Automatic Detection and Analysis of Impressive Japanese Sentences Using Supervised Machine Learning

摘要

It is important to write sentences that impress the listener or reader ("impressive sentences") in many cases, such as when drafting political speeches. The study reported here provides useful information for writing such sentences in Japanese. Impressive sentences in Japanese are collected and examined for characteristic words. A number of such words are identified that often appear in impressive sentences, including jinsei (human life), hitobito (people), koufuku (happiness), yujou (friendliness), seishun (youth), and ren 'ai (love). Sentences using these words are likely to impress the listener or reader. Machine learning (SVM) is also used to automatically extract impressive sentences. It is found that the use of machine learning enables impressive sentences to be extracted from a large amount of Web documents with higher precision than that obtained with a baseline method, which extracts all sentences as impressive sentences.