Linguistic Profiling of Texts Across Textual Genres and Readability Levels. An Exploratory Study on Italian Fictional Prose

Proceedings of Recent Advances in Natural Language Processing, pages 189–197, Hissar, Bulgaria, 7-13 September 2013. Linguistic Profiling of Texts Across Textual Genres and Readability Levels. An Exploratory Study on Italian Fictional Prose Felice Dell’Orletta, Simonetta Montemagni, Giulia Venturi Istituto di Linguistica Computazionale “Antonio Zampolli” (ILC–CNR) ItaliaNLP Lab - www.italianlp.it Via G. Moruzzi, 1 – Pisa (Italy) {felice.dellorletta,simonetta.montemagni,giulia.venturi}@ilc.cnr.it Abstract In this paper we present a case study fo- cusing on the literature genre, in partic- ular on Italian fictional prose, aimed at identifying the features characterizing this text type. Identified features were tested in two classification tasks, i.e. by genre and by readability, with promising results. Interestingly, the same multi–level set of linguistic features turned out to reliably capture variation within and across textual genres. 1 Introduction Over the last ten years, Natural Language Pro- cessing (NLP) techniques combined with machine learning algorithms started being used to investi- gate the “form” of a text rather than its content. The range of tasks sharing this approach to the analysis of texts is wide, ranging e.g. from na- tive language identification (see among the oth- ers Koppel et al. (2005) and Wong and Dras (2009)), author recognition and verification (see e.g. van Halteren (2004), authorship attribution (see Juola (2008) for a survey), genre identifica- tion (Mehler et al., 2011) to readability assess- ment (see Dell’Orletta et al. (2011a) for an up- dated survey). Besides obvious differences at the level of selected linguistic features and learning techniques, which are also motivated by the lan- guage varieties targeted by the different tasks, they share a common approach: they succeed in deter- mining the language variety, the author, the text genre or the level of readability of a text by ex- ploiting the distribution of features automatically extracted from texts. The issues typically dealt with in this type of studies can be summarised in two main research questions aimed at investigat- ing 1) which linguistic features work best for a given task, and 2) which type of machine learning algorithms are best suited for a given task. In this paper, we focus on the first issue, i.e. on the typology of linguistic features which could be reliably extracted from automatically analysed texts with particular attention to the potential im- pact of achieved results on two classification tasks. In particular, we identified the set of linguistic fea- tures characterizing classes of documents, based on their textual genre or the type of audience they target: to put it in van Halteren words (van Hal- teren, 2004), we carried out “linguistic profiling” of texts selected as representative of different gen- res and/or readability levels. Achieved theoretical results were tested in two text classification tasks, aimed at classifying texts by genre or readabil- ity level. This goal was pursued in a case study focusing on the literature genre, in particular on Italian fictional prose. First, we studied variation within and across genres, by carrying out a con- trastive linguistic analysis a) of a corpus of litera- ture texts with respect …

Content truncated to 500 words