Learning apparatus, information processing apparatus, learning method, information processing method and program
A learning apparatus executes receiving a text and a question associated with the text, and calculating an evidence score expressing a likelihood of a character string included in the text as evidence for an answer to the question by using a model parameter of a first neural network; extracting, by sampling from a predetermined distribution having the evidence score as a parameter, a first set indicating a set of the character strings as the evidence for the answer from the text; receiving the question and the first set and extracting the answer from the first set by using a model parameter of a second neural network; and learning the model parameters of the first and second neural networks by calculating a gradient through error back propagation by using a continuous relaxation and a first loss between the answer and a true answer to the question.
1 . A learning apparatus, comprising:
a memory; and
a processor configured to execute:
receiving a text and a question associated with the text as input, and calculating an evidence score expressing a likelihood of a character string included in the text as evidence for an answer to the question by using a model parameter of a first neural network;
extracting, by sampling from a predetermined distribution having the evidence score as a parameter, a first set indicating a set of the character strings as the evidence for the answer from the text;
receiving the question and the first set as input and extracting the answer from the first set by using a model parameter of a second neural network; and
learning the model parameter of the first neural network and the model parameter of the second neural network by calculating a gradient through error back propagation by using a continuous relaxation and a first loss between the answer and a true answer to the question.
2 . The learning apparatus according to claim 1 , wherein
the learning calculates the first loss by a loss function including a term concerning a penalty for the fact that information expressed by a character string extracted by the extracting does not include information mentioned by the question, and a term concerning a penalty for the fact that the answer is not included in the character string extracted by the extracting.
3 . The learning apparatus according to claim 1 , wherein the receiving of the text and the question receives the text, the question associated with the text, and a second set indicating a set of character strings as a correct evidence for the answer to the question as input, and the extracting extracts a character string included in the first set from the second set by using the model parameter of the first neural network, and
wherein the processor is further configured to execute:
learning the model parameter of the first neural network by using a second loss between the first set and the second set and learn the model parameter of the second neural network by using a third loss between the answer and the true answer,
wherein the learning of the model parameters by calculating the gradient is executed after the learning of the model parameter of the first neural network by using the second loss.
4 . A non-transitory computer-readable recording medium having computer-readable instructions stored thereon, which when executed, cause a computer to operate as the learning apparatus according to claim 1 .
5 . An information processing apparatus, comprising:
a memory; and
a processor configured to execute:
extracting, based on a text and a question about the text as input, by using a first model parameter of a first neural network, a set of character strings from the text as evidence for an answer to the question from the text; and
extracting, based on the evidence and the question about the text, the answer from the set of character strings by using a second model parameter of a second neural network, wherein the first model parameter of the first neural network and the second model parameter of the second neural network have been trained by calculating a gradient through error back propagation by using a continuous relaxation and a first loss between the answer and a true answer to the question.
6 . A non-transitory computer-readable recording medium having computer-readable instructions stored thereon, which when executed, cause a computer to operate as the information processing apparatus according to claim 5 .
7 . A learning method executed by a computer including a memory and a processor, the learning method comprising:
receiving a text and a question associated with the text as input, calculating an evidence score expressing a likelihood of a character string included in the text as evidence for an answer to the question by using a model parameter of a first neural network, and extracting, by sampling from a predetermined distribution having the evidence score as a parameter, a first set indicating a set of the character strings as the evidence for the answer from the text;
receiving the question and the first set as input and extract the answer from the first set by using a model parameter of a second neural network; and
learning the model parameter of the first neural network and the model parameter of the second neural network by calculating a gradient through error back propagation by using a continuous relaxation and a first loss between the answer and a true answer to the question.
8 . An information processing method executed by a computer including a memory and a processor, the information processing method comprising:
extracting, based on a text and a question about the text as input, by using a first model parameter of a first neural network, a set of character strings from the text as being evidence for an answer to the question from the text; and
extracting, based on the evidence and the question about the text, the answer from the set of character strings by using a second model parameter of a second neural network, wherein the first model parameter of the first neural network and the second model parameter of the second neural network have been trained by calculating a gradient through error back propagation by using a continuous relaxation and a first loss between the answer and a true answer to the question.