Back to the Roots of Genres: Text Classification by Language Function

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 632–640, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Back to the Roots of Genres: Text Classification by Language Function Henning Wachsmuth University of Paderborn, s-lab Paderborn, Germany hwachsmuth@s-lab.upb.de Kathrin Bujna University of Paderborn Paderborn, Germany kabu@mail.upb.de Abstract The term “genre” covers different aspects of both texts and documents, and it has led to many classification schemes. This makes different approaches to genre iden- tification incomparable and the task itself unclear. We introduce the linguistically motivated text classification task language function analysis, LFA, which focuses on one well-defined aspect of genres. The aim of LFA is to determine whether a text is predominantly expressive, appellative, or informative. LFA can be used in search and mining applications to efficiently fil- ter documents of interest. Our approach to LFA relies on fast machine learning classi- fiers with features from different research areas. We evaluate this approach on a new corpus with 4,806 product texts from two domains. Within one domain, we correctly classify up to 82% of the texts, but differ- ences in feature distribution limit accuracy on out-of-domain data. 1 Introduction Text classification has been successfully applied to various natural language processing tasks, among which some of the most popular are topic de- tection, authorship attribution, sentiment analysis, and genre identification. While the first three refer to single aspects of a text, genres cover different properties of both documents and texts, such as their form, function, purpose, and target audience. As a consequence, many different genre classifica- tion schemes exist, which makes most approaches to genre identification badly comparable as a re- cent study showed (Sharoff et al., 2010). Corre- spondingly, the question which features work best in genre identification still remains open. We ar- gue that one major reason behind is a missing common understanding of genres and that we need to focus on the single aspects of genres in order to overcome this situation. In this paper, we investigate why a text was writ- ten. Therefore, we introduce the ambitious task language function analysis, abbreviated as LFA, i.e., to classify the predominant function of a text as intended by its author. In line with the work of the psychologist Karl B¨uhler (1934), we dis- tinguish three abstract and very general classes, namely, expressive, appellative, and informative texts. Several search and text mining applications can benefit from applying LFA as a document fil- tering step with respect to both efficiency and ef- fectiveness. A search engine, for instance, might restrict its result list to hits that mainly serve an in- formative purpose. Similarly, LFA can be used in opinion mining to cancel out promotional texts in favor of personal attitudes. While, of course, LFA does not replace genre identification, we think that language functions constitute one root of a com- mon and clear genre concept. Language functions are well-studied in linguis- tic pragmatics, but we analyze whether they also correlate with statistical text characteristics. For this …

Content truncated to 500 words