Intention estimation equipment and intention estimation system

• US 9,569,427 B2
• Filed: 12/15/2015
• Issued: 02/14/2017
• Est. Priority Date: 12/25/2014
• Status: Active Grant

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Clarion Co., Ltd. (Faurecia SA)
• Inventors: Obuchi, Yasunari, Shima, Kazuaki, Homma, Takeshi
• Primary Examiner(s): AZAD, ABUL K
• Application Number:
  US14/969,109
• Publication Number:
  US 20160188574A1
• Time in Patent Office: 427 Days
• Field of Search: 704 1- 10, 704231-257, 704/270, 704/275
• US Class Current: 1/1
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding

Abstract

An intention estimation equipment includes: a first training data group; a second training data group; a model creation unit that creates first and second statistical models that estimates an intention of an input text; an error data extraction unit that extracts, from the second training data group, training data corresponding to a text, of which an intention estimation result based on the first and the second statistical models is correct and erroneous, respectively, as error data; an opposite data extraction unit that extracts, from the second training data group, training data that is a cause for an intention estimation result of the error data based on the second statistical model becoming erroneous as opposite data; and a data correction unit that performs correction of the second training data group so that an influence of the error data or of the opposite data on creation of the statistical model is changed.

Claims

1. An intention estimation device comprising:

a processor; and

a memory connected to the processor, and storing a first training data group including a plurality of pieces of training data each being constituted by a text and a correct intention expression correlated with each other, and a second training data group including the plurality of pieces of training data in the first training data group and a plurality of pieces of training data not included in the first training data group,

wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute;

a model creation unit that creates a plurality of statistical models that, when text is input, estimate an intention of the input text using the training data and outputs an intention estimation result, and the statistical models include a first statistical model is created by the model creation unit using the first training data group and a second statistical model is created by the model creation unit using the second training data group,

an error data extraction unit that extracts training data from the second training data group corresponding to text of which an intention estimation result output based on the first statistical model is correct and an intention estimation result output based on the second statistical model is erroneous, as error data;

an opposite data extraction unit that extracts, from the second training data group, training data that is a cause for an intention estimation result of the error data output based on the second statistical model becoming erroneous as opposite data; and

a data correction unit that performs correction of the second training data group so that an influence of the error data or an influence of the opposite data on creation of the statistical models is changed.

2. The intention estimation device according to claim 1, wherein the training data further includes a priority indicating an index of importance of the text in intention estimation, and the data correction unit compares a priority of the error data with a priority of the opposite data and performs correction of the second training data group based on a result of the comparison.

3. The intention estimation device according to claim 2, wherein the data correction unit, when the priority of the error data is higher than the priority of the opposite data, performs correction of the second training data group to increase an influence of the error data or decrease an influence of the opposite data in creating the statistical model.

4. The intention estimation device according to claim 2, wherein the data correction unit, when the priority of the error data is lower than the priority of the opposite data, performs correction of the second training data group to decrease an influence of the error data or increase an influence of the opposite data in creating the statistical model.

5. The intention estimation device according to claim 2,

wherein the memory further stores priority basic information that includes information relating to an intention importance table indicating relations among a correct intention, an error intention estimation result, and a degree of disadvantage, which represents a magnitude of disadvantage a user receives when the error intention estimation result is output, and

wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute:

a priority setting unit that refers to the intention importance table and sets the priority based on the magnitude of the degree of disadvantage.

6. The intention estimation device according to claim 5, wherein the intention estimation result includes a function estimation result and a variable estimation result, the intention importance table includes a first intention importance table corresponding to an error of the function estimation result and a second intention importance table corresponding to an error of the variable estimation result, and

the priority setting unit, when the function estimation result is erroneous, refers to the first intention importance table or when variable estimation result is erroneous, refers to the second intention importance table, and sets the priority based on the magnitude of the degree of disadvantage.

7. The intention estimation device according to of claim 2, further comprising:

a communication unit that performs transmission/reception of information to/from a terminal that includes an input unit that detects an input based on a behavior of a user and receives, from the terminal, input information based on the input detected by the input unit; and

a priority setting unit that sets the priority based on a frequency in which a text is created based on input information received by the communication unit.

8. An intention estimation device according to claim 2, further comprising:

a communication unit that performs transmission/reception of information to/from a terminal including an input unit that detects an input based on a behavior of a user and receives, from the terminal, input information based on the input detected by the input unit, and

wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute;

a priority setting unit that determines the priority based on at least one of (A) a preset importance of the terminal, (B) a preset importance of the user, (C) a frequency at which the terminal transmits the information, (D) an elapsed time from last transmission of the information by the terminal, and (E) an elapsed time from initial transmission of the information.

9. The intention estimation device according to claim 1, wherein at least one of the first training data and the second training data further includes a multiplying factor that represents a weight of the text in creating the statistical models by the model creation unit, and the data correction unit performs correction of the second training data group by changing the multiplying factor of the error data and the opposite data.

10. The intention estimation device according to claim 9,

wherein the memory further stores a rule table in which a text and a correct intention correlated with the text are stored,

wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute:

an intention estimation unit that outputs an intention estimation result for an input text based on the first statistical model or the rule table,

wherein the data correction unit transfers training data that is included in the second training data group and has the multiplying factor outside a predetermined range to the rule table, and the intention estimation unit, when the input text is included in the rule table, outputs the intention estimation result based on the rule table without using the first statistical model.

11. The intention estimation device according to claim 1, wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute:

a model evaluation unit that judges whether the second statistical model satisfies predetermined criteria based on an intention estimation result obtained by inputting any text into the second statistical model,

wherein the data correction unit, when evaluation unit judges that the second statistical model satisfies the predetermined criteria, performs correction of the second training data group, and the model creation unit repeats creation of the second statistical model using the second training data group corrected by the data correction unit.

12. The intention estimation device according to claim 1, wherein the second training data group includes the first training data group and additional training data created based on information transmitted from a terminal that includes an input unit.

13. The intention estimation device according to claim 1, further comprising:

a communication unit that performs transmission/reception of information to/from a terminal that includes an input unit that detects an input based on a behavior of a user,

wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute;

an intention estimation unit that outputs an intention estimation result for the input text based on the first statistical model; and

a correct judgment unit that judges whether the intention estimation result output by the intention estimation unit is true or false,

wherein the communication unit transmits the intention estimation result output from the intention estimation unit to the terminal and receives from the terminal input information based on an input detected by the input unit within a predetermined period of time from receipt of the intention estimation result, and the correct judgment unit judges whether the intention estimation result output by the intention estimation unit based on the input information received by the communication unit is true or false.

14. The intention estimation device according to claim 1, wherein the memory further stores instructions that, when executed by the processor, causes to processor to execute:

a notification unit that notifies a text of which an intention estimation result output based on the first statistical model is erroneous and an intention estimation result output based on the second statistical model is correct.

15. An intention estimation system comprising:

an intention estimation device connected to a network and which includes;

a processor; and

a memory connected to the processor, and storing a first training data group including a plurality of pieces of training data each being constituted by a text and a correct intention expression correlated with each other, and a second training data group including the plurality of pieces of training data in the first training data group and a plurality of pieces of training data not included in the first training data group, and

the memory further stores instructions that, when executed by the processor, causes to processor to execute;

a model creation unit that creates a plurality of statistical models that, when text is input, estimate an intention of the input text using the training data and outputs an intention estimation result, and the statistical models include a first statistical model is created by the model creation unit using the first training data group and a second statistical model is created by the model creation unit using the second training data group,

an error data extraction unit that extracts training data from the second training data group corresponding to text of which an intention estimation result output based on the first statistical model is correct and an intention estimation result output based on the second statistical model is erroneous, as error data;

an opposite data extraction unit that extracts, from the second training data group, training data that is a cause for an intention estimation result of the error data output based on the second statistical model becoming erroneous as opposite data; and

a data correction unit that performs correction of the second training data group so that an influence of the error data or an influence of the opposite data on creation of the statistical models is changed; and

a terminal connectable to the intention estimation device via the network,

wherein the terminal includes;

an input unit that receives an input by a user;

a transmission unit that transmits the input that is input to the input unit by the user to the intention estimation device; and

an execution unit that receives an intention estimation result on the input by the user performed by the intention estimation device and acts according to the intention estimation result.