METHOD AND DEVICE FOR INFORMATION RETRIEVAL

• US 20110302115A1
• Filed: 05/24/2011
• Published: 12/08/2011
• Est. Priority Date: 06/02/2010
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: KEMP, Thomas
• Granted Patent: US 8,666,915 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06F 16/3346 using probabilistic model
• CPC Class Codes: H04N 21/4756 for rating content, e.g. scoring a recommended movie

Abstract

A method of information retrieval, comprising: determining, using a microprocessor, Q generative models (λ) in accordance with Probabilistic Latent Semantic Indexing (PLSI), said Q generative models being determined in an offline training; receiving a user query (q); choosing, using said microprocessor, N generative models out of the Q generative models, with N≦Q; determining, using said microprocessor, a content item (d) based on said query and a combination of the N generative models.

Claims

1. A method of information retrieval, comprising:

determining, using a microprocessor, Q generative models (λ

) in accordance with Probabilistic Latent Semantic Indexing (PLSI), said Q generative models being determined in an offline training;

receiving a user query (q);

choosing, using said microprocessor, N generative models out of the Q generative models, with N<

Q;

determining, using said microprocessor, a content item (d) based on said query and a combination of the N generative models.

2. The method of claim 1, wherein Q is larger than N by a factor between 2 to 6.

3. The method of claim 1, wherein said step of choosing includes

determining, for each of the Q generative models, a value based on a quality estimation function, wherein the N generative models are chosen depending on the value.

4. The method of claim 3, wherein the quality estimation function is determined based on a cross validation data set, leaving-one-out prediction, and/or half-query prediction.

5. The method of claim 3, wherein the quality estimation function depends on the query, such that the choosing of the N generative models depends on the query, wherein the step of choosing is performed at runtime when the query is received.

6. The method of claim 5, wherein the query is split into a first part q1 and a second part q2 and the quality estimation function depends on a first probability p1 and a second probability p2, wherein

p1=p(q2|q1), and

p2=p(q1|q2).

7. The method of claim 6, wherein, before determining p1, the folding-in technique of PLSI is used to estimate the PLSI decomposition model p(aspect|q1) which is then used to determine p1.

8. The method of claim 6, wherein, before determining p2, the folding-in technique of PLSI is used to estimate the PLSI decomposition model p(aspect|q2) which is then used to determine p2.

9. The method of claim 6, wherein the first part q1 corresponds substantially to the first half of the query and the second part q2 corresponds substantially to the second half of the query.

10. The method of claim 5, wherein the query comprises a plurality of N words w1, . . . , w−

N, and the quality estimation function depends on a probability p-loo, wherein for calculating p-loo a probability p3(x) is calculated, wherein

p3(x)=p(w−

x|r),

wherein w−

x is a word selected from the plurality of words w1, . . . , w−

N, and

r represents all of the plurality of words w1, . . . , w−

N except w−

x.

11. The method of claim 5, wherein the query comprises a plurality of N words w1, . . . , w−

N, and the quality estimation function is given by

12. A method of claim 1, wherein a training material for said off-line training includes Electronic Program Guide (EPG) data and said user query is related to finding EPG data.

13. A device comprising:

a microprocessor adapted to calculate Q generative models (λ

) in accordance with Probabilistic Latent Semantic Indexing (PLSI), said Q generative models being determined in an offline training;

a storage adapted to store said Q generative models;

an input interface adapted to receive a user query (q);

a selection unit adapted to select N generative models out of the Q generative models, with N<

Q;

wherein said microprocessor is further adapted to determine a content item (d) based on said query and a combination of the N generative models.

14. The device of claim 13, wherein said selection unit is further adapted to determine, at runtime when a query has been received by the input interface, for each of the Q generative models, a value based on a quality estimation function, wherein the N generative models are chosen depending on the value, wherein the quality estimation function depends on the query, wherein the query is split into a first part q1 and a second part q2 and the quality estimation function depends on a first probability p1 and a second probability p2, wherein

p1=p(q2|q1), and

p2=p(q1|q2).

15. The device of claim 13, wherein said selection unit is further adapted to determine, at runtime when a query has been received by the input interface, for each of the Q generative models, a value based on a quality estimation function, wherein the N generative models are chosen depending on the value, wherein the query comprises a plurality of N words w1, . . . , w−

N, and the quality estimation function is given by

16. The device of claim 13, wherein the device is a server and the input interface is adapted to receive the user query from a client.

17. A television equipment comprising:

a data processor adapted to calculate Q generative models (λ

) in accordance with Probabilistic Latent Semantic Indexing (PLSI), said Q generative models being determined in an offline training, wherein a training material for said offline training comprises EPG data;

a storage adapted to store said Q generative models;

an input interface adapted to receive a user query (q), the user query being related to television program information;

a selection unit adapted to select N generative models out of the Q generative models, with N<

Q;

wherein said data processor is further adapted to determine a content item (d) based on said query and a combination of the N generative models, said content item being related to a television program.