System and method for generating contrastive explanations for text guided by attributes
A method, computer program product and system are provided to generate perturbed text is provided. A processor receives a string of text from a user. A processor determines one or more classifications for at least one word in the string of text by a classification model. A processor determines a plurality of perturbations of the at least one word based on the one or more classifications, where the plurality of perturbations do not share the same one or more classifications as the at least one word in the string of text. A processor selects a perturbation of the string of text based on (i) an edit distance between the string of text and the plurality of perturbations, and (ii) a fluency metric for each of the plurality of perturbations. A processor provides the perturbation of the string of text to the user.
1. A computer-implemented method comprising:
receiving a string of text from a user;
determining one or more classifications for at least one word in the string of text by a classification model;
determining a plurality of perturbations of the at least one word in the string of text such that the plurality of perturbations do not share a same classification with the one or more classifications of the at least one word in the string of text;
selecting, from the plurality of perturbations, a perturbation of the string of text that includes a respective perturbed classification that is different from the one or more classifications of the at least one word in the string of text based on (i) an edit distance between the string of text and the plurality of perturbations, and (ii) a fluency metric for each of the plurality of perturbations; and
providing, to the user, the perturbation of the string of text that includes the respective perturbed classification that is different from the one or more classifications of the at least one word in the string of text, wherein the providing identifies one or more attribute changes between the string of text and the perturbation of the string of text corresponding to the respective perturbed classification.
2. The computer-implemented method of claim 1 , wherein the perturbation of the string of text and the one or more attribute changes are provided to the classification model for reclassification of the perturbation.
3. The computer-implemented method of claim 2 , the method further comprising:
determining, by the one or more processors, one or more perturbed classifications for the perturbation of the string of text by the classification model; and
identifying, by the one or more processors, bias in the classification model based on the one or more perturbed classifications.
4. The computer-implemented method of claim 1 , wherein the fluency metric for each of the plurality of perturbations is based on output from a masked-language modeling bidirectional encoder representations from transformers (MLM-BERT) Language model.
5. The computer-implemented method of claim 1 , wherein at least one word in the plurality of perturbations is selected based on an integrated gradient neural network model of the string of text.
6. The computer-implemented method of claim 1 , wherein the edit distance indicates a number of words changed from the string of text to a perturbation of the string of text of the plurality of perturbations.
7. A computer program product comprising:
one or more computer-readable storage media and program instructions stored on the one or more computer-readable storage media, the program instructions comprising:
program instructions to receive a string of text from a user;
program instructions to determine one or more classifications for at least one word in the string of text by a classification model;
program instructions to determine a plurality of perturbations of the at least one word in the string of text such that the plurality of perturbations do not share a same classification with the one or more classifications of the at least one word in the string of text;
program instructions to select, from the plurality of perturbations, a perturbation of the string of text that includes a respective perturbed classification that is different from the one or more classifications of the at least one word in the string of text based on (i) an edit distance between the string of text and the plurality of perturbations, and (ii) a fluency metric for each of the plurality of perturbations; and
program instructions to provide, to the user, the perturbation of the string of text that includes the respective perturbed classification that is different from the one or more classifications of the at least one word in the string of text, wherein the program instructions to provide identifies one or more attribute changes between the string of text and the perturbation of the string of text corresponding to the respective perturbed classification.
8. The computer program product of claim 7 , wherein the perturbation of the string of text and the one or more attribute changes are provided to the classification model for reclassification of the perturbation.
9. The computer program product of claim 8 , the program instructions further comprising:
program instructions to determine one or more perturbed classifications for the perturbation of the string of text by the classification model; and
program instructions to identify bias in the classification model based on the one or more perturbed classifications.
10. The computer program product of claim 7 , wherein the fluency metric for each of the plurality of perturbations is based on output from a masked-language modeling bidirectional encoder representations from transformers (MLM-BERT) Language model.
11. The computer program product of claim 7 , wherein at least one word in the plurality of perturbations is selected based on an integrated gradient neural network model of the string of text.
12. The computer program product of claim 7 , wherein the edit distance indicates a number of words changed from the string of text to a perturbation of the string of text of the plurality of perturbations.
13. A computer system comprising:
one or more computer processors;
one or more computer readable storage media; and
program instructions stored on the computer readable storage media for execution by at least one of the one or more processors, the program instructions comprising:
program instructions to receive a string of text from a user;
program instructions to determine one or more classifications for at least one word in the string of text by a classification model;
program instructions to determine a plurality of perturbations of the at least one word in the string of text such that the plurality of perturbations do not share a same classification with the one or more classifications of the at least one word in the string of text;
program instructions to select, from the plurality of perturbations, a perturbation of the string of text that includes a respective perturbed classification that is different from the one or more classifications of the at least one word in the string of text based on (i) an edit distance between the string of text and the plurality of perturbations, and (ii) a fluency metric for each of the plurality of perturbations; and
program instructions to provide, to the user, the perturbation of the string of text that includes the respective perturbed classification that is different from the one or more classifications of the at least one word in the string of text, wherein the program instructions to provide identifies one or more attribute changes between the string of text and the perturbation of the string of text corresponding to the respective perturbed classification.
14. The computer system of claim 13 , wherein the perturbation of the string of text and the one or more attribute changes are provided to the classification model for reclassification of the perturbation.
15. The computer system of claim 14 , the program instructions further comprising:
program instructions to determine one or more perturbed classifications for the perturbation of the string of text by the classification model; and
program instructions to identify bias in the classification model based on the one or more perturbed classifications.
16. The computer system of claim 13 , wherein the fluency metric for each of the plurality of perturbations is based on output from a masked-language modeling bidirectional encoder representations from transformers (MLM-BERT) Language model.
17. The computer system of claim 13 , wherein at least one word in the plurality of perturbations is selected based on an integrated gradient neural network model of the string of text.