Training a ranking component

• US 20070136281A1
• Filed: 01/05/2006
• Published: 06/14/2007
• Est. Priority Date: 12/13/2005
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Li, Hang, Cao, Yunbo, Gao, Jianfeng
• Granted Patent: US 7,783,629 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/313 Selection or weighting of terms for indexing

Abstract

A query and a factoid type selection are received from a user. An index of passages, indexed based on factoids, is accessed and passages that are related to the query, and that have the selected factoid type, are retrieved. The retrieved passages are ranked and provided to the user based on a calculated score, in rank order.

Claims

1. A method of training a ranking component configured to rank text passages retrieved from a corpus based on a factoid type selection input and based on a textual input, the method comprising:

constructing a plurality of passages from a training corpus;

matching a predefined set of queries against documents in the training corpus based on the constructed passages; and

training the ranking component based on an accuracy measure indicative of an accuracy of matching the predefined set of queries against the documents in the training corpus.

2. The method of claim 1 wherein matching comprises:

identifying matching passages in the documents that match the constructed passages.

3. The method of claim 1 and further comprising:

calculating the accuracy measure based on how closely the matching passages match the constructed passages.

4. The method of claim 2 wherein identifying matching passages comprises:

for a selected constructed passage, generating a feature vector indicative of whether a content word in the selected constructed passage resides in a selected potentially matching passage, and whether a content word in the selected potentially matching passage resides in the selected constructed passage.

5. The method of claim 4 wherein training the ranking component comprises:

calculating a ranking function corresponding to each constructed passage given a training query.

6. The method of claim 5 wherein training the ranking component further comprises:

calculating a ranking score for each constructed passage based on the corresponding ranking function.

7. The method of claim 6 wherein training the ranking component further comprises:

ranking each constructed passage relative to each training query.

8. The method of claim 7 wherein the accuracy measure is indicative of an accuracy of the ranking of the constructed passages relative to a given training query and wherein training the ranking component further comprises:

applying a weighting vector to the ranking scores based on the accuracy measure to obtain desired ranking results for the given training query.

9. The method of claim 8 wherein training the ranking component comprises:

learning the weighting vector based on the ranking of each constructed passage relative to the given training query and the accuracy measure.

10. The method of claim 9 wherein learning the weighting vector comprises:

conducting machine learning by iterating through a plurality of different values for components of the weighting vector; and

for each different value, calculating the accuracy measure.

11. The method of claim 10 wherein learning the weighting vector comprises:

choosing values for the components of the weighting vector based on whether the accuracy measure indicates an improvement in accuracy that meets a threshold value.

12. A computer readable medium storing computer readable instructions which, when executed by a computer, cause the computer to perform a method of training a ranking component that ranks passages in a corpus relative to input passages based on a factoid type selection input, the method comprising:

identifying the passages in the corpus based on a set of training queries;

ranking the identified passages relative to each training query based on a ranking score calculated by applying a weighting vector to a ranking function; and

using machine learning to learn desired values for components of the weighting vector based on an accuracy measure indicative of an accuracy of the ranking of the identified passages relative to the training queries.

13. The computer readable medium of claim 12 wherein the training function includes a multi-dimensional feature vector and wherein using machine learning to learn desired values for components of the weighting vector comprises:

setting the component values in the weighting vector corresponding to a given dimension in the feature vector to a desired value;

ranking the identified passages relative to the training queries; and

calculating the accuracy measure for the ranking using the desired value of the component values of the weighting vector.

14. The computer readable medium of claim 13 wherein using machine learning comprises:

repeating steps of setting the component values to a desired value, ranking the identified passages relative to the training queries, and calculating the accuracy measure for the ranking using the desired value, a plurality of times for each dimension in the feature vector.

15. The computer readable medium of claim 12 wherein identifying the passages in the corpus comprises:

constructing the passages from text in the corpus based on the input passages and a factoid type selection input.

16. The computer readable medium of claim 12 wherein matching comprises:

obtaining the training queries; and

matching the training queries against documents in a training corpus based on the identified passages.