Leveraging annotation bias to improve annotations
View Patent ↗In order to leverage annotation bias in batch annotations, obtained via crowdsourcing, on a set of comments on user posts in a social network, a system may select a subset of the comments for annotation based on how informative expected annotations for the comments in the subset are for the one or more classifiers and probabilities of occurrence of the expected annotations based on a predetermined annotation probability distribution. Note that the classifier may predict how likely the expected annotations are accurate for the comments in a given subset. Moreover, the predetermined annotation probability distribution may specify the annotation bias. In this way, the system may use the annotation bias to select the subset that is likely to receive expected annotations and, thus, are that are easier to use in training the classifier.
1. A computer-implemented method for selecting a subset of a set of comments associated with a group of documents, the method comprising:
accessing, at memory locations, the set of comments and a predetermined annotation probability distribution of annotations for another set of comments associated with another group of documents, wherein the annotation probability distribution specifies biases in the annotations for the other set of comments; and
using a computer processor that is coupled to the memory location and programmed to select the subset:
selecting the subset based on the annotation probability distribution and an objective function from a supervised-learning technique, wherein the objective function is optimized by maximizing an expression comprising the annotation probability distribution and the objective function.
2. The method of claim 1 , wherein the selecting exploits the biases in the annotations for the other set of comments.
3. The method of claim 1 , wherein
after selecting the subset, the method further comprises obtaining annotations for the subset by:
providing the subset to reviewers; and
receiving the annotations for the subset from the reviewers.
4. The method of claim 3 , wherein the classifier is dynamically updated based on the annotations for the subset.
5. The method of claim 1 , wherein:
after selecting the subset, the method further comprises obtaining annotations for the subset by:
masking a remainder of the set of comments so that only the subset is presented to reviewers; and
receiving the annotations for the subset from the reviewers.
6. The method of claim 5 , wherein the classifier is dynamically updated based on the annotations for the subset.
7. The method of claim 1 , wherein the set of comments are currently unannotated.
8. The method of claim 1 , wherein at least some of the comments in the set of comments are different than the comments in the other set of comments.
9. The method of claim 1 , wherein at least some of the documents in the group of documents are different than the documents in the other group of documents.
10. The method of claim 1 , wherein:
selecting the subset involves optimizing an objective function over batches of k comments in the set of comments;
and
during the optimizing, the objective function is weighted by the predetermined annotation probability distribution.
11. An apparatus, comprising:
one or more processors;
memory; and
a program module, wherein the program module is stored in the memory and, during operation of the apparatus, is executed by the one or more processors to select a subset of a set of comments associated with a group of documents, the program module including:
instructions for accessing, at memory locations in the memory, the set of comments and a predetermined annotation probability distribution of annotations for another set of comments associated with another group of documents, wherein the annotation probability distribution specifies biases in the annotations for the other set of comments; and
instructions for selecting the subset based on the annotation probability distribution and an objective function from a supervised-learning technique, wherein the objective function is optimized by maximizing an expression comprising the annotation probability distribution and the objective function.
12. The apparatus of claim 11 , wherein the selecting exploits the biases in the annotations for the other set of comments.
13. The apparatus of claim 11 , wherein the program module further includes instructions for obtaining annotations for the subset after selecting the subset, by:
providing the subset to reviewers; and
receiving the annotations for the subset from the reviewers.
14. The apparatus of claim 13 , wherein the classifier is dynamically updated based on the annotations for the subset.
15. The apparatus of claim 11 , wherein the program module further includes instructions for obtaining annotations for the subset after selecting the subset, by:
masking a remainder of the set of comments so that only the subset is presented to reviewers; and
receiving the annotations for the subset from the reviewers.
16. The apparatus of claim 15 , wherein the classifier is dynamically updated based on the annotations for the subset.
17. The apparatus of claim 11 , wherein the set of comments are currently unannotated.
18. The apparatus of claim 11 , wherein:
at least some of the comments in the set of comments are different than the comments in the other set of comments; and
at least some of the documents in the group of documents are different than the documents in the other group of documents.
19. The apparatus of claim 11 , wherein:
selecting the subset involves optimizing an objective function over batches of k comments in the set of comments; and
during the optimizing, the objective function is weighted by the predetermined annotation probability distribution.
20. A system, comprising:
a processing module comprising a non-transitory computer readable medium storing instructions that, when executed, cause the system to:
access, at memory locations, a set of comments and a predetermined annotation probability distribution of annotations for another set of comments associated with another group of documents, wherein the annotation probability distribution specifies biases in the annotations for the other set of comments; and
selecting a subset of the set of comments associated with a group of documents based on the annotation probability distribution and an objective function from a supervised-learning technique, wherein the objective function is optimized by maximizing an expression comprising the annotation probability distribution and the objective function.