Summary generating apparatus, summary generating method and computer program

A summary generating method of generating a summary of text using a computer having a processor and memory, the method comprising the computer performing the steps of:

• storing in the memory the text with information indicating a portion to be focused on in the text in a text storage device;
• analyzing the text to determine the portion to be focused on in the text including:
• receiving target text of the text;
• performing morphological analysis on the received target text;
• performing dependency analysis of the target text based on the result of the morphological analysis;
• identifying a problem description portion of the target text based on the dependency analysis of the target text;
• specifying a predicate template of the problem description portion
• identifying content words in the target text as candidate important portions;
• for each candidate important portion:
• determining if the candidate important portion is positioned before a predicate in the problem description portion;
• when the candidate important portion is positioned before the predicate in the problem description portion:
• generating word vector sequences including word vectors from the candidate important portion by a first method;
• inputting the generated word vector sequences into a first method neural network trained to identify important portions;
• for each of the word vectors, adding an element indicating whether or not the word vector is the portion to be focused on based on whether the first method neural network identifies the word vector as an important portion;
• when the candidate important portion is not positioned before the predicate in the problem description portion:
• generating the word vector sequences including the word vectors from the candidate important portion by a second method;
• inputting the generated word vector sequences into a second method neural network trained to identify the important portions;
• for each of the word vectors, adding the element indicating whether or not the word vector is the portion to be focused on based on whether the second method neural network identifies the word vector as an important portion;
• performing sequence-to-sequence type conversion by it a converting neural network trained beforehand by machine learning such that it outputs, in response to input of each of the word vectors of the word vector sequence in a prescribed order, a summary of the text consisting of words represented by the word vector sequence;
• and
• inputting each of the word vectors of the converted word vector sequence to the converting neural network in the prescribed order, and obtaining a word sequence output by the converting neural network in response to the input, as a summary of the text.