Resolving latent status from dense information using machine learning
Aspects of the present disclosure are directed to resolving a latent status of document(s) that contain dense information using machine learning. Documents can be organized as a set of informational pages that together convey data associated with an organizational concept. Resolving a latent status from the pages of documents comprising dense information can be challenging, for example when indicator(s) of this latent status are hidden within the dense information. Implementations predict the latent status of document(s) via machine learning by performing a feature extraction processing pipeline. For example, a feature extraction processing pipeline can reduce the dimensionality of the document data, thus providing the machine learning model(s) a simplified problem. The machine learning model(s) can process the extracted data and, given the simplified problem, accurately predict the latent status of the document(s). In some implementations, the machine learning model(s) are trained using real-world training instances and synthetic training instances.
1. A method for resolving a latent status of one or more documents, that contain dense information, using machine learning, the method comprising:
receiving one or more documents comprising a plurality of line items, wherein the one or more documents are organized into pages;
determining match instances between text from a subset of the line items and text descriptive of an issue category;
extracting, for each of the match instances, one or more distance and/or angle measurements between: a) a location of the matched text of the match instance on a corresponding page of the one or more documents, and b) a location of symbols associated with the matched text on the corresponding page of the one or more documents;
generating status predictions for the issue category by inputting, to one or more trained machine learning models, instances of input that correspond to the match instances, wherein each instance of input comprises at least (1) a portion of the matched text from the match instances, (2) distance and/or angle measurements extracted between the matched text and one or more of the associated symbols, and (3) one or more indicators representative of the associated symbols; and
analyzing the status predictions to generate an overall status prediction for the issue category.
2. The method of claim 1 , wherein the symbols associated with the matched text from the match instances comprise numerical values.
3. The method of claim 2 , wherein the extracted measurements comprise one or more angles with respect to a) a location of the matched text of the match instances on the corresponding page of the one or more documents, and b) a location of one or more of the numerical values associated with the matched text on the corresponding page of the one or more documents.
4. The method of claim 1 , wherein determining the matches is based on a Levenshtein distance between the text from a subset of the line items and the text descriptive of the issue category.
5. The method of claim 1 , wherein the trained one or more machine learning models comprise one or more gradient boosted trees.
6. The method of claim 1 , wherein the trained machine learning models are trained using real-world training instances comprising extracted features from one or more training documents and synthetic training instances.
7. The method of claim 6 , wherein the real-world training instances and synthetic training instances comprise data labels, and the trained machine learning model is trained via supervised learning.
8. The method of claim 6 , wherein the synthetic training instances comprise features that correspond to skewed, warped, or manipulated versions of the real-world training instances.
9. The method of claim 1 , wherein the one or more documents comprise one or more vehicle loan documents and the overall status prediction corresponds to a coverage or liability for the vehicle loan documents.
10. The method of claim 9 , wherein the issue category relates to gap coverage and the overall status prediction is a prediction about whether the vehicle loan documents comprises gap coverage.
11. A computer-readable storage medium storing instructions that, when executed by a computing system, cause the computing system to perform a process for resolving a latent status of one or more documents, that contain dense information, using machine learning, the process comprising:
receiving one or more documents comprising a plurality of line items, wherein the one or more documents are organized into pages;
determining match instances between text from a subset of the line items and text descriptive of an issue category;
extracting, for each of the match instances, one or more distance and/or angle measurements between: a) a location of the matched text of the match instance on a corresponding page of the one or more documents, and b) a location of symbols associated with the matched text on the corresponding page of the one or more documents;
generating status predictions for the issue category by inputting, to one or more trained machine learning models, instances of input that correspond to the match instances, wherein each instance of input comprises at least (1) a portion of the matched text from the match instances, (2) distance and/or angle measurements extracted between the matched text and one or more of the associated symbols, and (3) one or more indicators representative of the associated symbols; and
analyzing the status predictions to generate an overall status prediction for the issue category.
12. The computer-readable storage medium of claim 11 , wherein the symbols associated with the matched text from the match instances comprise numerical values.
13. The computer-readable storage medium of claim 12 , wherein the extracted measurements comprise one or more angles with respect to a) a location of the matched text of the match instances on the corresponding page of the one or more documents, and b) a location of one or more of the numerical values associated with the matched text on the corresponding page of the one or more documents.
14. The computer-readable storage medium of claim 11 , wherein determining the matches is based on a Levenshtein distance between the text from a subset of the line items and the text descriptive of the issue category.
15. The computer-readable storage medium of claim 11 , wherein the trained one or more machine learning models comprise one or more gradient boosted trees.
16. The computer-readable storage medium of claim 11 , wherein the trained machine learning models are trained using real-world training instances comprising extracted features from one or more training documents and synthetic training instances.
17. The computer-readable storage medium of claim 16 , wherein the real-world training instances and synthetic training instances comprise data labels, and the trained machine learning model is trained via supervised learning.
18. The computer-readable storage medium of claim 11 , wherein the one or more documents comprise one or more vehicle loan documents and the overall status prediction corresponds to a coverage or liability for the vehicle loan documents.
19. The computer-readable storage medium of claim 18 , wherein the issue category relates to gap coverage and the overall status prediction is a prediction about whether the vehicle loan documents comprises gap coverage.
20. A computing system for resolving a latent status of one or more documents, that contain dense information, using machine learning, the computing system comprising:
one or more processors; and
one or more memories storing instructions that, when executed by the one or more processors, cause the computing system to perform a process comprising:
receiving one or more documents comprising a plurality of line items, wherein the one or more documents are organized into pages;
determining match instances between text from a subset of the line items and text descriptive of an issue category;
extracting, for each of the match instances, one or more distance and/or angle measurements between: a) a location of the matched text of the match instance on a corresponding page of the one or more documents, and b) a location of symbols associated with the matched text on the corresponding page of the one or more documents;
generating status predictions for the issue category by inputting, to one or more trained machine learning models, instances of input that correspond to the match instances, wherein each instance of input comprises at least (1) a portion of the matched text from the match instances, (2) distance and/or angle measurements extracted between the matched text and one or more of the associated symbols, and (3) one or more indicators representative of the associated symbols; and
analyzing the status predictions to generate an overall status prediction for the issue category.