In this study, we applied a machine learning method to model the sentence-final expressions in modern Japanese novels, and examined what kind of items exhibit remarkable fluctuations. Our experiments were conducted on a diachronic corpus consisting of contemporary novels published from 1910 to 2014. Sentence-final expression data were extracted from the corpus and used for analysis. First, we investigated changes in the diversity of sentence-final expressions using the index of vocabulary richness. Then, we constructed modeling related to sentence-final expressions, and extracted the instrumental items playing important roles in model construction. The results showed that the sentence-final expressions became more diverse with time. Furthermore, statistical modeling revealed characteristic tendencies suggesting changes in the narrative methods and techniques of novels.

View full abstract

In this study, we use Balanced Corpus of Contemporary Written Japanese with annotations of predicate-argument structures to quantify word order freedom, case marking richness, and the correlation between the two in six text genres. The results of the study clarified the following three findings. Firstly, it was found that the word order freedom tended to increase in the order of white papers, Yahoo! Blogs, Yahoo! Chiebukuro, newspapers, magazines and books. Secondly, it was clarified that the case marking richness increases in the order of Yahoo! Blogs, Yahoo! Chiebukuro, newspapers, magazines, books and white papers. Thirdly, it was revealed that a positive correlation was found between word order freedom and case marking richness in five text genres other than white papers. This research is the first to quantitatively analyze the relationship between word order freedom and case marking richness in Japanese. The results of this study will be useful for a deeper understanding of Japanese in terms of the relationship between syntax and morphology, and in terms of language universals.

View full abstract

Quantitative analysis of language is often stated using the difference of the proportion calculated from frequency of various survey items such as words, part of speech, etymological types, character types, co-occurrence words, and the number of people who answered "yes", by corpus surveys and questionnaire surveys. However, most corpus surveys and questionnaire surveys are sample surveys, not complete surveys. The proportion obtained from the sample survey should not be treated the same as the proportion of the population. Also, even when comparing proportions, the sample size must be stated. The ratio of 10/50 and 20/50 and the ratio of 4/20 and 8/20 are the same for the former 20% and the latter 40%, and the difference in ratio is 20 points, but in the latter case, it is not a statistically meaningful difference. In this paper, while showing some concrete examples together with the calculation formulas, we describe the points to be noted about whether there is a difference in the proportion. The statistical explanation is kept to a minimum, and it is shown by a calculation formula that can be understood by knowledge of four arithmetic operations and square roots.

View full abstract