Systems and methods for a scalable continuous active learning approach to information classification
View Patent ↗Systems and methods for classifying electronic information are provided by way of a Technology-Assisted Review (“TAR”) process. In certain embodiments, the TAR process is a Scalable Continuous Active Learning (“S-CAL”) approach. In certain embodiments, S-CAL selects an initial sample from a document collection, trains a classifier by using a default classification for a portion of the initial sample, scores the initial sample, selects a sub-sample from the initial sample for review, removes the reviewed sub-sample from the initial sample, and repeats the process by re-training the classifier until the initial sample is exhausted. In certain embodiments, a classification threshold is determined using a calculated estimate of the prevalence of relevant information such that the threshold classifies the information in accordance with a determined target criteria. In certain embodiments, the estimate of prevalence is determined from the results of iterations of a TAR process such as S-CAL.
1. A system for classifying information, the system comprising:
at least one computing device having a processor and physical memory, the physical memory storing instructions that cause the processor to:
receive an identification of a relevant document;
select a set of documents U from a document collection, wherein the document collection is stored on a non-transitory storage medium;
assign a default classification to one or more documents in U to be used as a training set along with the relevant document, wherein the default classification is a presumptively assigned classification not based on a user review;
train a classifier using the training set;
score the documents in U using the classifier;
remove one or more documents from the training set;
select a first batch size documents from U to form a set V;
select a first sub-sample of documents from V to form a set W;
present one or more documents in W to a reviewer;
receive from the reviewer one or more user coding decisions associated with the presented documents;
add one or more of the documents presented to the reviewer to the training set and remove said documents from U;
estimate a number of relevant documents in V using the number of relevant documents identified in the user coding decision received from the reviewer;
update the classifier using one or more documents in the training set;
estimate a prevalence of relevant documents in the document collection; and
upon determining that a stopping criteria has been reached, calculate a threshold for the classifier using the estimated prevalence, and classify the documents in the document collection using the classifier and the calculated threshold.
2. The system of claim 1 , wherein the instructions further cause the processor to:
upon determining that a stopping criteria has not been reached,
score the documents in U using the updated classifier;
select a second batch size documents from U to form a set V;
select a second sub-sample size of documents from V to form a set W; and
repeat the steps of presenting documents to a reviewer, receiving user coding decisions, adding reviewed documents to the training set and removing said documents from U, estimating the number of relevant documents in V, updating the classifier, estimating the prevalence of relevant documents, and determining whether a stopping criteria has been reached.
3. The system of claim 2 , wherein the second batch size is calculated as
first
batch
size
+
⌈
first_batch
_size
10
⌉
.
4. The system of claim 2 , wherein the size of the second sub-sample is varied between iterations.
5. The system of claim 2 , wherein the number of relevant documents in V is estimated as
R
^
=
R
^
prev
+
r
·
B
b
,
where {circumflex over (R)} prev is the value of the estimate of {circumflex over (R)} from a prior iteration, r is the number of relevant documents identified by the reviewer, B is a batch size, and b is the number of documents in W presented to a reviewer.
6. The system of claim 5 , wherein b is computed as b=second batch size if {circumflex over (R)} prev <=1 or the second batch size second sub sample size, otherwise b=the second sub-sample size.
7. The system of claim 2 , wherein the prevalence of relevant documents is estimated using intermediate results of a plurality of iterations of a TAR process.
8. The system of claim 1 , wherein the documents in set V are selected by random sampling.
9. The system of claim 1 , wherein the documents in set W are the highest scoring documents from V.
10. The system of claim 1 , wherein the stopping criteria is the exhaustion of the set U.
11. The system of claim 1 , wherein the threshold is calculated using a targeted level of recall.
12. The system of claim 1 , wherein the threshold is calculated by maximizing F 1 .
13. The system of claim 1 , wherein the documents in set Ware documents randomly sampled from V.
14. A computerized method for classifying information, the method comprising:
receiving an identification of a relevant document;
selecting a set of documents U from a document collection, wherein the document collection is stored on a non-transitory storage medium;
assigning a default classification to one or more documents in U to be used as a training set along with the relevant document, wherein the default classification is a presumptively assigned classification not based on a user review;
training a classifier using the training set;
scoring the documents in U using the classifier;
removing one or more documents from the training set;
selecting a first batch size documents from U to form a set V;
selecting a first sub-sample of documents from V to form a set W;
presenting one or more documents in W to a reviewer;
receiving from the reviewer one or more user coding decisions associated with the presented documents;
adding one or more of the documents presented to the reviewer to the training set and removing said documents from U;
estimating a number of relevant documents in V using the number of relevant documents identified in the user coding decision received from the reviewer;
updating the classifier using one or more documents in the training set;
estimating a prevalence of relevant documents in the document collection; and
upon determining that a stopping criteria has been reached, calculating a threshold for the classifier using the estimated prevalence, and classifying the documents in the document collection using the classifier and the calculated threshold.
15. The method of claim 14 , further comprising:
upon determining that a stopping criteria has not been reached,
scoring the documents in U using the updated classifier;
selecting a second batch size documents from U to form a set V;
selecting a second sub-sample size of documents from V to form a set W; and
repeating the steps of presenting documents to a reviewer, receiving user coding decisions, adding reviewed documents to the training set and removing said documents from U, estimating the number of relevant documents in V, updating the classifier, estimating the prevalence of relevant documents, and determining whether a stopping criteria has been reached.
16. The method of claim 15 , wherein the second batch size is calculated as
first
batch
size
+
⌈
first_batch
_size
10
⌉
.
17. The method of claim 15 , wherein the size of the second sub-sample is varied between iterations.
18. The method of claim 15 , wherein the number of relevant documents in V is estimated as
R
^
=
R
^
prev
+
r
·
B
b
,
where {circumflex over (R)} prev is the value of the estimate of {circumflex over (R)} from a prior iteration, r is the number of relevant documents identified by the reviewer, B is a batch size, and b is the number of documents in W presented to a reviewer.
19. The method of claim 18 , wherein b is computed as b=second batch size if {circumflex over (R)} prev <=1 or the second batch size second sub sample size, otherwise b=the second sub-sample size.
20. The method of claim 14 , wherein the prevalence of relevant documents is estimated using intermediate results of a plurality of iterations of a TAR process.
21. The method of claim 14 , wherein the documents in set V are selected by random sampling.
22. The method of claim 14 , wherein the documents in set W are the highest scoring documents from V.
23. The method of claim 14 , wherein the stopping criteria is the exhaustion of the set U.
24. The method of claim 14 , wherein the threshold is calculated using a targeted level of recall.
25. The method of claim 14 , wherein the threshold is calculated by maximizing F 1 .
26. The method of claim 14 , wherein the documents in set Ware documents randomly sampled from V.