Behind the Times: Detecting Epoch Changes using Large Corpora

International Joint Conference on Natural Language Processing, pages 347–355, Nagoya, Japan, 14-18 October 2013. Behind the Times: Detecting Epoch Changes using Large Corpora Octavian Popescu and Carlo Strapparava FBK-irst, Trento, Italy {popescu,strappa}@fbk.eu Abstract Using large corpora of chronologically or- dered language, it is possible to explore diachronic phenomena, identifying previ- ously unknown correlations between lan- guage usage and time periods, or epochs. We focused on a statistical approach to epoch delimitation and introduced the task of epoch characterization. We investi- gated the significant changes in the distri- bution of terms in the Google N-gram cor- pus and their relationships with emotion words. The results show that the method is reliable and the task is feasible. 1 Introduction Traditionally, scholars of history define epochs ac- cording to their deep knowledge and understand- ing of facts over a long stretch of time. Intuitively, in order to define a new epoch, both a big social impact of a series of events and new issues, which arouse the social interest, must be observed. How- ever, it is hard to define what makes a feature “dis- tinctive” or an event a “great change”. It is even harder to evaluate and measure the impact of a series of changes in society in an objective way. Since the advent of regular newspapers and the in- dustry of mass media, written information has rep- resented a mirror of the interests of society. A so- cial event is relevant only if people pay attention to it and comment on it. A major change in so- ciety is reflected in the frequencies with which a set of topics is mentioned in mass media, some of them becoming mentioned more often than previ- ously, while some others are no more of interest. Furthermore, specific epochs typically develop a particular form of wording or rhetorical style. In this paper we describe a computational ap- proach to epoch delimitation on the basis of word distribution over certain periods of time. A big quantity of data, chronologically ordered, allows accurate statistical statements regarding the co- variance between the frequencies of two or more terms over a certain period of time. By discov- ering significant statistical changes in word usage behavior, it is possible to define epoch boundaries. We show that it is possible to distinguish a se- ries of limited periods of time, spanning at most three years, within which non-random changes af- fect the joint distribution of terms. Between two such short periods (i.e. the boundaries) no statis- tical significant changes are observed for decades, and thus we can refer to it as an epoch. The dis- tributions of the considered terms before and after boundaries are distinctly different. We also introduce the task of epoch character- ization. Certain words carry with them an emo- tional charge, like joy, fear, disgust etc. Within a given epoch, we can analyze the distribution of emotion words and their co-occurrences with the set of terms considered indicative for epoch def- inition. The pattern of these co-occurrences …

Content truncated to 500 words