The term "Big Data" has become very popular in recent years, and this concept has also had a significant influence on the study of languages. By utilizing vast archives of digital text and newly developed analysis methods, linguists now have more options to do research. In particular, we can cross-check several sources at the same time, and in greater detail than was previously possible. The digital humanities are a rapidly growing research field all around the world. But even with the help of computers, humans are still central to such studies, and we rely on expert opinion to determine the key elements for data analysis. Perhaps this is the main reason why text mining has only developed relatively slowly, and one of the main difficulties encountered in conducting text mining is the data input. For computing purposes, texts are a form of unstructured data which first needs to be quantified, and since, so far, there has been no Standard Operating Procedure (SOP) for the quantification process, experts play an important role in selecting relevant information (i.e. variables) for data analysis. In this study, our goal is to establish an SOP for text mining, and then to use it to study the historical change in Chinese writing style, from classical Chinese to modern Chinese, which occurred about a century ago. The study material is Volumes 1-7 of New Youth (新青年), by the end of which, modern vernacular Chinese had almost completely replaced classical Chinese. We adapt the idea of unsupervised learning from statistical learning theory to define and identify important variables. In particular, we use the approach of Exploratory Data Analysis (EDA) to evaluate potential variables as candidates for differentiating between language styles. Also, following a previous study of ours, numerous variable and data reduction methods are needed. Thus, we use principle component analysis to reduce the number of variables, and then apply classification methods, such as logistic regression, to judge whether the style of an article is closer to classical or to modern Chinese. Also, to avoid over-parameterization (i.e. using more variables than are necessary), we use cross-validation to select the most feasible model. This cross-validation separates the data into a training set (or in-sample) and a testing set (or out-sample); the training set is used to construct the model and this model is then applied to the testing set to calculate the model's accuracy. Our study shows a gradual change in the writing style of New Youth articles from classical Chinese to modern Chinese. Thus, only 1% of the articles in Volume 1 are classified as modern Chinese compared with 98% of those in Volume 7, and about 60% of the articles in an intermediate volume, number 4, are classified as modern Chinese. Our model has a prediction accuracy for the articles in Volume 4 of about 84%, as determined by cross-checking with expert Chinese linguists. The results of our quantitative numerical analysis are clearly promising, suggesting that this approach should be continued for the study of modern Chinese writing. [ABSTRACT FROM AUTHOR]