Categorization for a global taxonomy

1. A method performed by a computing system for generating training data for training a classifier to assign nodes of a taxonomy graph to items based on item descriptions, each node having a label, the method comprising:

accessing item descriptions of items and the taxonomy graph,

wherein an item description includes a title and one or more category labels;

for each item,

identifying for that item one or more candidate paths within the taxonomy graph that are relevant to that item, the candidate paths identified based on content of the item description of that item matching a feature of the nodes, a candidate path being a sequence of nodes starting from a root node of the taxonomy graph,

wherein the identifying of the candidate paths for an item comprises:

identifying one or more candidate paths that have a leaf node with a label that matches a tail portion of the title of the item description for that item;

for each identified candidate path,

 generating a relevance score for that candidate path;

 determining whether the candidate path includes a partial path that is a category path, a category path being a path of the taxonomy graph with a leaf node that has a label that matches a category label; and

 upon determining that the candidate path includes a partial path that is a category path, adjusting the relevance score of that candidate path to indicate an increased relevance; and

discarding candidate paths with a relevance score that does not satisfy a threshold relevance; and

for each identified candidate path, labeling the item description with the candidate path,

wherein the labeled item descriptions compose the training data.

2. The method of claim 1 wherein the identifying of the candidate paths further comprises, prior to discarding the candidate paths:

for each candidate path,

determining whether a category path includes a partial path that is the candidate path; and

upon determining that a category path includes a partial path that is the candidate path, adjusting the relevance score of that candidate path to indicate an increased relevance.

3. The method of claim 1 further comprising:

determining whether no candidate path has a leaf node with a label that matches a tail portion of the title of the item description for that item;

upon determining that no candidate path has such a leaf node, identifying one or more candidate paths that have a leaf node with a label that matches any portion of the title of the item description for that item.

4. The method of claim 1 wherein an item description includes fields having field names and having field content and further comprising, prior to labeling:

accessing rules, each rule associated with a node, each rule having a field name test, a content test, and a relevance adjustment;

for each candidate path,

for each node along the candidate path,

for each rule associated with that node,

for each field with field name satisfies the field name test of that rule and with field content that satisfies the field content test of that rule, adjusting the relevance score of that candidate path based on the relevance adjustment of that rule.

5. The method of claim 1 further comprising, prior to labeling:

determining whether multiple candidate paths have been identified; and

upon determining that multiple candidate paths have been identified,

for each candidate path, generating a similarity score for that candidate path based on similarity between a word vector for the item description and a word vector for the candidate path; and

discarding candidate paths other than the candidate paths with the similarity score above a threshold.

6. The method of claim 1 further comprising training a classifier using the labeled item descriptions as training data.

7. The method of claim 6 further comprising:

for each item, applying the classifier to the item description of that item to assign a classification label and a classification score to that item;

for each classification label, calculating a label threshold for the classification label based on the classification scores of items with that classification label, the label threshold based on accuracy of the classifier at assigning the classification label.

8. A method performed by a computing system for identifying a path of a taxonomy graph to assign to an item description of an item, the method comprising:

identifying one or more candidate paths within the taxonomy graph that are relevant to the item, the candidate paths identified based on content of the item description of the item matching labels of nodes, each candidate path having a relevance score, a path being a sequence of nodes starting from a root node of the taxonomy graph;

applying a classifier to the item description to identify classification paths, each classification path having a classification score;

for each candidate path, generating a final relevance score for that candidate path by combining the relevance score for that candidate path with the classification score for the corresponding classification path if any; and

assigning to the item one or more candidate paths with a final relevance score indicating that the candidate path is relevant to the item.

9. The method of claim 8 wherein an item description includes a title and one or more category labels and wherein the identifying of the candidate paths for an item comprises:

identifying one or more candidate paths that have a leaf node with a label that matches a tail portion of the title of the item description for that item;

for each identified candidate path,

generating a relevance score for that candidate path;

determining whether the candidate path includes a partial path that is a category path, a category path being a path of the taxonomy graph with a leaf node that has a label that matches a category label; and

upon determining that the candidate path includes a partial path that is a category path, adjusting the relevance score of that candidate path to indicate an increased relevance; and

discarding candidate paths with a relevance score that does not satisfy a threshold relevance.

10. The method of claim 9 wherein the identifying of the candidate paths further comprises, prior to discarding the candidate paths:

for each candidate path,

determining whether a category path includes a partial path that is the candidate path; and

upon determining that a category path includes a partial path that is the candidate path, adjusting the relevance score of that candidate path to indicate an increased relevance.

11. The method of claim 9 further comprising:

determining whether no candidate path has a leaf node with a label that matches a tail portion of the title of the item description for that item;

upon determining that no candidate path has such a leaf node, identifying one or more candidate paths that have a leaf node with a label that matches any portion of the title of the item description for that item.

12. The method of claim 9 wherein an item description includes fields having field names and having field content and further comprising, prior to selecting the candidate paths:

accessing rules, each rule associated with a node, each rule having a field name test, a content test, and a relevance adjustment;

for each candidate path,

for each node along the candidate path,

for each rule associated with that node,

for each field with field name satisfies the field name test of that rule and with field content that satisfies the field content test of that rule, adjusting the relevance score of that candidate path based on the relevance adjustment of that rule.

13. The method of claim 8 further comprising for each candidate path, adjusting the relevance score of the candidate path based on a normalization threshold for the classification label of that candidate path, the normalization threshold based on accuracy of the classifier at assigning the classification label.

14. One or more computing systems for identifying a path of a taxonomy graph to assign to an item description of an item, the one or more computing systems comprising:

one or more computer-readable storage mediums storing computer-executable instructions for controlling the one or more computing systems to:

identify one or more candidate paths within the taxonomy graph that are relevant to the item, each candidate path having a relevance score;

apply a classifier to the item description to identify classification paths, each classification path having a classification score;

for each candidate path, generate a final relevance score for that candidate path by combining the relevance score for that candidate path with the classification score for the corresponding classification path; and

assign to the item one or more candidate paths with a final relevance score indicating that the candidate path is relevant to the item; and

one or more processors for executing the computer-executable instructions stored in the one or more computer-readable storage mediums.

15. The one or more computing systems of claim 14 wherein an item description includes a title and one or more category labels and wherein the instructions that identify of the candidate paths for an item comprises instructions to:

identify one or more candidate paths that have a leaf node with a label that matches a tail portion of the title of the item description for that item;

for each identified candidate path,

generate a relevance score for that candidate path;

determine whether the candidate path includes a partial path that is a category path, a category path being a path of the taxonomy graph with a leaf node that has a label that matches a category label; and

upon determining that the candidate path includes a partial path that is a category path, adjust the relevance score of that candidate path to indicate an increased relevance; and

discard candidate paths with a relevance score that does not satisfy a threshold relevance.

16. The one or more computing systems of claim 15 wherein the instructions that identify of the candidate paths for an item comprises instructions to, prior to discarding the candidate paths:

for each candidate path,

determine whether a category path includes a partial path that is the candidate path; and

upon determining that a category path includes a partial path that is the candidate path, adjust the relevance score of that candidate path to indicate an increased relevance.

17. The one or more computing systems of claim 15 wherein the instructions further comprise instructions to:

determine whether no candidate path has a leaf node with a label that matches a tail portion of the title of the item description for that item;

upon determining that no candidate path has such a leaf node, identify one or more candidate paths that have a leaf node with a label that matches any portion of the title of the item description for that item.

18. The one or more computing systems of claim 15 wherein an item description includes fields having field names and having field content and wherein the instructions further comprise instructions to, prior to selecting the candidate paths:

access rules, each rule associated with a node, each rule having a field name test, a content test, and a relevance adjustment;

for each candidate path,

for each node along the candidate path,

[1] for each rule associated with that node,

for each field with field name satisfies the field name test of that rule and with field content that satisfies the field content test of that rule, adjust the relevance score of that candidate path based on the relevance adjustment of that rule.

19. The one or more computing systems of claim 14 wherein the instructions further comprise instructions to, for each candidate path, adjust the relevance score of the candidate path based on a normalization threshold for the classification label of that candidate path, the normalization threshold based on accuracy of the classifier at assigning the classification label.