Named entity recognition with convolutional networks

1. A method comprising:

receiving a document image and a recognized text of the document image;

creating a feature map of the document image comprising one or more features of the recognized text;

creating a tagged map including one or more tagged cells;

estimating, with a machine learning model, one or more locations of a tag in the tagged map; and

recognizing one or more named entities within the recognized text based on the one or more locations of the tag.

2. The method of claim 1 , wherein creating the feature map includes:

creating the feature map comprising a grid with a plurality of cells, wherein a subset of the cells correspond to portions of the document image; and

determining, for the subset of the cells of the feature map, one or more features of the recognized text contained in the corresponding portion of the document image.

3. The method of claim 1 , wherein the features include a recognized word of the recognized text that intersects a section of the document image, a text feature, and a spatial feature.

4. The method of claim 3 wherein the text feature includes a feature selected from the group consisting of syntactical properties of the recognized word and orthographic properties of the recognized word, and

wherein the spatial feature includes a feature selected from the group consisting of a location of a recognized word and a boundary of the recognized word.

5. The method of claim 1 , wherein estimating the location of one or more predicted tags in the feature map further comprises:

estimating a probability that one or more of the tagged cells correspond to a portion of the document that contains the tag.

6. The method of claim 1 , wherein the recognized text includes a tabular text formatted in a table within the document image, and wherein at least one of the one or more named entities is recognized in the tabular text.

7. The method of claim 1 , wherein the recognized text includes a nested texted formatted to be nested beneath another portion of the recognized text, and wherein at least one of the one or more named entities is recognized in the nested text.

8. The method of claim 1 , further comprising:

receiving (i) a training document image, (ii) a training text recognized from the training document image, (iii) a training feature map associated with the training image document, (iv) a labeled tagged map indicating one or more locations of the tag within the training feature map; and (v) a labeled training text indicating one or more named entities located within the training text;

estimating, with the machine learning model, one or more locations of a tag within a training tagged map;

recognizing one or more named entities within the training text based on the training tagged map;

comparing the one or more locations of the tag within the training tagged map with the one or more locations of the tag within the labeled tagged map to identify one or more tag errors in the locations of the tag within the training tagged map;

comparing the one or more named entities recognized within the training text with the named entities indicated in the labeled training text to identify one or more named entity errors in the named entities recognized within the training text; and

updating one or more parameters of the machine learning model based on one or both of the tag errors and the named entity errors.

9. The method of claim 8 , wherein the machine learning model is initially, prior to updating, configured to estimate tag locations in documents of a first document type and the machine learning model estimates tag locations in documents of a second document type using updates of the one or more parameters of the machine learning mode.

10. The method of claim 8 , wherein the machine learning model estimates locations of a new tag using updates of the one or more parameters of the machine learning model.

11. A system comprising:

a processor; and

a document intake system comprising a feature map creator, a tag predictor, and a named entity recognizer,

wherein the document intake system is configured, when executed by the processor, to:

receive a document image and a recognized text of the document image;

create, with the feature map creator, a feature map of the document image comprising one or more features of the recognized text;

create, with a tag predictor, a tagged map including one or more tagged cells;

estimate, with a machine learning model, one or more locations of a tag in the tagged map; and

recognize, with a named entity recognizer, one or more named entities within the recognized text based on the one or more locations of the tag.

12. The system of claim 11 , wherein the feature map creator is configured, when executed by the processor, to:

create the feature map comprising a grid with a plurality of cells, wherein a subset of the cells correspond to portions of the document image; and

determine, for the subset of the cells of the feature map, one or more features of the recognized text contained in the corresponding portion of the document image.

13. The system of claim 11 , wherein the features include a recognized word of the recognized text that intersects a section of the document image, a text feature, and a spatial feature,

wherein the text feature includes a feature selected from the group consisting of syntactical properties of the recognized word and orthographic properties of the recognized word, and

wherein the spatial feature includes a feature selected from the group consisting of a location of the recognized word and a boundary of the recognized word.

14. The system of claim 11 , wherein the tag predictor is configured, when executed by a processor, to:

estimate a probability that one or more of the tagged cells correspond to a portion of the document that contains the tag.

15. The system of claim 11 , wherein the recognized text includes a tabular text formatted in a table within the document image, and wherein at least one of the one or more named entities is recognized in the tabular text.

16. The system of claim 11 , wherein the recognized text includes a nested texted formatted to be nested beneath another portion of the recognized text, and wherein at least one of the one or more named entities is recognized in the nested text.

17. The system of claim 11 , further comprising a training system configured, when executed by the processor, to:

receive (i) a training document image, (ii) a training text recognized from the training document image, (iii) a training feature map associated with the training image document, (iv) a labeled tagged map indicating one or more locations of the tag within the training feature map; and (v) a labeled training text indicating one or more named entities located within the training text;

estimate, with the machine learning model, one or more locations of a tag within a training tagged map;

recognize one or more named entities within the training text based on the training tagged map;

compare the one or more locations of the tag within the training tagged map with the one or more locations of the tag within the labeled tagged map to identify one or more tag errors in the locations of the tag within the training tagged map;

compare the one or more named entities recognized within the training text with the named entities indicated in the labeled training text to identify one or more named entity errors in the named entities recognized within the training text; and

update one or more parameters of the machine learning model based on one or both of the tag errors and the named entity errors.

18. The system of claim 17 , wherein the machine learning model is initially configured to estimate tag locations in documents of a first document type and updating one or more parameters of the machine learning model configures the machine learning model to better estimate tag locations in documents of a second document type.

19. The system of claim 17 , wherein updating one or more parameters of the machine learning model configures the machine learning model to better estimate locations of a new tag.

20. A non-transitory computer-readable medium storing instructions which, when executed by one or more processors, cause the one or more processors to:

receive a document image and a recognized text of the document image;

create a feature map of the document image comprising one or more features of the recognized text;

create a tagged map including one or more tagged cells;

estimate one or more locations of a tag in the tagged map; and

recognize one or more named entities within the recognized text based on the one or more locations of the tag.