Text-based relation extraction method and device

1 . A method for relation extraction (RE) from a text, the method comprising:

predicting, by a first relation prediction network, first relation information included in an entity pair of an input text based on a first label set for pre-defined relation information;

predicting, by a second relation prediction network, second relation information included in the entity pair of the input text based on a second label set including a labeling bias for the pre-defined relation information;

estimating, by at least one difference estimation network for a disagreement probability distribution of relation information, difference information between the first relation information and the second relation information; and

training the first relation prediction network, the second relation prediction network, and the at least one difference estimation network by using a loss function based on the difference information.

2 . The method of claim 1 , wherein the entity pair of the input text corresponds to a combination of a head entity and a tail entity, and

wherein the method further includes determining a word corresponding to the head entity and a word corresponding to the tail entity among a series of words of the input text by executing a feature encoder.

3 . The method of claim 2 , wherein the head entity and the tail entity are expressed as feature vectors of the words corresponding thereto, respectively.

4 . The method of claim 1 , wherein the first label set includes a human annotated label, and

the second label set includes a distantly supervised label.

5 . The method of claim 1 , wherein the first relation prediction network and the second relation prediction network are configured to share an initial weight with each other.

6 . The method of claim 1 , wherein the loss function includes a loss term for the difference information, and

the loss term for the difference information is a function of a ratio of the first relation information and the second relation information, and the difference information.

7 . The method of claim 1 , wherein the training includes determining the loss function based on a type of label for the relation information included in the entity pair of the input text.

8 . The method of claim 1 , wherein the disagreement probability distribution is a log normal distribution for a ratio of the first relation information and the second relation information, and

the at least one difference estimation network includes a first difference estimation network for an average of the log normal distribution and a second difference estimation network for a standard deviation of the log normal distribution.

9 . The method of claim 1 , further comprising:

estimating, by the first relation prediction network trained by the training, relation information included in the entity pair of a second input text.

10 . A computer readable non-transitory recording medium storing a computer program including at least one command configured to allow a computer to execute a method for relation extraction from a text of claim 1 .

11 . An apparatus for relation extraction from a text, the apparatus comprising:

a processor; and

a memory operably connected to the processor and storing at least one code performed by the processor,

wherein when the memory is executed through the processor, the memory stores a code which allows the processor to

predict, by a first relation prediction network, first relation information included in an entity pair of an input text based on a first label set for pre-defined relation information,

predict, by a second relation prediction network, second relation information included in the entity pair of the input text based on a second label set including a labeling bias for the pre-defined relation information,

estimate, by at least one difference estimation network for a disagreement probability distribution of relation information, difference information between the first relation information and the second relation information, and

train the first relation prediction network, the second relation prediction network, and the at least one difference estimation network by using a loss function based on the difference information.

12 . The apparatus of claim 11 , wherein the entity pair of the input text corresponds to a combination of a head entity and a tail entity, and

when the memory is executed through the processor, the memory stores a code configured to allow the processor to determine a word corresponding to the head entity and a word corresponding to the tail entity among a series of words of the input text by executing a feature encoder.

13 . The apparatus of claim 11 , wherein the first label set includes a human annotated label, and

the second label set includes a distantly supervised label.

14 . The apparatus of claim 11 , wherein the loss function includes a loss term for the difference information, and

the loss term for the difference information is a function of a ratio of the first relation information and the second relation information, and the difference information.

15 . The apparatus of claim 11 , wherein when the memory is executed through the processor, the memory stores a code configured to allow the processor to determine the loss function based on a type of label for relation information included in the entity pair of the input text.

16 . The apparatus of claim 11 , wherein the disagreement probability distribution is a log normal distribution for a ratio of the first relation information and the second relation information, and

the at least one difference estimation network includes a first difference estimation network for an average of the log normal distribution and a second difference estimation network for a standard deviation of the log normal distribution.

17 . The apparatus of claim 12 , wherein when the memory is executed through the processor, the memory stores a code configured to allow the processor to estimate the relation information included in the entity pair of a second input text by the trained first relation prediction network.