Electronic communications signature recognition for privacy preserving computer operations
Mechanisms are provided for executing a privacy preserving computing operation. Features are extracted from line(s) of an electronic communication. Trained machine learning (ML) computer model(s) process the extracted features and generate classification output(s) for the line(s). The ML computer models process the features and generate a classification output for each line which specifies, for corresponding lines, a corresponding classification of the corresponding line as to whether it is a signature line or non-signature line, based on patterns in the features. Each line is annotated with metadata specifying the corresponding classification of the line based on the classification output, to generate an annotated set of data. The privacy preserving computing operation is executed on the electronic communication based on the annotated set of data.
1 . A computer-implemented method in a data processing system, for executing a privacy preserving computing operation, the method comprising:
retrieving a set of data, from a data storage, that comprises a first electronic communication corresponding to the privacy preserving computing operation;
extracting first features from one or more first lines of the first electronic communication;
processing, by one or more trained machine learning computer models, the extracted first features as input to the one or more trained machine learning computer models and which generate classification outputs for the one or more first lines, wherein the one or more trained machine learning computer models process the first features and generate a classification output for each first line in the one or more first lines, which specifies, for a corresponding first line in the one or more first lines, a corresponding classification of the corresponding first line as to whether it is a signature line or non-signature line, based on patterns in the first features;
processing, by the one or more trained machine learning computer models, extracted second features extracted from one or more second lines of one or more other second electronic communications in the set of data, to generate the classification output for each first line in the one or more first lines of the first electronic communication, wherein the first features are local features corresponding to a first electronic communication thread in which the first electronic communication is a part, and wherein the second features are global features corresponding to the one or more other second electronic communications which are not part of the first electronic communication thread;
annotating each first line in the one or more first lines with metadata specifying the corresponding classification of the first line based on the classification output, to thereby generate an annotated set of data; and
executing the privacy preserving computing operation on the first electronic communication based on the annotated set of data.
2 . The computer-implemented method of claim 1 , wherein executing the privacy preserving computing operation comprises:
processing the one or more first lines of the first electronic communication that are annotated as being signature lines to identify one or more instances of personal information present in the one or more first lines; and
obfuscating the one or more instances of personal information to generate obfuscated electronic communication data.
3 . The computer-implemented method of claim 2 , wherein executing the privacy preserving computing operation further comprises executing one or more operations to service a data subject access request (DSAR) based on the obfuscated electronic communication data.
4 . The computer-implemented method of claim 1 , wherein processing, by the one or more trained machine learning computer models, the extracted first features as input to the trained machine learning computer model which generates classification outputs for the one or more first lines, comprises processing, by the one or more machine learning computer models, first features from a plurality of the first lines of the first electronic communication in combination to identify a signature block of the first electronic communication based on a pattern of features from the plurality of first lines.
5 . The computer-implemented method of claim 4 , wherein identifying the signature block of the first electronic communication based on the pattern of features from the plurality of first lines comprises executing named entity recognition and natural language processing on a subset of the plurality of first lines to extract named entities and part of speech information for text in the subset of the plurality of first lines, wherein the subset comprises a portion of the plurality of first lines identified as being part of the signature block.
6 . The computer-implemented method of claim 5 , wherein identifying the pattern of features by executing the named entity recognition and natural language processing on the subset of the plurality of first lines comprises identifying, by the named entity recognition, a person name followed by one or more subsequent first lines that do not have verbs or auxiliary verbs present in the one or more subsequent first lines, as determined by the natural language processing.
7 . The computer-implemented method of claim 1 , further comprising maintaining a knowledge base with frequency information specifying, for each electronic communication sender, and for each phrase or sentence present in lines of electronic communications sent by the electronic communication sender, a frequency of occurrence, and wherein processing, by the one or more trained machine learning computer models, the extracted first features and extracted second features, further comprises processing the frequency information in the knowledge base to classify the one or more first lines of the first electronic communication.
8 . The computer-implemented method of claim 1 , further comprising generating, for each unique sender identifier of each sender of electronic communications in the set of data, and for each phrase or sentence present in lines of electronic communications sent by each sender of electronic communications, a frequency of email lines per sender identifier (FEL/SID) metric, wherein the FEL/SID metric is a ratio of a number of appearances of the corresponding phrase or sentence in an electronic communication line over a number of different electronic communications sent by the same unique sender identifier, and wherein the FEL/SID metric is processed by the one or more trained machine learning computer models along with the first features to classify the one or more first lines.
9 . The computer-implemented method of claim 1 , further comprising training the one or more machine learning computer models through an iterative machine learning training process, until a convergence condition is reached, based on a training dataset that is input to the one or more machine learning computer models, wherein:
the training dataset comprises electronic communication lines for multiple different electronic communication threads and corresponding ground truth labels for each electronic communication line in the training dataset specifying whether the corresponding electronic communication line is a signature line or not a signature line, and
the one or more machine learning computer models are trained to classify the input electronic communication lines of the electronic communications in the training dataset as to whether they are signature lines or non-signature lines at least by processing features of the input electronic communication lines, generating a predicted classification output for each input electronic communication line, comparing the predicted classification output to a corresponding ground truth label for each input electronic communication line to determine an error, and executing a machine learning training algorithm on the error to adjust operational parameters of the one or more machine learning computer models to reduce the error.
10 . A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:
retrieve a set of data, from a data storage, that comprises a first electronic communication corresponding to a privacy preserving computing operation;
extract first features from one or more first lines of the first electronic communication;
process, by one or more trained machine learning computer models, the extracted first features as input to the one or more trained machine learning computer models and which generate classification outputs for the one or more first lines, wherein the one or more trained machine learning computer models process the first features and generate a classification output for each first line in the one or more first lines, which specifies, for a corresponding first line in the one or more first lines, a corresponding classification of the corresponding first line as to whether it is a signature line or non-signature line, based on patterns in the first features;
process, by the one or more trained machine learning computer models, extracted second features extracted from one or more second lines of one or more other second electronic communications in the set of data, to generate the classification output for each first line in the one or more first lines of the first electronic communication, wherein the first features are local features corresponding to a first electronic communication thread in which the first electronic communication is a part, and wherein the second features are global features corresponding to the one or more other second electronic communications which are not part of the first electronic communication thread;
annotate each first line in the one or more first lines with metadata specifying the corresponding classification of the first line based on the classification output, to thereby generate an annotated set of data; and
execute the privacy preserving computing operation on the first electronic communication based on the annotated set of data.
11 . The computer program product of claim 10 , wherein the computer readable program further causes the computing device to execute the privacy preserving computing operation at least by:
processing the one or more first lines of the first electronic communication that are annotated as being signature lines to identify one or more instances of personal information present in the one or more first lines; and
obfuscating the one or more instances of personal information to generate obfuscated electronic communication data.
12 . The computer program product of claim 11 , wherein the computer readable program further causes the computing device to execute the privacy preserving computing operation at least by executing one or more operations to service a data subject access request (DSAR) based on the obfuscated electronic communication data.
13 . The computer program product of claim 10 , wherein the computer readable program further causes the computing device to process, by the one or more trained machine learning computer models, the extracted first features as input to the trained machine learning computer model which generates classification outputs for the one or more first lines, at least by processing, by the one or more machine learning computer models, first features from a plurality of the first lines of the first electronic communication in combination to identify a signature block of the first electronic communication based on a pattern of features from the plurality of first lines.
14 . The computer program product of claim 13 , wherein identifying the signature block of the first electronic communication based on the pattern of features from the plurality of first lines comprises executing named entity recognition and natural language processing on a subset of the plurality of first lines to extract named entities and part of speech information for text in the subset of the plurality of first lines, wherein the subset comprises a portion of the plurality of first lines identified as being part of the signature block.
15 . The computer program product of claim 14 , wherein identifying the pattern of features by executing the named entity recognition and natural language processing on the subset of the plurality of first lines comprises identifying, by the named entity recognition, a person name followed by one or more subsequent first lines that do not have verbs or auxiliary verbs present in the one or more subsequent first lines, as determined by the natural language processing.
16 . The computer program product of claim 10 , wherein the computer readable program further causes the computing device to maintain a knowledge base with frequency information specifying, for each electronic communication sender, and for each phrase or sentence present in lines of electronic communications sent by the electronic communication sender, a frequency of occurrence, and wherein processing, by the one or more trained machine learning computer models, the extracted first features and extracted second features, further comprises processing the frequency information in the knowledge base to classify the one or more first lines of the first electronic communication.
17 . The computer program product of claim 10 , wherein the computer readable program further causes the computing device to generate, for each unique sender identifier of each sender of electronic communications in the set of data, and for each phrase or sentence present in lines of electronic communications sent by each sender of electronic communications, a frequency of email lines per sender identifier (FEL/SID) metric, wherein the FEL/SID metric is a ratio of a number of appearances of the corresponding phrase or sentence in an electronic communication line over a number of different electronic communications sent by the same unique sender identifier, and wherein the FEL/SID metric is processed by the one or more trained machine learning computer models along with the first features to classify the one or more first lines.
18 . An apparatus comprising:
at least one processor; and
at least one memory coupled to the at least one processor, wherein the at least one memory comprises instructions which, when executed by the at least one processor, cause the at least one processor to:
retrieve a set of data, from a data storage, that comprises a first electronic communication corresponding to a privacy preserving computing operation;
extract first features from one or more first lines of the first electronic communication;
process, by one or more trained machine learning computer models, the extracted first features as input to the one or more trained machine learning computer models and which generate classification outputs for the one or more first lines, wherein the one or more trained machine learning computer models process the first features and generate a classification output for each first line in the one or more first lines, which specifies, for a corresponding first line in the one or more first lines, a corresponding classification of the corresponding first line as to whether it is a signature line or non-signature line, based on patterns in the first features;
process, by the one or more trained machine learning computer models, extracted second features extracted from one or more second lines of one or more other second electronic communications in the set of data, to generate the classification output for each first line in the one or more first lines of the first electronic communication, wherein the first features are local features corresponding to a first electronic communication thread in which the first electronic communication is a part, and wherein the second features are global features corresponding to the one or more other second electronic communications which are not part of the first electronic communication thread;
annotate each first line in the one or more first lines with metadata specifying the corresponding classification of the first line based on the classification output, to thereby generate an annotated set of data; and
execute the privacy preserving computing operation on the first electronic communication based on the annotated set of data.