Segmentation and Clustering of Textual Sequences: a Typological Approach

Proceedings of Recent Advances in Natural Language Processing, pages 427–433, Hissar, Bulgaria, 12-14 September 2011. Segmentation and Clustering of Textual Sequences: a Typological Approach Christelle Cocco, Rapha¨el Pittier, Franc¸ois Bavaud and Aris Xanthos University of Lausanne, Switzerland {Christelle.Cocco,Raphael.Pittier, Francois.Bavaud,Aris.Xanthos}@unil.ch Abstract The long term goal of this research is to develop a program able to produce an automatic segmentation and categoriza- tion of textual sequences into discourse types. In this preliminary contribution, we present the construction of an algo- rithm which takes a segmented text as in- put and attempts to produce a categoriza- tion of sequences, such as narrative, argu- mentative, descriptive and so on. Also, this work aims at investigating a possible convergence between the typological ap- proach developed in particular in the field of text and discourse analysis in French by Adam (2008) and Bronckart (1997) and unsupervised statistical learning. 1 Introduction An increasing amount of research has been conducted concerning text genre detection us- ing POS (part-of-speech) tags since the work of Biber (1988). For instance, Malrieu and Rastier (2001) describe how to classify texts ac- cording to genres (comedy, tragedy, drama...) or discourses (literary, legal, political. . . ) using POS- tags. POS-tags can be determined in an unsupervised way (see e.g. Schmid (1994)) and their distri- bution happens to differ according to types of texts, such as narrative, explicative and so on. Hence, developing automatic discourse type de- tection, which is of interest to the linguistic com- munity, seems practicable. Thus, the purpose of the present study is to clus- ter clauses of a text into discourse types, i.e. to develop a tool for type detection with a limited quantity of annotated texts. We limit ourselves to the use of simple bag-of-words models on which fuzzy and K-means clustering are applied. Specifically, the aim is twofold: firstly, the con- struction of a program which takes a segmented text as input and produces a categorization of se- quences of clauses by clustering, based princi- pally on POS-tags; secondly, the comparison of this clustering with the typology proposed by a human expert, corresponding to discourse types. Thus, this preliminary work aims at investigat- ing a possible convergence between unsupervised statistical learning on the one hand, and the ty- pological approach developed in particular in the field of French linguistics by Adam (2008) and in language psychology by Bronckart (1997) on the other hand. As a first step, sample texts were manually an- notated, that is segmented (section 2.1) and classi- fied (section 2.2). Then, the clauses resulting from the previous segmentation were clustered on the basis of their POS distribution (sections 2.3 and 2.4). It appeared that the latter vary across the ty- pological classes proposed by the expert (section 3.1) which were compared to those resulting from fuzzy and K-means clustering processes (sections 3.2 and 4). Future developments are proposed in section 5. 2 Method 2.1 Segmentation The first step of this research was to create a cor- pus of annotated texts. For that purpose, a …

Content truncated to 500 words