Information processing apparatus, information processing method, and computer readable recording medium
An information processing apparatus ( 10 ) includes: a formal language query accepting unit ( 12 ) that accepts a query expression and correct answer data; a semi-structured data accepting unit ( 14 ) that accepts semi-structured data that includes text nodes; a node text extraction unit ( 16 ) that extracts natural language text from the text node, as node text; a node text expression generation unit ( 18 ) that receives the node text from the a converter ( 100 ) and obtains node text expressions; an answer calculation unit ( 20 ) that calculates an answer to the query expression with use of the node text expressions; and an update unit ( 22 ) that, if the answer calculated by the answer calculation unit ( 20 ) matches the correct answer data, updates parameters in the converter ( 100 ) such that the corresponding node text expression is more likely to be output in the converter ( 100 ).
1. An information processing apparatus that trains, through machine learning, a converter that receives an input of natural language text and outputs a formal language expression based on a parameter, the information processing apparatus comprising:
a formal language query accepting circuit configured to accept:
a query expression described in a formal language, and
correct answer data indicating a suitable answer to the query expression;
a semi-structured data accepting circuit configured to accept an input of semi-structured data, the semi-structured data including:
a text node including natural language text that has a semantic structure, and
a label associated with the text node;
a node text extraction circuit configured to extract the natural language text as node text from the text node of the semi-structured data accepted by the semi-structured data accepting circuit;
a node text expression generation circuit configured to input the node text extracted by the node text extraction circuit to the converter and obtain a node text expression from the converter, the node text expression being a formal language expression;
a correction circuit configured to correct the node text expression by combining the node text expression with a predicate corresponding to the label associated with the text node;
an answer calculation circuit configured to calculate an answer to the query expression accepted by the formal language query accepting circuit, the calculation based on the node text expression obtained by the node text expression generation circuit and corrected by the correction circuit; and
an update circuit configured to update the parameter of the converter such that the node text expression obtained by the node text expression generation circuit is more likely to be output in the converter, in a case where the answer calculated by the answer calculation circuit matches the correct answer data accepted by the formal language query accepting circuit.
2. The information processing apparatus according to claim 1 ,
wherein the semi-structured data can be expressed in a graph structure in which a plurality of nodes including the text node are associated with each other.
3. The information processing apparatus according to claim 1 , further comprising:
a natural language query accepting circuit configured to accept a query described in natural language and the correct answer data that indicates a suitable answer to the query; and
a query expression generation circuit configured to input the query accepted by the natural language query accepting circuit to the converter and obtain a query expression described in the formal language from the converter,
wherein the update circuit updates the parameter of the converter such that the query expression obtained by the query expression generation circuit is more likely to be output in the converter, in a case where the answer calculated by the answer calculation circuit matches the correct answer data.
4. The information processing apparatus according to claim 3 ,
wherein the converter includes a first converter that outputs the node text expression and a second converter that outputs the query expression.
5. The information processing apparatus according to claim 4 ,
wherein the first converter and the second converter include the same parameter.
6. The information processing apparatus according to claim 4 ,
wherein the parameter includes a first parameter used by the first converter and a second parameter used by the second converter.
7. The information processing apparatus according to claim 1 ,
wherein the parameter of the converter is a weight associated with a feature of a pair including natural language text input to the converter and a formal language expression output by the converter, and
the update circuit updates the parameter so as to increase a weight associated with a feature of a pair that includes the node text and the node text expression, in a case where the answer calculated by the answer calculation circuit matches the correct answer data.
8. The information processing apparatus according to claim 1 , further comprising:
an output circuit configured to output the parameter of the converter that was updated by the update circuit.
9. An information processing method for training, through machine learning, a converter that receives an input of natural language text and outputs a formal language expression based on a parameter, the information processing method comprising:
accepting a query expression described in a formal language;
accepting correct answer data indicating a suitable answer to the query expression;
accepting an input of semi-structured data, the semi-structured data including:
a text node including natural language text that has a semantic structure, and
a label associated with the text node;
extracting the natural language text as node text from the text node of the semi-structured data;
inputting the node text to the converter and obtaining a node text expression from the converter, the node text expression being a formal language expression;
correcting the node text expression by combining the node text expression with a predicate corresponding to the label associated with the text node;
calculating an answer to the query expression based on the corrected node text expression; and
updating the parameter of the converter such that the node text expression is more likely to be output in the converter, in a case where the answer calculated to the query expression matches the correct answer data.
10. The information processing method according to claim 9 ,
wherein the semi-structured data can be expressed in a graph structure in which a plurality of nodes including the text node are associated with each other.
11. The information processing method according to claim 9 , further comprising:
accepting a query described in natural language; and
inputting the query to the converter and obtaining a query expression described in the formal language from the converter,
wherein when the query expression is accepted, the query expression obtained from the converter is accepted, and
when the parameter of the converter is updated, the parameter of the converter is updated such that the query expression obtained from the converter is more likely to be output in the converter, in a case where the answer calculated to the query expression matches the correct answer data.
12. The information processing method according to claim 11 ,
wherein the converter includes a first converter that outputs the node text expression and a second converter that outputs the query expression.
13. The information processing method according to claim 12 ,
wherein the first converter and the second converter include the same parameter.
14. The information processing method according to claim 12 ,
wherein the parameter includes a first parameter used by the first converter and a second parameter used by the second converter.
15. The information processing method according to claim 9 ,
wherein the parameter of the converter is a weight associated with a feature of a pair including natural language text input to the converter and a formal language expression output by the converter, and
when the parameter of the converter is updated, the parameter is updated so as to increase a weight associated with a feature of a pair that includes the node text and the node text expression, in a case where the answer calculated to the query expression matches the correct answer data.
16. The information processing method according to claim 9 , further comprising:
outputting the parameter of the converter after the parameter of the converter is updated.
17. A non-transitory computer readable recording medium that includes a program recorded thereon, the program causing a computer to train, through machine learning, a converter that receives an input of natural language text and outputs a formal language expression based on a parameter, and the program including instructions that causes a computer to:
accept a query expression described in a formal language;
accept correct answer data indicating a suitable answer to the query expression;
accept an input of semi-structured data, the semi-structured data including: a text node which includes natural language text that has a semantic structure, and
a label associated with the text node;
extract the natural language text as node text from the text node of the semi-structured data;
input the node text to the converter and obtaining a node text expression from the converter, the node text expression being a formal language expression;
correct the node text expression by combining the node text expression with a predicate corresponding to the label associated with the text node;
calculate an answer to the query expression based on the corrected node text expression; and
update the parameter of the converter such that the node text expression is more likely to be output in the converter, in a case where the answer calculated to the query expression matches the correct answer data.