Classifying data from de-identified content
Methods, systems, and computer program products for classifying data from de-identified content are provided herein. A computer-implemented method includes applying one or more rules to identify one or more structural elements of a document; determining, based at least in part on the one or more structural elements, one or more pairs of words within the document having a hypernym relationship; extracting de-identified content within the document based on one or more de-identification techniques applied to the document; and applying a set of causal rules to the de-identified content and the one or more pairs of words to annotate at a least a portion of the de-identified content as belonging to a class of protected content.
1. A computer-implemented method, the method comprising:
applying one or more rules to identify one or more structural elements of a document, wherein the one or more structural elements are indicative of at least one of an organization of text within the document and a visual presentation of text within the document;
determining, based at least in part on the one or more structural elements, one or more pairs of words within the document having a hypernym relationship;
extracting de-identified content within the document based on one or more de-identification techniques applied to the document;
applying a set of causal rules to the de-identified content and the one or more pairs of words by computing mutual dependence and/or mutual independence between the de-identified content and the one or more pairs of words;
annotating, based at least in part on the computed mutual dependence and/or mutual independence, at least a portion of the de-identified content as belonging to a class of protected content; and
training a machine learning model on a set of training data to automatically identify at least a portion of the de-identification techniques in at least one other document, wherein the set of training data comprises the annotated portion of the de-identified content;
wherein the method is carried out by at least one computing device.
2. The computer-implemented method of claim 1 , wherein determining the one or more pairs of words comprises determining one or more contexts in which a given one of the words is used within the document.
3. The computer-implemented method of claim 2 , wherein the one or more contexts correspond to the one or more structural elements.
4. The computer-implemented method of claim 2 , wherein the determining the one or more contexts comprises computing a set of context vectors for the given word, wherein the set of context vectors comprises:
a first vector indicating the contexts in which the given word is used as a relational hypernym;
a second vector indicating the contexts in which the given word is used as a relational hyponym; and
a third vector indicating the contexts in which the given word is used as a general hypernym.
5. The computer-implemented method of claim 4 , wherein computing the set of context vectors comprises at least one of:
determining a number of times the given word appears in the document;
determining a number of times the given word precedes each of the one or more structural elements; and
determining a number of times the given word succeeds each of the one or more structural elements.
6. The computer-implemented method of claim 1 , wherein the one or more structural elements relate to at least one of:
at least one of a type and a style of font;
a type of spacing;
a type of indention;
at least one of: a shape or a position of a shape within the document;
one or more symbols; and
one or more keywords.
7. The computer-implemented method of claim 1 , wherein the de-identification techniques comprise at least one of: one or more anonymization techniques and one or more pseudonymization techniques.
8. The computer-implemented method of claim 7 , wherein the machine learning model comprises a deep neural network, and wherein the annotated portion of the de-identified content comprises one or more labeled samples to identify at least a portion of the de-identification techniques.
9. The computer-implemented method of claim 1 , wherein the set of causal rules comprises at least one of a CCU causality rule and a CCC causality rule.
10. The computer-implemented method of claim 1 , wherein
the mutual independence and/or the mutual dependence between the de-identified content and the one or more pairs of words is computed by applying one or more thresholds.
11. The computer-implemented method of claim 1 , wherein the class of protected content is defined by a regulatory document.
12. The computer-implemented method of claim 1 , wherein software is provided as a service in a cloud environment.
13. A computer program product comprising:
a computer readable storage medium; program instructions embodied therewithstored in the computer readable storage medium, for causing a computing device to perform the following computer operations:
apply one or more rules to identify one or more structural elements of a document, wherein the one or more structural elements are indicative of at least one of an organization of text within the document and a visual presentation of text within the document;
determine, based at least in part on the one or more structural elements, one or more pairs of words within the document having a hypernym relationship;
extract de-identified content within the document based on one or more de-identification techniques applied to the document;
apply a set of causal rules to the de-identified content and the one or more pairs of words by computing mutual dependence and/or mutual independence between the de-identified content and the one or more pairs of words;
annotate, based at least in part on the computed mutual dependence and/or mutual independence, at least a portion of the de-identified content as belonging to a class of protected content; and
train a machine learning model on a set of training data to automatically identify at least a portion of the de-identification techniques in at least one other document, wherein the set of training data comprises the annotated portion of the de-identified content.
14. The computer program product of claim 13 , wherein determining the one or more pairs of words comprises determining one or more contexts in which a given one of the words is used within the document.
15. The computer program product of claim 14 , wherein the one or more contexts correspond to the one or more structural elements.
16. The computer program product of claim 14 , wherein the determining the one or more contexts comprises computing a set of context vectors for the given word, wherein the set of context vectors comprises:
a first vector indicating the contexts in which the given word is used as a relational hypernym;
a second vector indicating the contexts in which the given word is used as a relational hyponym; and
a third vector indicating the contexts in which the given word is used as a general hypernym.
17. The computer program product of claim 13 , wherein the de-identification techniques comprise at least one of: one or more anonymization techniques and one or more pseudonymization techniques.
18. The computer program product of claim 13 , wherein the set of causal rules comprises at least one of a CCU causality rule and a CCC causality rule.
19. The computer program product of claim 13 , wherein
the mutual independence and/or the mutual dependence between the de-identified content and the one or more pairs of words is computed by applying one or more thresholds.
20. A system comprising:
a memory configured to store program instructions;
a processor operatively coupled to the memory to execute the program instructions to:
apply one or more rules to identify one or more structural elements of a document, wherein the one or more structural elements are indicative of at least one of an organization of text within the document and a visual presentation of text within the document;
determine, based at least in part on the one or more structural elements, one or more pairs of words within the document having a hypernym relationship;
extract de-identified content within the document based on one or more de-identification techniques applied to the document;
apply a set of causal rules to the de-identified content and the one or more pairs of words by computing mutual dependence and/or mutual independence between the de-identified content and the one or more pairs of words;
annotate, based at least in part on the computed mutual dependence and/or mutual independence, at least a portion of the de-identified content as belonging to a class of protected content; and
train a machine learning model on a set of training data to automatically identify at least a portion of the de-identification techniques in at least one other document, wherein the set of training data comprises the annotated portion of the de-identified content.