data, the processor utilizes two dense layers of rectified linear units (ReLUs), followed by a single layer of bi-directional Long Short-Term Memory (LSTM) units, and followed by two additional dense layers of ReLUs.
16. The method of claim 15 , wherein each dense layer of ReLUs includes at least five hundred layers, and the dense layer of LSTM units includes at least three hundred layers.
17. The method of claim 1 , wherein the particular data includes at least one of invoicing entity, an invoicing entity address, an invoice number, a line item number, a line item description, and a price.
18. An invoice processing system for extracting data associated with a data field of a database from a randomly formatted source document, the system comprising:
a memory; and
a processor in communication with the memory and configured to:
receive the document;
embed a text-based representation of text of the document into vector data associated with the text;
in parallel operations:
encode the vector data into reconstructed text activations; and
concatenate the vector data with an image-based representation of the document to provide first concatenated data and encode the first concatenated data through a dilated convolution neural network into extracted visual feature activations;
concatenate the reconstructed text activations with the extracted visual feature activations to provide second concatenated data;
decode the second concatenated data to identify particular data associated with the data field, wherein the data field; and
update an entry of the database with the particular data in the particular data includes at least one of invoicing entity, an invoicing entity address, an invoice number, a line item number, a line item description, and a price.