AI-based detection of contextual class description in document images

1. A non-transitory computer-readable medium having executable instructions stored thereon that, when executed by a processor, cause the processor to perform operations comprising:

obtaining a document image file including a first image corresponding to a first page of a document;

generating, using a first trained model, based on the first image, a first prediction that the first page includes a term identifying a class of people, the first prediction including a first location of the term within the first image;

in response to generating the first prediction that the first page includes the term identifying the class of people, generating, using a second trained model, based on the first image, a second prediction of whether or not the first page includes a section that uses a term identifying the class of people in a specific context.

2. The non-transitory computer-readable medium of claim 1 , wherein:

the second prediction is that the first page includes the section that uses the term identifying the class of people in the specific context; and

the second prediction comprises a second location of the section.

3. The non-transitory computer-readable medium of claim 2 , wherein the operations further comprise: in response to the second prediction being that the first page includes the section that uses the term identifying the class of people in the specific context, redacting, based on the second location of the section, the section within the first image.

4. The non-transitory computer-readable medium of claim 2 , wherein the operations further comprise: in response to the second prediction being that the first page includes the section that uses the term identifying the class of people in the specific context, storing the second location and a page number of the first page in a datastore.

5. The non-transitory computer-readable medium of claim 4 , wherein

the first prediction further comprises a first confidence score;

the second prediction further comprises a second confidence score; and

the operations further comprise: storing the first location, the first confidence score, and the second confidence score in the datastore.

6. The non-transitory computer-readable medium of claim 2 , wherein the operations further comprise: redacting, based on the second location of the section, the section within the first image.

7. The non-transitory computer-readable medium of claim 2 , wherein the specific context is a discriminatory context.

8. The non-transitory computer-readable medium of claim 7 , wherein:

the document refers to a real property;

the class of people corresponding to the term used in the section comprises a race, color, religion, national origin, sex, or sexual orientation; and

the section includes text restricting, based on race, color, religion, national origin, sex, or sexual orientation, the sale, use, lease, rent, or occupancy of the real property.

9. The non-transitory computer-readable medium of claim 2 , wherein:

the first prediction further includes a label corresponding to the term identifying the class of people; and

the operations further comprise:

determining that the first location is within the second location; and

in response to determining that the first location is within the second location, associating the label with the section.

10. The non-transitory computer-readable medium of claim 2 , wherein the operations further comprise:

determining that the first location is not within the second location; and

in response to determining that the first location is not within the second location, presenting a graphical user interface for a user to review the section.

11. The non-transitory computer-readable medium of claim 1 , wherein:

the document image file includes a second image of a second page of the document; and

the operations further comprise:

generating, using the first trained model, based on the second image, a third prediction that the second page does not include any terms identifying classes of people; and

in response to generating the third prediction, not using the second trained model to generate any predictions about the second page.

12. The non-transitory computer-readable medium of claim 1 , wherein the operations further comprise generating the first trained model by:

obtaining multiple textual data representations of multiple document image files;

identifying, using at least regular expression-based rules, and the multiple textual data representations, multiple text strings within the multiple textual data representations that identify classes of people;

determining a location of each of the multiple text strings within the multiple document image files; and

training, based on the multiple document image files and the locations of the multiple text strings, a model as the second trained model.

13. The non-transitory computer-readable medium of claim 12 , wherein generating the first trained model further includes:

transforming the multiple document image files to obtain multiple transformed versions of the document image files;

determining a location of each of the multiple text strings within the multiple transformed versions of the document image files; and

training, based on the multiple transformed versions of the document image files and the locations of the multiple text strings within the transformed versions of the document image files, the model as the second trained model.

14. A method, comprising:

obtaining, at a computing device, a document image file including a first image corresponding to a first page of a document;

generating, at the computing device, using a first trained model, based on the first image, a first prediction that the first page includes a term identifying a class of people, the first prediction including a first location of the term within the first image;

in response to generating the first prediction that the first page includes the term identifying the class of people, generating, at the computing device, using a second trained model, based on the first image, a second prediction of whether or not the first page includes a section that uses a term identifying the class of people in a specific context.

15. The method of claim 14 , wherein:

the second prediction is that the first page includes the section that uses the term identifying the class of people in the specific context; and

the second prediction comprises a second location of the section.

16. The method of claim 15 , further comprising: in response to the second prediction being that the first page includes the section that uses the term identifying the class of people in the specific context, redacting, at the computing device, based on the second location of the section, the section within the first image.

17. The method of claim 15 , further comprising: in response to the second prediction being that the first page includes the section that uses the term identifying the class of people in the specific context, storing the second location and a page number of the first page in a datastore.

18. The method of claim 15 , further comprising: redacting, based on the second location of the section, the section within the first image.

19. The method of claim 15 , wherein the specific context is a discriminatory context.

20. A system, comprising:

a processor; and

a non-transitory computer-readable medium having executable instructions stored thereon that, when executed by the processor, cause the processor to perform operations comprising:

obtaining a document image file including a first image corresponding to a first page of a document;

generating, using a first trained model, based on the first image, a first prediction that the first page includes a term identifying a class of people, the first prediction including a first location of the term within the first image; and

in response to generating the first prediction that the first page includes the term identifying the class of people, generating, using a second trained model, based on the first image, a second prediction of whether or not the first page includes a section that uses a term identifying the class of people in a specific context.