Extracting information from documents using automatic markup based on historical data

1. A method comprising:

receiving, by a processing device, a document;

identifying, in a data structure, a record corresponding to the document, the record comprising one or more entries, each entry containing data reflecting a respective item of information extracted from a corresponding part of the document;

for each entry of the record, determining a first embedding of a first character string contained by the entry;

identifying, among a set of embeddings corresponding to character strings of the document, a second embedding that is closest in a vector space to the first embedding;

identifying, in the document, a second character string encoded by the second embedding;

determining a degree of association between the entry and the second character string;

selecting, among a plurality of degrees of association between entries of the data structure and corresponding character strings, a set of degrees of association whose aggregate degree of association satisfies a criterion; and

training, using the set of degrees of association, a machine learning model to extract information from new documents.

2. The method of claim 1 , further comprising:

updating the degree of association by performing at least one of:

identifying character strings using a word search, detecting fields with a neural network model, or receiving identification of character strings from a user interface.

3. The method of claim 2 , wherein the word search comprises: searching a body of text in the document for a character string from the entry of the record by performing one or more of an exact string search, a prefix search, an approximate string search, or an approximate prefix search.

4. The method of claim 2 , wherein detecting the fields with the neural network model comprises: marking one or more areas in a document as fields, each field associated with a class of information, and matching a character string from the entry in the record with one or more fields based on the class of information associated with each field.

5. The method of claim 1 , further comprising performing at least one of:

eliminating degrees of association that are lower than a predefined threshold value or combining degrees of association resulting from different character string identification methods.

6. The method of claim 1 , wherein selecting the set of degrees of association that satisfies a criterion comprises: applying a bijective function to a set of entries from the record and a set of items of information identified in the document.

7. A system comprising:

a memory device comprising a data structure;

a processor coupled to the memory device, the processor configured to perform operations comprising:

receiving a document;

identifying, in a data structure, a record corresponding to the document, the record comprising one or more entries, each entry containing data reflecting a respective item of information extracted from a corresponding part of the document;

for each entry of the record, determining a first embedding of a first character string contained by the entry;

identifying, among a set of embeddings corresponding to character strings of the document, a second embedding that is closest in a vector space to the first embedding;

identifying, in the document, a second character string encoded by the second embedding;

determining a degree of association between the entry and the second character string;

selecting, among a plurality of degrees of association between entries of the data structure and corresponding character strings, a set of degrees of association whose aggregate degree of association satisfies a criterion; and

training, using the set of degrees of association, a machine learning model to extract information from new documents.

8. The system of claim 7 , wherein the operations further comprise:

updating the degree of association by performing at least one of:

identifying character strings using a word search, detecting fields with a neural network model, or receiving identification of character strings from a user interface.

9. The system of claim 8 , wherein the word search comprises: searching a body of text in the document for a character string from the entry from the record by performing one or more of an exact string search, a prefix search, an approximate string search, or an approximate prefix search.

10. The system of claim 8 , wherein detecting the fields with the neural network model comprises: marking one or more areas in a document as fields, each field associated with a class of information, and matching a character string from the entry in the record with one or more fields based on the class of information associated with each field.

11. The system of claim 7 , wherein the operations further comprise:

performing at least one of: eliminating corresponding degrees of association that are lower than a predefined threshold value or combining degrees of association resulting from different character string identification methods.

12. The system of claim 7 , wherein selecting the set of degrees of association that satisfies a criterion comprises: applying a bijective function to a set of entries from the record and a set of items of information identified in the document.