Text generation apparatus, text generation learning apparatus, text generation method, text generation learning method and program
A text generation apparatus includes a memory and a processor configured to, based on learned parameters of neural networks, acquire, as a reference text, a predetermined number of two or more sentences having a relatively high relevance to an input sentence from a set of sentences different from the input sentence and generate text based on the input sentence and the reference text, such that information to be considered when generating text can be added as text.
1 . A text generation apparatus comprising:
a memory; and
a processor configured to execute:
receive an input sentence from a source, wherein the input sentence is in textual form;
acquire a reference text in textual form as extracted from an externally connected knowledge source database according to first learned parameters of a first trained neural network, wherein the externally connected knowledge source database is distinct from the source, the reference text includes a predetermined number of two or more sentences in textual form, respective sentences in the reference text are distinct from the input sentence, and the first trained neural network generates a relevance score associated with a respective sentence of the reference text based on vectorized words associated with the input sentence, and the relevance score indicates semantic relevance between a respective sentence of the externally connected knowledge source database and the input sentence;
identify, based on the relevance score, one or more reference sentences of the reference text predicted as relevant to the input sentence;
identify, based on another relevance score, one or more parts of the input sentence for summarization; and
generate, using the identified one or more reference sentences predicted as relevant to the identified one or more parts of the input sentence, an output sentence that is a summary of the input sentence according to second learned parameters of a second trained neural network, wherein the second trained neural network generates the output sentence by summarizing a combination including the identified one or more parts of the input sentence and the identified one or more reference sentences using vectorized words associated with the input sentence, and the output sentence as a summary of the input sentence comprises a word at least in part obtained from one or more words of the identified one or more parts of the input sentence and another word at least in part obtained from at least a part of the identified one or more reference sentences.
2 . The text generation apparatus according to claim 1 , wherein the processor is configured to calculate a relevance measure indicating a degree of the relevance for each sentence included in the set or for each word of the sentence and select a sentence to be included in the reference text based on the relevance measure.
3 . The text generation apparatus according to claim 2 , wherein the processor is configured to, when generating text based on the reference text, weight an attention probability for each word included in the reference text by using the relevance measure.
4 . A text generation learning apparatus comprising:
a memory; and
a processor configured to, based on learned parameters of neural networks:
receive an input sentence from a source, wherein the input sentence is in text form;
acquire a reference text in textual form from as extracted an externally connected knowledge source database according to first learned parameters of a first trained neural network,
wherein the externally connected knowledge source database is distinct from the source,
the reference text includes a predetermined number of two or more sentences in textual form, respective sentences in the reference text are distinct from the input sentence, and
the first trained neural network generates a relevance score associated with a respective sentence of the reference text based on vectorized words associated with the input sentence, and the relevance score indicates semantic relevance between a respective sentence of the externally connected knowledge source database and the input sentence;
identify, based on the relevance score, one or more reference sentences of the reference text predicted as relevant to the input sentence;
identify, based on another relevance score, one or more parts of the input sentence for summarization; and
generate, using the identified one or more reference sentences predicted as relevant to the identified one or more parts of the input sentence, an output sentence that is a summary of the input sentence according to second learned parameters of a second trained neural network, wherein the second trained neural network generates the output sentence by summarizing a combination including the identified one or more parts of the input sentence and the identified one or more reference sentences using vectorized words associated with the input sentence, the output sentence as a summary of the input sentence comprises a word at least in part obtained from one or more words of the identified one or more parts of the input sentence and another word at least in part obtained from at least a part of the identified one or more reference sentences, and
wherein the processor is further configured to learn the parameters.
5 . A computer-implemented text generation method, performed based on learned parameters of neural networks comprising operations to:
receive an input sentence from a source, wherein the input sentence is in textual form;
acquire a reference text in textual form as extracted from an externally connected knowledge source database according to first learned parameters of a first trained neural network, wherein the externally connected knowledge source database is distinct from the source, the reference text includes a predetermined number of two or more sentences in textual form, respective sentences in the reference text are distinct from an input sentence, and the first trained neural network generates a relevance score associated with a respective sentence of the reference text based on vectorized words associated with the input sentence, and the relevance score indicates semantic relevance between a respective sentence of the externally connected knowledge source database and the input sentence;
identify, based on the relevance score, one or more reference sentences of the reference text predicted as relevant to the input sentence;
identify, based on another relevance score, one or more parts of the input sentence for summarization; and
generate, using the identified one or more reference sentences predicted as relevant to the input sentence, based on the identified one or more parts of the input sentence and the reference text, an output sentence that is a summary of the input sentence according to second learned parameters of a second trained neural network, wherein the second trained neural network generates the output sentence by summarizing a combination including the identified one or more parts of the input sentence and the identified one or more reference sentences using vectorized words associated with the input sentence, and the output sentence as a summary of the input sentence comprises a word at least in part obtained from one or more words of the identified one or more parts of the input sentence and another word at least in part obtained from at least a part of the identified one or more reference sentences.
6 . A non-transitory computer-readable recording medium having a program stored thereon causing a computer to execute the text generation method of claim 5 .