High scale adaptive search systems and methods

• US 7,627,564 B2
• Filed: 06/21/2005
• Issued: 12/01/2009
• Est. Priority Date: 06/21/2005
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Liu, Jun, Katariya, Sanjeev, Yao, Qi
• Primary Examiner(s): Jalil; Neveen Abel
• Assistant Examiner(s): Pulliam; Christyann R
• Application Number:
  US11/157,600
• Publication Number:
  US 20060287993A1
• Time in Patent Office: 1,624 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06F 16/24578 using ranking
• CPC Class Codes: Y10S 707/99934 Query formulation, input preparation, or translation
• CPC Class Codes: Y10S 707/99935 Query augmenting and refining, e.g. inexact access
• CPC Class Codes: Y10S 707/99937 Sorting

Abstract

The subject invention relates to systems and methods that employ automated learning techniques to database and information retrieval systems in order to facilitate knowledge capabilities for users and systems. In one aspect, an adaptive information retrieval system is provided. The system includes a database component to store structured and unstructured data values. A search component queries the data values from the database, wherein a learning component associated with the search component or the database component is provided to facilitate retrieval of desired information.

Claims

1. A computer comprising:

a processing unit; and

a computer readable storage medium storing;

a database component that stores a plurality of assets, the plurality of assets including structured assets and unstructured assets, the unstructured assets not containing metadata describing data in the unstructured assets, the structured assets containing metadata describing data in the structured assets;

a search engine that, when executed by the processing unit, causes the computer to perform a query on the plurality of assets; and

a learning component that, when executed by the processing unit, causes the computer to;

calculate ranks for assets in a result set, each asset in the result set being an asset in the database component satisfying the query, the ranks for assets in the result set based at least in part on feedback, the result set including at least one of the structured assets and at least one of the unstructured assets, the ranks for each structured asset in the result set being based at least in part on the metadata in the structured asset, higher ranked assets in the result set having a greater probability of being associated with a user'"'"'s actual information goal, the user'"'"'s actual information goal being that which a user actually wants to accomplish when the user submits the query; and

take an action directly related to the user'"'"'s actual information goal,

wherein the learning component causes the computer to calculate the ranks for the assets in the result set at least in part by causing the computer to;

calculate a static rank factor for each asset in the result set, the static rank factor for an asset being a measure of how well linked to the asset is, the static rank of the asset being independent of terms in the query;

calculate a quality of asset factor for each asset in the result set, the quality of asset factor for an asset being a measure of an inherent quality of the asset, the quality of the asset being independent of the terms in the query;

calculate a quality of match factor for each asset in the result set, the quality of match factor for an asset being a measure of how well contents of the asset and metadata of the asset match the query, the quality of match factor for the asset based at least in part on the past query click-through data; and

calculate the ranks for each asset in the result set based at least in part on;

the static rank factor for each asset in the result set, the quality of asset factor for each asset in the result set, and the quality of match factor for each asset in the result set.

2. The computer of claim 1, wherein the learning component, when executed by the processing unit, causes the computer to employ at least one Bayesian classifier to calculate the ranks for the assets in the result set.

3. The computer of claim 1, wherein the search engine, when executed by the processing unit, causes the computer to process natural language input to generate the query.

4. The computer of claim 1, wherein the feedback is implicit feedback.

5. The computer of claim 4, wherein the implicit feedback comprises the past query click-through data.

6. The computer of claim 1, wherein the database component is a relational database.

7. The computer of claim 1, wherein the learning component, when executed by the processing unit, causes the computer to calculate the ranks for assets in the result set based on types of queries, the structure of data in the database, or from implicit queries.

8. The computer of claim 7, wherein the learning component, when executed by the processing unit, causes the computer to perform the action by causing the computer to execute an application function based at least in part upon an inference of the user'"'"'s actual information goal.

9. The computer of claim 1, wherein the computer readable storage medium further stores a dialog component that, when executed by the processing unit, causes the computer to initiate a dialog with the user to mitigate system uncertainty, the dialog includes at least one of stating an intent or clarifying a slot.

10. The computer of claim 1, wherein the learning module further comprises a semantic reasoning engine that, when executed by the processing unit, causes the computer to analyze implicit or explicit feedback from the user.

11. The computer of claim 1, wherein the computer readable medium further comprises instructions that, when executed by the processing unit, cause the computer to provide a user interface that collects the feedback from a user.

12. An automated information retrieval method comprising:

receiving, by a computer, a query submitted by a user;

obtaining, by the computer, a result set by performing the query on a relational database that stores a plurality of assets, the plurality of assets including unstructured assets and structured assets, the result set comprising each asset in the plurality of assets that satisfies the query, the result set comprising at least one of the structured assets and at least one of the unstructured assets, the unstructured assets not containing metadata describing data in the unstructured assets, the structured assets containing metadata describing data in the structured assets;

calculating, by the computer, ranks for each asset in the result set, higher ranked assets in the result set having a greater probability of being associated with a user'"'"'s actual information goal, the ranks for each structured asset in the result set being based at least in part on the metadata in the structured asset, the user'"'"'s actual information goal being that which a user actually wants to accomplish when the user submits the query, the ranks for each asset in the result set based at least in part on past query click-through data; and

presenting, by the computer, the assets in the result set to the user in an order based on the ranks for the assets in the result set,

wherein calculating the ranks for each asset in the result set comprises;

calculating, by the computer, a static rank factor for each asset in the result set, the static rank factor for an asset being a measure of how well linked to the asset is, the static rank of the asset being independent of terms in the query;

calculating, by the computer, a quality of asset factor for each asset in the result set, the quality of asset factor for an asset being a measure of an inherent quality of the asset, the quality of the asset being independent of the terms in the query;

calculating, by the computer, a quality of match factor for each asset in the result set, the quality of match factor for an asset being a measure of how well contents of the asset and metadata of the asset match the query, the quality of match factor for the asset based at least in part on the past query click-through data; and

calculating, by the computer, the ranks for each asset in the result set based at least in part on;

the static rank factor for each asset in the result set, the quality of asset factor for each asset in the result set, and the quality of match factor for each asset in the result set.

13. The method of claim 12, wherein calculating the ranks for each asset in the result set comprises:

for each asset in the result set, calculating, by the computer, a normalized weighted sum of the static rank of the asset, the quality of page factor of the asset, a junk page probability of the asset, and the quality of match factor of the asset.

14. A system memory comprising software that, when executed by a processing unit of a computer, causes the computer to:

use at least one index to obtain, at the computer, a result set by performing a query on a relational database, the query submitted by a user of the computer, the query containing query terms, the relational database storing a plurality of assets, the plurality of assets comprising unstructured assets and structured assets, the unstructured assets not containing metadata describing data in the unstructured assets, the structured assets containing metadata describing data in the structured assets, the result set comprising each asset in the plurality of assets that contains the query terms;

calculate, at the computer, an overall rank for each asset in the result set, assets in the result set having higher overall ranks are more likely to be associated with a user'"'"'s actual information goal, the user'"'"'s actual information goal being that which the user actually wants to accomplish when the user submits the query, the overall rank for each asset in the result set being equal to a quotient that results from dividing a numerator by a denominator, the numerator being a sum of;

a product of a static rank coefficient and a static rank factor for the asset, the static rank factor for the asset being a measure of how well linked to the asset is, the static rank factor of the asset being independent of the query terms, the static rank factor for the asset being a real number between 0 and 1;

a product of a quality of page coefficient and a quality of asset factor for the asset, the quality of asset factor for the asset being a measure of an inherent quality of the asset, the quality of the asset factor being independent of the query terms, the quality of asset factor for the asset being a real number between 0 and 1;

a product of a junk page probability coefficient and a junk page probability factor for the asset, the junk page probability factor for the asset being a probability that the asset is a junk page, the junk page probability factor for the asset being a real number between 0 and 1;

a product of a query weight and a quality of match factor for the asset, the query weight based on a length of the query and rarity of the query terms, the quality of match factor for an asset being a measure of how well contents of the asset and metadata of the asset match the query, the quality of match factor for the asset being a real number between 0 and 1, the quality of match factor for the asset being a sum of;

a product of an information retrieval rank coefficient and a similarity function factor, the similarity function factor measuring similarity between the query and the asset, the similarity function factor being a real number between 0 and 1,

a product of a query feedback probability coefficient, a confidence level for feedback model training, and a probability of the asset given the query, wherein the probability of the asset given the query is based on past query click-through data, and

a product of a structured information retrieval coefficient, a confidence level for structural model training, and a structural similarity factor for the asset, wherein the structural similarity factor for the asset is a measure of structural similarity between the query and a structure of the asset, the structural similarity factor for the asset being a real number between 0 and 1;

wherein the denominator is equal to a sum of;

the static rank coefficient,

the quality of page coefficient,

the junk page probability coefficient,

a product of the query weight and the information retrieval rank coefficient,

a product of the query weight and the query feedback probability coefficient, and

a product of the query weight and the structured information retrieval coefficient; and

present, by the computer on an output device, assets in the result set to the user in an order based on the overall ranks for the assets in the result set.

15. The system memory of claim 14, wherein the software, when executed by the processing unit, further causes the computer to sort assets in the result set based on the overall ranks of the assets in the result set.