Diversity sampling for technology-assisted document review
The technology implements diversity sampling for a technology-assisted review of documents. An apparatus obtains an unlabeled set of documents and constructs a first batch of documents. The apparatus obtains labels for the documents and constructs a classification model using the labeled documents. The apparatus logs a found rate of a subsequent batch of documents from the unlabeled set of documents, the subsequent batch of documents being selected based on a comparison to the classification model. The apparatus determines that the classification model requires further training based on the found rate of the subsequent batch of documents and constructs a second batch of documents that includes an amount of diversity, which may be based on the found rate. The apparatus obtains labels for the second batch of documents and updates the model using the labeled second batch of documents. The method may be repeated to continue to refine the classification model.
1. A computer program product, comprising:
a non-transitory computer-readable storage device having computer-executable program instructions embodied thereon that, when executed by a computer, cause the computer to implement diversity sampling for technology-assisted review (“TAR”) of documents by providing additional documents not selected based on initial criteria for a classification model to select documents, the computer-executable instructions comprising:
constructing a classification model based on training data comprising documents identified as likely to be relevant, the constructed classification model being configured to identify documents that are potentially relevant based on initial criteria for the constructed classification model to select documents;
executing the constructed classification model to identify documents based on the initial criteria for the classification model to select documents;
providing at least a portion of the documents identified by the constructed classification model for review by a reviewer;
receiving an input of a label from the reviewer indicating at least one of relevant or not relevant for each of the documents identified by the constructed classification model provided for review;
determining a found rate of the documents identified by the constructed classification model provided for review that were labeled as relevant, the found rate based on a number of documents labeled as relevant and a total number of documents provided for review;
determining that the found rate meets a defined threshold;
determining, in response to the found rate meeting the defined threshold, to provide diverse documents for review with the identified documents that are provided for review, the diverse documents comprising documents that do not meet the initial criteria for selection by the constructed classification model;
executing the constructed classification model to identify the diverse documents that do not meet the initial criteria for selection by the constructed classification model; and
providing the diverse documents for review with at least another portion of the identified documents that are provided for review.
2. The computer program product of claim 1 , wherein determining that the found rate meets the defined threshold comprises determining that the found rate has dropped below a configured amount.
3. The computer program product of claim 1 , wherein determining that the found rate meets the defined threshold comprises determining that the found rate has dropped at a rate greater than a configured amount.
4. The computer program product of claim 1 , wherein determining that the found rate meets the defined threshold comprises determining that the found rate has not increased by a configured amount in a configured amount of time.
5. The computer program product of claim 1 , the computer-executable program instructions further comprising:
receiving an input of a label from the reviewer for each of the diverse documents provided for review; and
updating the initial criteria used by the established classification model to select documents based on the label received for each of the diverse documents.
6. The computer program product of claim 5 , wherein updating the initial criteria used by the classification model comprises creating training data based on the label received for each of the diverse documents and labels for other documents, and updating the classification model based on the training data.
7. The computer program product of claim 1 , further comprising determining an amount of the diverse documents to provide for review with the at least another portion of the identified documents, wherein a larger amount of diverse documents is provided for review based on the found rate having a larger deviation from the defined threshold.
8. The computer program product of claim 1 , further comprising determining an amount of the diverse documents to provide for review with the at least another portion of the identified documents, wherein an amount of the diverse documents provided for review is based on a magnitude of change in the found rate.
9. The computer program product of claim 1 , wherein the classification model is generated via a machine-learning function.
10. The computer program product of claim 1 , wherein the diverse documents provided for review are selected based on a diversity sampler process or a biased probabilistic sampler process.
11. The computer program product of claim 1 , wherein the documents are associated with a document review in a legal proceeding.
12. The computer program product of claim 1 , wherein the classification model is a hyperplane.
13. The computer program product of claim 12 , wherein the hyperplane is created by a support vector machine.
14. A computer-implemented method to implement diversity sampling for technology-assisted review (“TAR”) of documents, comprising:
constructing, using one or more computing devices, a classification model based on training data comprising documents identified as likely to be relevant, the constructed classification model being configured to identify documents that are potentially relevant based on initial criteria for the constructed classification model to select documents;
executing, using one or more computing devices, the constructed classification model to identify documents based on the initial criteria for the classification model to select documents;
providing, using the one or more computing devices, at least a portion of the documents identified by the constructed classification model for review by a reviewer;
receiving, using the one or more computing devices, an input of a label from the reviewer indicating at least one of relevant or not relevant for each of the documents identified by the constructed classification model provided for review;
determining, using the one or more computing devices, a found rate of the documents identified by the constructed classification model provided for review that were labeled as relevant, the found rate based on a number of documents labeled as relevant and a total number of documents provided for review;
determining, using the one or more computing devices, that the found rate meets a defined threshold;
determining, using the one or more computing devices and in response to the found rate meeting the defined threshold, to provide diverse documents for review with the identified documents that are provided for review, the diverse documents comprising documents that do not meet the initial criteria for selection by the constructed classification model; and
executing, using the one or more computing devices, the constructed classification model to identify the diverse documents that do not meet the initial criteria for selection by the constructed classification model; and
providing, using the one or more computing devices, the diverse documents for review with at least another portion of the identified documents that are provided for review.
15. The computer-implemented method of claim 14 , wherein determining that the found rate meets the defined threshold comprises determining that the found rate has dropped below a configured amount.
16. The computer-implemented method of claim 14 , wherein determining that the found rate meets the defined threshold comprises determining that the found rate has dropped at a rate greater than a configured amount.
17. The computer-implemented method of claim 14 , further comprising:
receiving an input of a label from the reviewer for each of the diverse documents provided for review; and
updating the initial criteria used by the established classification model to select documents based on the label received for each of the diverse documents.
18. The computer-implemented method of claim 14 , wherein determining that the found rate meets the defined threshold comprises determining that the found rate has not increased by a configured amount in a configured amount of time.
19. The computer-implemented method of claim 14 , further comprising determining, based on a magnitude of the found rate, an amount of the diverse documents to provide for review with the at least another portion of the identified documents.
20. A system to implement diversity sampling for technology-assisted review (“TAR”) of documents, comprising:
a storage resource; and
a processor communicatively coupled to the storage resource, wherein the processor executes application code instructions that are stored in the storage resource to cause the processor to:
construct a classification model based on training data comprising documents identified as likely to be relevant, the constructed classification model being configured to identify documents that are potentially relevant based on initial criteria for the constructed classification model to select documents;
execute the constructed classification model to identify documents based on the initial criteria for the classification model to select documents;
provide at least a portion of the documents identified by the constructed classification model for review by a reviewer;
receive an input of a label from the reviewer indicating at least one of relevant or not relevant for each of the documents identified by the constructed classification model provided for review;
determine a found rate of the documents identified by the constructed classification model provided for review that were labeled as relevant, the found rate based on a number of documents labeled as relevant and a total number of documents provided for review;
determine that the found rate meets a defined threshold;
determine, in response to the found rate meeting the defined threshold, to provide diverse documents for review with the identified documents that are provided for review, the diverse documents comprising documents that do not meet the initial criteria for selection by the classification model;
execute the constructed classification model to identify the diverse documents that do not meet the initial criteria for selection by the constructed classification model; and
provide the diverse documents for review with at least another portion of the identified documents that are provided for review.