QUALITY OF LABELED TRAINING DATA
Methods, systems, apparatus, and tangible non-transitory carrier media encoded with one or more computer programs for classifying an item. In accordance with particular embodiments, a labeling task is issued to workers participating in a crowdsourcing system. The labeling task includes evaluating an inferred classification that includes one or more of the class labels in a hierarchical classification taxonomy based at least in part on a description of the item and the class labels in the classification. Evaluation decisions are received from the crowdsourcing system. The classification is validated based on the evaluation decisions to obtain a validation result. The validating includes applying at least one consensus criterion to an aggregation of the received evaluation decisions. Data corresponding to one or more of the class labels in the classification is routed to respective destinations based on the validation result.
1 . A computer-implemented method of labeling items, comprising:
receiving an item record comprising a description of an item;
based on one or more machine learning based classifiers, inferring for the item a classification in a hierarchical classification taxonomy comprising successive levels of nodes associated with respective class labels, wherein the classification path comprises one or more of the class labels in the hierarchical classification taxonomy;
issuing, over a communications network, a labeling task to a plurality of workers participating in a crowdsourcing system, wherein the labeling task comprises evaluating the classification based at least in part on the description of the item and the one or more class labels in the classification;
receiving evaluation decisions from the crowdsourcing system;
validating the classification to obtain a validation result, wherein the validating comprises applying at least one consensus criterion to an aggregation of the received evaluation decisions;
routing, over a communications network, data corresponding to one or more of the class labels in the classification to respective destinations based on the validation result.
2 . The method of claim 1 , wherein the inferring is based on the item record.
3 . The method of claim 1 , wherein the classification comprises a classification path corresponding to an ordered sequence of respective ones of the class labels in successive levels of the hierarchical classification taxonomy.
4 . The method of claim 3 , wherein the labeling task comprises confirming the classification path based at least in part on the description of the item and an ordered sequence of the class labels in the classification path.
5 . The method of claim 4 , wherein the confirming of the classification path is additionally based on results of an online search query comprising the description of the item.
6 . The method of claim 4 , wherein the item record comprises a merchant associated with the item, and the confirming of the classification path is additionally based on the merchant.
7 . The method of claim 4 , wherein the item record comprises a price associated with the item, and the confirming of the classification path is additionally based on the price.
8 . The method of claim 1 , wherein:
the validating comprises, responsive to failure to satisfy at first consensus criterion, issuing the labeling task to at least one additional worker participating in the crowdsourcing system, and receiving a respective evaluation decision from the at least one additional worker; and
the applying comprises applying a second consensus criterion to an aggregation of the received evaluation decisions.
9 . The method of claim 1 , wherein, responsive to a validation of the classification path, designating one or more of the class labels in the classification as training data for one or more of the machine learning based classifiers.
10 . The method of claim 1 , wherein, responsive to an invalidation of the classification, the routing comprises issuing the labeling task over a communications network to at least one domain expert for relabeling.
11 . The method of claim 10 , further comprising receiving, from the at least one domain expert, a relabeled one of the one or more of the class labels in the classification, and designating the relabeled class label in the classification as training data for one or more of the machine learning based classifiers.
12 . The method of claim 1 , further comprising filtering out duplicate tasks prior to the issuing.
13 . The method of claim 1 , wherein the inferred classification extends through successive levels in the hierarchical classification taxonomy from one level in the hierarchical classification taxonomy to another level in the hierarchical classification taxonomy.
14 . The method of claim 13 , wherein the other level in the hierarchical classification taxonomy corresponds to a leaf node level in the hierarchical classification taxonomy.
15 . The method of claim 1 , wherein the inferred classification extends through successive levels in the hierarchical classification taxonomy but terminates prior to the leaf node level.
16 . The method of claim 1 , wherein the item record comprises a description of product.
17 . A computer-readable data storage apparatus comprising a memory component storing executable instructions that are operable to be executed by a processor, wherein the memory component includes:
executable instructions to infer for the item a classification in a hierarchical classification taxonomy comprising successive levels of nodes associated with respective class labels based on one or more machine learning based classifiers, wherein the classification path comprises one or more of the class labels in the hierarchical classification taxonomy;
executable instructions to issue, over a communications network, a labeling task to a plurality of workers participating in a crowdsourcing system, wherein the labeling task comprises evaluating the classification based at least in part on the description of the item and the one or more class labels in the classification;
executable instructions to receive evaluation decisions regarding to labeling task from the crowdsourcing system;
executable instructions to validate the classification to obtain a validation result, wherein the executable instructions to validate comprise executable instructions to apply at least one consensus criterion to an aggregation of the received evaluation decisions;
executable instructions to route, over a communications network, data corresponding to one or more of the class labels in the classification to respective destinations based on the validation result.
18 . The computer-readable data storage apparatus of claim 17 , wherein the classification comprises a classification path corresponding to an ordered sequence of respective ones of the class labels in successive levels of the hierarchical classification taxonomy.
19 . A system, comprising
a communication interface arranged to:
issue, over a communications network, a labeling task to a plurality of workers participating in a crowdsourcing system, wherein the labeling task comprises evaluating an inferred classification comprising an ordered sequence of respective class labels in successive levels of a hierarchical classification taxonomy based at least in part on a description of the item and the class labels in the classification path; and
receive respective evaluation decisions from the crowdsourcing system;
a processor arranged to:
validate the classification to obtain a validation result, wherein the validating comprises applying at least one consensus criterion to an aggregation of the received evaluation decisions; and
route, over a communications network, data corresponding to one or more of the class labels in the classification to respective destinations based on the validation result.
20 . The system of claim 19 , wherein, responsive to an invalidation of the classification path, the processor is arranged to transmit the labeling task over a communications network to at least one domain expert for relabeling.