Information processing device, information processing method, and program

1. An information processing device, comprising:

a processing circuitry configured to receive as input a document and a question, and to execute processing to output an answer range as a range of a string that can be an answer to the question in the document, the range being defined with position numbers of start and end words in the document, or an answer suitability of the document with respect to the question, by using neural networks, the answer suitability indicating a degree of a document as being answerable to the question,

wherein the processing circuitry includes a first neural network configured to calculate the answer range, and a second neural network configured to calculate the answer suitability, and between the first neural network and the second neural network, both neural networks share a plurality of layers that process attributes of the inputted document and question, while only the first neural network includes a reading comprehension layer as a final layer that outputs the answer range and only the second neural network includes an information retrieval layer as a final layer that outputs the answer suitability.

2. The information processing device as claimed in claim 1 , wherein the processing circuitry obtains, for each document in an input group of documents, the answer range and the answer suitability of the document with respect to the question.

3. The information processing device as claimed in claim 1 ,

wherein the processing circuitry obtains the answer range and the answer suitability for each document in an input group of documents, to obtain a string as an answer candidate from the answer range, and

wherein the processing circuitry is further configured, based on answer suitabilities for documents having a same string extracted as the answer candidate, to calculate an answer candidate score for the string obtained for said each document.

4. The information processing device as claimed in claim 3 , wherein the processing circuitry calculates the answer candidate score by a sum total of the answer suitabilities, a weighted mean of the answer suitabilities, or a weighted sum of the answer suitabilities for the documents having the same string extracted as the answer candidate.

5. The information processing device as claimed in claim 1 , wherein the processing circuitry generates a matching matrix based on the document and the question, by the shared part of the layers constituting said both neural networks, and

calculates the answer range by the first neural network, and calculates the answer suitability by the second neural network, based on the matching matrix.

6. An information processing device, comprising:

a processing circuitry configured to receive as input a document and a question, and to execute processing to output an answer range as a range of a string that can be an answer to the question from the document, the range being defined with position numbers of start and end words in the document, or an answer suitability with respect to the answer range, by using neural networks, the answer suitability indicating a degree of a document as being answerable to the question,

wherein the processing circuitry includes a first neural network configured to calculate the answer range, and a second neural network configured to calculate the answer suitability, and between the first neural network and the second neural network, both neural networks share a plurality of layers that process attributes of the inputted document and question, while only the first neural network includes a reading comprehension layer as a final layer that outputs the answer range and only the second neural network includes an information retrieval layer as a final layer that outputs the answer suitability,

wherein the processing circuitry is further configured to learn parameters of said both neural networks from outputs from said both neural networks and training data as right answers, and

wherein the training data as the right answers is constituted with tuples of question, document, answer range, and answer suitability.

7. The information processing device as claimed in claim 1 , wherein the first neural network is constituted with a plurality of neural networks each of which corresponds to one of a plurality of ranges as the answer range indicating answers at a plurality of levels of granularity, and between the second neural network and the plurality of neural networks, part of the layers constituting these neural networks is shared.

8. The information processing device as claimed in claim 7 , wherein the processing circuitry is further configured to output a range having a most appropriate level of granularity as the answer range, based on the plurality of ranges indicating the answers having the plurality of levels of granularity calculated by the plurality of neural networks.

9. The information processing device as claimed in claim 8 , wherein the range having the most appropriate level of granularity corresponds to one of a phrase, a paragraph, and a document.

10. A non-transitory computer-readable recording medium having a program stored thereon for causing a computer to function as the processing circuitry in the information processing device as claimed in claim 1 .

11. An information processing method executed by an information processing device including a processing circuitry configured to receive as input a pair of a document and a question, and to execute processing to output an answer range as a range of a string that can be an answer to the question in the document, the range being defined with position numbers of start and end words in the document, or an answer suitability of the document with respect to the question, the answer suitability indicating a degree of a document as being answerable to the question,

wherein the processing circuitry includes a first neural network configured to calculate the answer range, and a second neural network configured to calculate the answer suitability, and between the first neural network and the second neural network, both neural networks share a plurality of layers that process attributes of the inputted document and question, while only the first neural network includes a reading comprehension layer as a final layer that outputs the answer range and only the second neural network includes an information retrieval layer as a final layer that outputs the answer suitability,

the method comprising:

a step executed by the processing circuitry for generating a matching matrix based on the document and the question, by the shared part of the layers constituting said both neural networks; and

a step executed by the processing circuitry for calculating the answer range by the first neural network based on the matching matrix, or calculating the answer suitability by the second neural network based on the matching matrix.

12. The information processing method as claimed in claim 11 ,

wherein the first neural network is constituted with a plurality of neural networks each of which corresponds to one of a plurality of ranges as the answer range indicating answers at a plurality of levels of granularity, and between the second neural network and the plurality of neural networks, part of the layers constituting these neural networks is shared.

13. A non-transitory computer-readable recording medium having a program stored thereon for causing a computer to function as the processing circuitry in the information processing device as claimed in claim 6 .