Item recommendation device, item recommendation method, and computer program product

• US 10,089,364 B2
• Filed: 10/29/2015
• Issued: 10/02/2018
• Est. Priority Date: 10/31/2014
• Status: Active Grant

• Current Assignee: Toshiba Digital Solutions Corporation (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation), Toshiba Solutions Corporation (Toshiba Corporation)
• Inventors: Hamada, Shinichiro
• Primary Examiner(s): Choi, Yuk Ting
• Application Number:
  US14/926,219
• Publication Number:
  US 20160125048A1
• Time in Patent Office: 1,069 Days
• CPC Class Codes: G06F 16/243 Natural language query formulation
• CPC Class Codes: G06F 16/24575 using context
• CPC Class Codes: G06F 16/24578 using ranking

Abstract

An item recommendation device includes a context information generator and a ranker. The context information generator generates and outputs context information including user intention by performing semantic analysis on a natural language request which is input. The ranker ranks candidates of items to be presented to a user based on the context information, user information representing an attribute of the user, and history information representing item usage history of the user.

Claims

1. An item recommendation device comprising:

processing circuitry configured to function as;

a search information generator that generates search information by performing semantic analysis on a natural language request which is input;

a candidate extractor that performs searching with respect to a storing unit storing therein item information using the generated search information, and extracts candidates of items to be presented to a user;

a context information generator that generates and outputs context information including user intention by performing semantic analysis on the natural language request;

a ranker that ranks the extracted candidates of items based on the generated context information, user information representing an attribute of the user, and history information representing item usage history of the user; and

a context tag dictionary that stores therein a plurality of context tags, each context tag of the plurality of context tags representing a context and stored in association with at least one word or phrase acting as a trigger for generating the each context tag, wherein

when a word or phrase extracted from the natural language request matches with a context tag stored in the context tag dictionary, the context information generator outputs the context tag as the generated context information,

when the word or phrase extracted from the natural language request matches with a word or phrase stored in the context tag dictionary, the context information generator outputs a corresponding context tag stored in the context tag dictionary, as the generated context information,

the history information includes the context information that is used for ranking, when an item has been extracted as a candidate, the item that has been used, and

the ranking unit

generates a first feature vector related to the user and the generated context information based on the attribute of the user and the context information included in the history information,

generates a second feature vector related to items based on attributes of the items,

calculates rating estimation values of the extracted candidates of items using an equation to which the first feature vector and the second feature vector are applied,

sorts the extracted candidates of items in a descending order of rating estimation value, and

outputs the sorted candidates of items, and

the plurality of context tags include tags related to scenes, tags related to accompanying persons, tags related to objectives, tags related to situations, and tags related to timeslots.

2. The device according to claim 1, further comprising a search tag dictionary that stores therein a plurality of search tags, each search tag of the plurality of search tags serving as a search keyword and stored in association with at least one word or phrase acting as a trigger for generating the each search tag, and an application destination attribute representing an attribute to which the search tag is applied, wherein

when a word or phrase extracted from the natural language request matches with a search tag stored in the search tag dictionary, the search information generator outputs the search tag and a corresponding application destination attribute stored in the search tag dictionary, as the generated search information, and

when the word or phrase extracted from the natural language request matches with a word or phrase stored in the search tag dictionary, the search information generator outputs a corresponding search tag and a corresponding application destination attribute stored in the search tag dictionary, as the generated search information.

3. The device according to claim 1, wherein the processing circuitry is further configured to function as:

a human relationship information generator that analyzes an image which is input, and generates and outputs human relationship information which represents human relationship in a group of persons included in the image; and

an integrating unit that integrates the human relationship information into the generated context information, wherein

the ranker ranks the extracted candidates of items based on the generated context information into which the human relationship information has been integrated, the user information, and the history information.

4. The device according to claim 3, wherein the human relationship information generator

detects persons from the image and estimates an attribute of each of the persons, and

collates a set of attributes of the persons detected from the image with a predetermined rule and generates the human relationship information.

5. The device according to claim 4, wherein

from among persons detected from the image, the human relationship information generator selects a person having a largest size in the image, or selects a person detected to be closest to a center of the image, or selects a person whose lip area is moving, and

the ranker uses the attribute of the person, who is selected by the human relationship information generator, as the user information.

6. An item recommendation method implemented in an item recommendation device, comprising:

generating search information by performing semantic analysis on a natural language request which is input;

performing searching with respect to a storing unit storing therein item information using the search information and extracting candidates of items to be presented to a user;

generating and outputting context information including user intention by performing semantic analysis on the natural language request;

ranking the extracted candidates of items based on the generated context information, user information representing an attribute of the user, and history information representing item usage history of the user;

storing in a context tag dictionary a plurality of context tags, each context tag of the plurality of context tags representing a context and stored in association with at least one word or phrase acting as a trigger for generating the each context tag;

when a word or phrase extracted from the natural language request matches with a context tag stored in the context tag dictionary, outputting the context tag as the generated context information; and

when the word or phrase extracted from the natural language request matches with a word or phrase stored in the context tag dictionary, outputting a corresponding context tag stored in the context tag dictionary, as the generated context information, wherein

the history information includes the context information that is used for ranking, when an item has been extracted as a candidate, the item that has been used, and further comprising

generating a first feature vector related to the user and the generated context information based on the attribute of the user and the context information included in the history information,

generating a second feature vector related to items based on attributes of the items,

calculating rating estimation values of the extracted candidates of items using an equation to which the first feature vector and the second feature vector are applied,

sorting the extracted candidates of items in a descending order of rating estimation value, and

outputting the sorted candidates of items, and

the plurality of context tags include tags related to scenes, tags related to accompanying persons, tags related to objectives, tags related to situations, and tags related to timeslots.

7. A computer program product comprising a non-transitory computer readable medium including programmed instructions for item recommendation, wherein the instructions, when executed by a computer, cause the computer to perform:

generating search information by performing semantic analysis on a natural language request which is input;

performing searching with respect to a storing unit storing therein item information using the generated search information and extracting candidates of items to be presented to a user;

generating and outputting context information including user intention by performing semantic analysis on the natural language request;

ranking the extracted candidates of items based on the generated context information, user information representing an attribute of the user, and history information representing item usage history of the user;

storing in a context tag dictionary a plurality of context tags, each context tag of the plurality of context tags representing a context and stored in association with at least one word or phrase acting as a trigger for generating the each context tag;

when a word or phrase extracted from the natural language request matches with a context tag stored in the context tag dictionary, outputting the context tag as the generated context information; and

when the word or phrase extracted from the natural language request matches with a word or phrase stored in the context tag dictionary, outputting a corresponding context tag stored in the context tag dictionary, as the generated context information, wherein

the history information includes the context information that is used for ranking, when an item has been extracted as a candidate, the item that has been used, and further comprising

generating a first feature vector related to the user and the generated context information based on the attribute of the user and the context information included in the history information,

generating a second feature vector related to items based on attributes of the items,

calculating rating estimation values of the extracted candidates of items using an equation to which the first feature vector and the second feature vector are applied,

sorting the extracted candidates of items in a descending order of rating estimation value, and

outputting the sorted candidates of items, and

the plurality of context tags include tags related to scenes, tags related to accompanying persons, tags related to objectives, tags related to situations, and tags related to timeslots.