Learning system, learning method, and recording medium

1 . A learning system comprising:

at least one memory that is configured to store instructions; and

at least one processor that is configured to execute the instructions to:

obtain document data;

generate a key phrase from the document data;

restore the document data from the generated key phrase;

calculate a degree of similarity between the document data and the restored document data by using a language model that outputs a probability for an inputted word string;

calculate a first evaluation value on the basis of the calculated degree of similarity, the first evaluation value being a value indicating a degree of restoration of the document data from the key phrase;

learn parameters of a generation unit which generates the key phrase on the basis of the document data and the restored document data;

learn the parameters on the basis of a first evaluation value calculated by a degree of similarity between the document data and the restored document data;

update parameters of the generation unit to maximize the first evaluation value; and

repeatedly perform the generating, the restoring, the calculating, and the updating until a predetermined termination condition is satisfied.

2 . The learning system according to claim 1 , wherein: the at least one processor is configured to execute the instructions to calculate the degree of similarity by using a predetermined language model, and the first evaluation value is a value indicating a degree of restoration of the document data from the key phrase.

3 . The learning system according to claim 1 , wherein the at least one processor is configured to execute the instructions to learn parameters of a restoration unit which restores the document data from the key phrase on the basis of the document data and the restored document data.

4 . The learning system according to claim 1 , wherein the at least one processor is configured to execute the instructions to obtain first document data that do not include the key phrase and second document data that include a predetermined key phrase.

5 . The learning system according to claim 4 , wherein the at least one processor is configured to execute the instructions to:

calculate the first evaluation value on the basis of a degree of similarity between the document data and the restored document data, for the first document data, and

calculate a second evaluation value on the basis of a degree of matching between the key phrase generated by the generation unit and the predetermined key phrase, for the second document data.

6 . The learning system according to claim 1 , wherein the at least one processor is configured to execute the instructions to obtain the parameters of a generation unit which generates the key phrase, in addition to the document data.

7 . A learning method, performed by at least one processor and comprising:

obtaining document data;

generating a key phrase from the document data by using a generation unit;

restoring the document data from the generated key phrase;

calculating a degree of similarity between the document data and the restored document data by using a language model that outputs a probability for an inputted word string;

calculating a first evaluation value on the basis of the calculated degree of similarity, the first evaluation value being a value indicating a degree of restoration of the document data from the key phrase;

learning parameters of a generation unit which generates the key phrase on the basis of the document data and the restored document data;

learning the parameters on the basis of a first evaluation value calculated by a degree of similarity between the document data and the restored document data;

updating parameters of the generation unit to maximize the first evaluation value; and

repeatedly performing the generating, the restoring, the calculating, and the updating until a predetermined termination condition is satisfied.

8 . A non-transitory recording medium on which a computer program that when executed by at least one processor, causes the at least one processor to:

obtain document data;

generate a key phrase from the document data by using a generation unit;

restore the document data from the generated key phrase;

calculate a degree of similarity between the document data and the restored document data by using a language model that outputs a probability for an inputted word string;

calculate a first evaluation value on the basis of the calculated degree of similarity, the first evaluation value being a value indicating a degree of restoration of the document data from the key phrase;

learn parameters of a generation unit which generates the key phrase on the basis of the document data and the restored document data;

learn the parameters on the basis of a first evaluation value calculated by a degree of similarity between the document data and the restored document data;

update parameters of the generation unit to maximize the first evaluation value; and

repeatedly perform the generating, the restoring, the calculating, and the updating until a predetermined termination condition is satisfied.