Granted Patent B2

US 12,566,818 · App. 17/118,700 · Granted Mar 3, 2026

Inventors: Huaiyu Zhu (Fremont, CA); Yunyao Li (San Jose, CA); Youxuan Jiang (Ypsilanti, MI)

Assignee: INTERNATIONAL BUSINESS MACHINES CORPORATION

G06F18/2155G06F18/2113G06F40/205G06F40/30G06N20/00

Abstract

Claims (24)

1 . A computer-implemented method comprising:

receiving, by a device operatively coupled to a processor, a training data set through an online communication network from a client subsystem of a plurality of candidate subsystems, wherein the training data sets comprise image data;

filtering out, by the device, one or more rejected candidate labels from a plurality of candidate labels based upon receipt of electronic information transmitted over a computer network from one or more disparate locations and indicative of the one or more rejected candidate labels being chosen less than a defined frequency as a correct label by a plurality of nonexpert annotators, wherein the filtering out results in a reduced subset of candidate labels;

output, from the device, to an expert annotator external to the device, one or more polls to determine what each expert annotator believes to be the correct description of the image of each of the training data sets;

receiving, by the device, feedback from outside of the device, of one or more responses from the one or more polls determining what each expert annotator believes to be the correct description of the image of each of the training data sets;

selecting, by the device, from the reduced subset of candidate labels, between a plurality of optimal labels based upon the one or more responses received from at least one expert annotator of a plurality of expert annotators, wherein each expert annotator of the plurality of expert annotators has an expertise level above a defined expert value;

adding, by the device, as metadata, an optimal label of the plurality of optimal labels, to the training data set;

selecting, by the device, a one of the plurality of candidate subsystems to which to transmit, over the communication network, the optimal label, wherein the selection is based on an expertise of the plurality of expert annotators and subject matter of the optimal label;

transmitting, by the device, over the communication network, to a selected one of the plurality of candidate subsystems, the optimal label, wherein the selected one of the plurality of candidate subsystems; and

applying, by the device, labeled training data sets to the training data sets; and

training, by the device, a machine learning algorithm, employing supervised learning, based on applying the labeled training data sets.

2 . The computer-implemented method of claim 1 , wherein the selecting the optimal answer comprises:

exposing the reduced subset of candidate labels to at least one of the expert annotators; and

receiving, from at least one of the expert annotators, and identification of the optimal label from the reduced subset of candidate labels.

3 . The computer-implemented method of claim 1 , wherein the filtering out is iteratively performed at least twice, and wherein each iteration of the filtering removes at least one candidate label from the reduced subset of candidate labels.

4 . The computer-implemented method of claim 1 , wherein the filtering out is iteratively performed until only a single label remains in the reduced subset of candidate labels.

5 . The computer-implemented method of claim 1 , wherein the filtering out comprises electronically voting, on an electronic device displaying the plurality of labels, by a crowd of the nonexpert annotators from the one or more disparate locations over the computer network, which comprise at least a portion of the plurality of nonexpert annotators, so that only labels deemed unlikely by a defined proportion of the nonexpert annotators are removed to obtain the reduced subset of candidate labels.

6 . The computer-implemented method of claim 1 , wherein the selecting comprises electronically voting by a crowd of nonexpert annotators, in response to a computer poll sent over a computer network to respective electronic devices associated with the crowd of nonexpert annotators, wherein the crowd of nonexpert annotators comprise at least a portion of the plurality of nonexpert annotators, so that a candidate answer, of the reduced subset of candidate labels, deemed correct by a defined proportion of the plurality of nonexpert annotators are selected as the optimal label.

7 . The computer-implemented method of claim 1 , further comprising:

receiving, by the device, from the plurality of nonexpert annotators, an indication of a relative level of uncertainty or difficulty of choosing a label;

determining, by the device, whether the selection of candidate labels is too difficult for the plurality of nonexpert annotators, wherein the determining is based upon the indication of the relative level of uncertainty or difficulty being greater than a defined uncertainty or difficulty threshold value; and

responsive to the determining that the selection of candidate labels is too difficult for the plurality of nonexpert annotators, communicating, by the device, an entirety of the plurality of candidate labels to at least one of the plurality of-expert annotators.

8 . The computer-implemented method of claim 1 , further comprising:

applying, by the device, the optimal label as metadata for use in semantic role labeling, wherein a task is to identify a correct label for a predicate, or a correct label for an argument role.

Assignments (1)

Continuity (1)

References Cited (46)