Word semantic relation estimation device and word semantic relation estimation method

1. A word semantic relation estimation device, comprising:

a processing circuitry

to search, as a word searcher, a database for each of words included in a word pair input to the word semantic relation estimation device;

to extract, as a word vector extractor, a word vector corresponding to each of the words searched by the word searcher from the database;

to convert characters included in the word pair into a character vector;

to acquire, as a feature amount acquirer, a feature amount by combining the word vector of each of the words and the character vector;

to estimate, as an estimator, a semantic relation between the words from the feature amount acquired by the feature amount acquiring unit; and

to switch, as an estimator switcher, on a basis of whether the word searcher retrieved each of the words from the database, between a first estimator estimating the semantic relation between the words using the word vector and the character vector as the feature amount and a second estimator estimating the semantic relation between the words using the character vector alone as the feature amount,

wherein the feature amount acquirer acquires the feature amount by combining the word vector and the character vector in a case where the word searcher retrieved each of the words from the database, and acquires the character vector as the feature amount in a case where the word searcher cannot retrieve each of the words from the database, and

the estimator estimates the semantic relation between the words from the feature amount acquired by the feature amount acquirer by using one of the first estimator and the second estimator depending on the switching of the estimator switcher.

2. The word semantic relation estimation device according to claim 1 ,

wherein the processing circuitry searches, as an additional searcher, the database for a word that partially matches characters included in an unknown word included in the word pair in a case where the word searcher cannot retrieve each of the words from the database,

wherein the word vector extractor searches the database for a word vector corresponding to the word that partially matches the characters included in the unknown word searched by the additional searcher.

3. The word semantic relation estimation device according to claim 1 ,

wherein the processing circuitry searches, as an additional searcher, the database for a word that partially matches characters included in an unknown word included in the word pair in a case where the word searcher cannot retrieve each of the words from the database,

wherein the estimator switcher switches between the first estimator and the second estimator on a basis of whether the additional searcher retrieved a word that partially matches characters included in the unknown word from the database, and

the feature amount acquirer acquires a feature amount obtained by combining the word vector and the character vector in a case where the additional searcher retrieved a word that partially matches characters included in the unknown word from the database and, in a case where no word that partially matches a character included in the unknown word is retrieved from the database, acquires the character vector as a feature amount.

4. The word semantic relation estimation device according to claim 1 , wherein the estimator estimates a degree of extraction indicating a vertical distance in a hierarchical relation between the words, for which the semantic relation is estimated, on a basis of structure information of training data of an estimator that estimates the semantic relation between the words.

5. A word semantic relation estimation method comprising:

searching, by a word searcher, a database for each of words included in a word pair which is input;

extracting, by a word vector extractor, a word vector corresponding to each of the words searched by the word searcher from the database,

converting, by a character vector converter, characters included in the word pair into a character vector;

acquiring, by a feature amount acquirer, a feature amount by combining the word vector of each of the words extracted by the word vector extractor and the character vector generated by converting the characters by the character vector converter;

estimating, by an estimator, a semantic relation between the words from the feature amount acquired by the feature amount acquirer; and

switching, by an estimator switcher, on a basis of whether the word searcher retrieved each of the words from the database, between a first estimator estimating the semantic relation between the words using the word vector and the character vector as the feature amount and a second estimator estimating the semantic relation between the words using the character vector alone as the feature amount,

wherein the feature amount acquirer acquires the feature amount by combining the word vector and the character vector in a case where the word searcher retrieved each of the words from the database, and acquires the character vector as the feature amount in a case where the word searcher cannot retrieve each of the words from the database, and

the estimator estimates the semantic relation between the words from the feature amount acquired by the feature amount acquired by using one of the first estimator and the second estimator depending on the switching of the estimator switcher.