Granted Patent B2

US 12,572,580 · App. 18/269,180 · Granted Mar 10, 2026

Inventors: Kosuke Akimoto (Tokyo, JP); Kazeto Yamamoto (Tokyo, JP); Kunihiko Sadamasa (Tokyo, JP)

Assignee: NEC CORPORATION

G06F16/353G06F16/3347G06F40/12G06F40/289G06N3/044

Abstract

Claims (76)

1 . An information processing apparatus, comprising at least one processor, the at least one processor carrying out:

a relation vector generation process of generating a relation vector that represents a relation between a plurality of entities of interest from at least one relation vector generation sentence which has been selected from a sentence set and in which the plurality of entities of interest occur;

a feature vector generation process of generating, for each of the plurality of entities of interest, a feature vector that represents a feature of that entity of interest from at least one feature vector generation sentence which has been selected from the sentence set and in which that entity of interest occurs; and

a relation categorization process of categorizing a relation between the plurality of entities of interest with use of a relation vector which has been generated in the relation vector generation process and feature vectors which have been generated in the feature vector generation process,

wherein:

in the feature vector generation process, the at least one processor generates, for each of the plurality of entities of interest, the feature vector from the at least one feature vector generation sentence with use of an algorithm that includes a plurality of parameters; and

the at least one processor further carries out a feature vector generation parameter updating process of updating the plurality of parameters such that a degree of similarity increases between a feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest, and

wherein:

in the feature vector generation process, the at least one processor carries out, for each of the plurality of entities of interest and for each feature vector generation sentence,

(1) a second word vector sequence generation process of generating a word vector sequence that corresponds to that feature vector generation sentence by replacing a word representing that entity of interest with a predetermined word vector and replacing a word representing an entity other than that entity of interest with a word vector representing the word,

(2) a second Recurrent Neural Network (RNN) output vector sequence generation process of generating an Recurrent Neural Network (RNN) output vector sequence that corresponds to that feature vector generation sentence by inputting, into a recurrent neural network, the word vector sequence which has been generated in the second word vector sequence generation process, the recurrent neural network using a word vector sequence as input, and vectors corresponding to respective word vectors constituting that word vector sequence as output,

(3) a sentence feature vector setting process of setting, from among vectors constituting the RNN output vector sequence generated in the second Recurrent Neural Network (RNN) output vector sequence generation process, a vector corresponding to that entity of interest to be a sentence feature vector that corresponds to that feature vector generation sentence, and

(4) a feature vector calculation process of calculating the feature vector by averaging, for each element, sentence feature vectors which have been set in the sentence feature vector setting process and which correspond to respective feature vector generation sentences; and

in the feature vector generation parameter updating process, the at least one processor updates a parameter of the recurrent neural network such that a degree of similarity increases between a sentence feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest.

2 . The information processing apparatus according to claim 1 , wherein:

in the relation vector generation process, the at least one processor generates the relation vector from the at least one relation vector generation sentence with use of an algorithm that includes a plurality of parameters; and

the at least one processor further carries out a relation vector generation parameter updating process of updating the plurality of parameters such that a degree of similarity increases between a plurality of relation vectors which are generated in the relation vector generation process from a plurality of sentences in all of which the plurality of entities of interest occur.

3 . The information processing apparatus according to claim 2 , wherein:

in the relation vector generation process, the at least one processor carries out, for each relation vector generation sentence,

(1) a first word vector sequence generation process of generating a word vector sequence that corresponds to that relation vector generation sentence by replacing a word representing each of the plurality of entities of interest with a predetermined word vector and by replacing a word representing an entity other than the plurality of entities of interest with a word vector representing that word,

(2) a first Recurrent Neural Network (RNN) output vector sequence generation process of generating an Recurrent Neural Network (RNN) output vector sequence that corresponds to that relation vector generation sentence by inputting, into a recurrent neural network, the word vector sequence which has been generated in the first word vector sequence generation process, the recurrent neural network using a word vector sequence as input, and vectors corresponding to respective word vectors constituting the word vector sequence as output,

(3) a sentence relation vector calculation process of calculating a sentence relation vector that corresponds to that relation vector generation sentence by averaging, for each element, vectors constituting the Recurrent Neural Network (RNN) output vector sequence which has been generated in the first Recurrent Neural Network (RNN) output vector sequence generation process, and

(4) a relation vector calculation process of calculating the relation vector by averaging, for each element, sentence relation vectors which have been calculated in the sentence relation vector calculation process and which correspond to respective relation vector generation sentences; and

in the relation vector generation parameter updating process, the at least one processor updates a parameter of the recurrent neural network such that a degree of similarity increases between a plurality of relation vectors that are generated in the relation vector generation process from a plurality of sentences in all of which a plurality of entities occur.

4 . The information processing apparatus according to claim 1 , wherein:

in the relation categorization process, the at least one processor carries out

a relation vector categorization process of categorizing a relation vector which has been generated in the relation vector generation process,

a feature vector categorization process of categorizing each of feature vectors which have been generated in the feature vector generation process, and

a categorization result combining process of categorizing a relation between the plurality of entities of interest by combining a categorization result obtained in the relation vector categorization process and a categorization result obtained in the feature vector categorization process together.

5 . The information processing apparatus according to claim 4 , wherein:

the at least one processor further carries out a relation vector clustering process of generating a cluster of relation vectors;

the at least one processor further carries out a feature vector clustering process of generating a cluster of feature vectors;

in the relation categorization process, the at least one processor carries out the relation vector categorization process by determining a cluster to which a relation vector generated in the relation vector generation process belongs, the cluster having been generated in the relation vector clustering process; and

in the relation categorization process, the at least one processor carries out the feature vector categorization process by determining a cluster to which each of feature vectors generated in the feature vector generation process belongs, the cluster having been generated in the feature vector clustering process.

6 . An information processing method, comprising:

generating a relation vector that represents a relation between a plurality of entities of interest from at least one relation vector generation sentence which has been selected from a sentence set and in which the plurality of entities of interest occur;

generating, for each of the plurality of entities of interest, a feature vector that represents a feature of that entity of interest from at least one feature vector generation sentence which has been selected from the sentence set and in which that entity of interest occurs; and

categorizing a relation between the plurality of entities of interest with use of the relation vector and the feature vectors which have been respectively generated for the plurality of entities of interest,

wherein:

in the feature vector generation process, generating, for each of the plurality of entities of interest, the feature vector from the at least one feature vector generation sentence with use of an algorithm that includes a plurality of parameters; and

carrying out a feature vector generation parameter updating process of updating the plurality of parameters such that a degree of similarity increases between a feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest, and

wherein:

in the feature vector generation process, carrying out, for each of the plurality of entities of interest and for each feature vector generation sentence,

(1) a second word vector sequence generation process of generating a word vector sequence that corresponds to that feature vector generation sentence by replacing a word representing that entity of interest with a predetermined word vector and replacing a word representing an entity other than that entity of interest with a word vector representing the word,

(2) a second Recurrent Neural Network (RNN) output vector sequence generation process of generating an Recurrent Neural Network (RNN) output vector sequence that corresponds to that feature vector generation sentence by inputting, into a recurrent neural network, the word vector sequence which has been generated in the second word vector sequence generation process, the recurrent neural network using a word vector sequence as input, and vectors corresponding to respective word vectors constituting that word vector sequence as output,

(3) a sentence feature vector setting process of setting, from among vectors constituting the RNN output vector sequence generated in the second Recurrent Neural Network (RNN) output vector sequence generation process, a vector corresponding to that entity of interest to be a sentence feature vector that corresponds to that feature vector generation sentence, and

(4) a feature vector calculation process of calculating the feature vector by averaging, for each element, sentence feature vectors which have been set in the sentence feature vector setting process and which correspond to respective feature vector generation sentences; and

in the feature vector generation parameter updating process, updating a parameter of the recurrent neural network such that a degree of similarity increases between a sentence feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest.

7 . A non-transitory storage medium storing a program for causing a computer to function as an information processing apparatus, the program causing the computer to carry out:

generating a relation vector that represents a relation between a plurality of entities of interest from at least one relation vector generation sentence which has been selected from a sentence set and in which the plurality of entities of interest occur;

generating, for each of the plurality of entities of interest, a feature vector that represents a feature of that entity of interest from at least one feature vector generation sentence which has been selected from the sentence set and in which that entity of interest occurs; and

categorizing a relation between the plurality of entities of interest with use of the relation vector and the feature vectors which have been respectively generated for the plurality of entities of interest,

wherein:

in the feature vector generation process, generating, for each of the plurality of entities of interest, the feature vector from the at least one feature vector generation sentence with use of an algorithm that includes a plurality of parameters; and

carrying out a feature vector generation parameter updating process of updating the plurality of parameters such that a degree of similarity increases between a feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest, and

wherein:

in the feature vector generation process, carrying out, for each of the plurality of entities of interest and for each feature vector generation sentence,

(1) a second word vector sequence generation process of generating a word vector sequence that corresponds to that feature vector generation sentence by replacing a word representing that entity of interest with a predetermined word vector and replacing a word representing an entity other than that entity of interest with a word vector representing the word,

(2) a second Recurrent Neural Network (RNN) output vector sequence generation process of generating an Recurrent Neural Network (RNN) output vector sequence that corresponds to that feature vector generation sentence by inputting, into a recurrent neural network, the word vector sequence which has been generated in the second word vector sequence generation process, the recurrent neural network using a word vector sequence as input, and vectors corresponding to respective word vectors constituting that word vector sequence as output,

(3) a sentence feature vector setting process of setting, from among vectors constituting the RNN output vector sequence generated in the second Recurrent Neural Network RNN) output vector sequence generation process, a vector corresponding to that entity of interest to be a sentence feature vector that corresponds to that feature vector generation sentence, and

(4) a feature vector calculation process of calculating the feature vector by averaging, for each element, sentence feature vectors which have been set in the sentence feature vector setting process and which correspond to respective feature vector generation sentences; and

in the feature vector generation parameter updating process, updating a parameter of the recurrent neural network such that a degree of similarity increases between a sentence feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest.

8 . An information processing apparatus, comprising at least one processor, the at least one processor carrying out:

a relation vector generation process of generating a relation vector with use of an algorithm including a plurality of parameters from at least one relation vector generation sentence which has been selected from a sentence set and in which a plurality of entities of interest occur, the relation vector representing a relation between the plurality of entities of interest;

a relation vector generation parameter updating process of updating the plurality of parameters such that a degree of similarity increases between a plurality of relation vectors that are generated in the relation vector generation process from a plurality of sentences in all of which the plurality of entities of interest occur; and

a relation categorization process of categorizing a relation between the plurality of entities of interest with use of a relation vector which has been generated in the relation vector generation process,

wherein:

in a feature vector generation process, the at least one processor generates, for each of the plurality of entities of interest, the feature vector from the at least one feature vector generation sentence with use of an algorithm that includes a plurality of parameters; and

the at least one processor further carries out a feature vector generation parameter updating process of updating the plurality of parameters such that a degree of similarity increases between a feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest, and

wherein:

in the feature vector generation process, the at least one processor carries out, for each of the plurality of entities of interest and for each feature vector generation sentence,

(1) a second word vector sequence generation process of generating a word vector sequence that corresponds to that feature vector generation sentence by replacing a word representing that entity of interest with a predetermined word vector and replacing a word representing an entity other than that entity of interest with a word vector representing the word,

(2) a second Recurrent Neural Network (RNN) output vector sequence generation process of generating an Recurrent Neural Network (RNN) output vector sequence that corresponds to that feature vector generation sentence by inputting, into a recurrent neural network, the word vector sequence which has been generated in the second word vector sequence generation process, the recurrent neural network using a word vector sequence as input, and vectors corresponding to respective word vectors constituting that word vector sequence as output,

(3) a sentence feature vector setting process of setting, from among vectors constituting the RNN output vector sequence generated in the second Recurrent Neural Network (RNN) output vector sequence generation process, a vector corresponding to that entity of interest to be a sentence feature vector that corresponds to that feature vector generation sentence, and

(4) a feature vector calculation process of calculating the feature vector by averaging, for each element, sentence feature vectors which have been set in the sentence feature vector setting process and which correspond to respective feature vector generation sentences; and

in the feature vector generation parameter updating process, the at least one processor updates a parameter of the recurrent neural network such that a degree of similarity increases between a sentence feature vector that is generated in the feature vector generation process from a sentence in which an entity of interest occurs and a word vector that represents the entity of interest.

Assignments (1)

Continuity (1)

References Cited (37)