Sentence Structure and Discourse Structure: Possible Parallels

The present contribution represents the first step in comparing the nature of syntactico-semantic relations present in the sentence structure to their equivalents in the discourse structure. The study is carried out on the basis of a Czech manually annotated material collected in the Prague Dependency Treebank (PDT). According to the semantic analysis of the underlying syntactic structure of a sentence (tectogrammatics) in the PDT, we distinguish various types of relations that can be expressed both within a single sentence (i.e. in a tree) and in a larger text, beyond the sentence boundary (between trees). We suggest that, on the one hand, each type of these relations preserves its semantic nature both within a sentence and in a larger text (i.e. a causal relation remains a causal relation) but, on the other hand, according to the semantic properties of the relations, their distribution in a sentence or between sentences is very diverse. In this study, this observation is analyzed for two cases (relations of condition and specification) and further supported by similar behaviour of the English data from the Penn Discourse Treebank. 1 Motivation and Background Although the annotation in the Prague Dependency Treebank 2.0 (PDT, Hajič et al., 2006; Mikulová et al., 2005) in principle does not surpass the sentence boundaries, i.e. each sentence is represented by a single dependency tree structure, to a certain extent, the information about the context has always been one of its concerns. First, the context of every sentence is reflected in one attribute of the nodes in the syntactico-semantic (tectogrammatical) structure: the information structure of the sentence (Topic-Focus Articulation, TFA, cf. Sgall, Hajičová and Panevová, 1986; Hajičová, Partee and Sgall, 1998), second, some basic coreference relations are marked (especially the grammatical coreference and some types of the textual coreference). In recent years, the interest in analyzing the structure of discourse in a more complex way has increased, and also the PDT is being enriched with this type of information. After having annotated the anaphoric chains and also the so-called bridging relations (or the association anaphora, see Nedoluzhko et al., 2009), the annotation of semantic relations between text spans indicated by certain discourse markers is now in progress. This annotation has two linguistic resources: besides the Prague (syntactico-semantic) approach it is inspired also by the Penn Discourse Treebank 2.0 approach based on identifying discourse connectives and their arguments (Prasad et. al, 2007 and 2008). One of the benefits of annotating discourse semantic relations on tectogrammatical trees is a possibility to exploit the syntactico-semantic information already captured in the corpus. This fact also enables us to compare the nature of relations expressed both within a single sentence (in a single tree) and in a larger text (between trees). Since the discourse annotation of the PDT is still a work in progress, it is premature to make some final conclusions in this respect. On the other hand, a majority of the corpus has already been processed and some tendencies are evident. In the present contribution we therefore want to introduce some observations about the nature of these corresponding relations and support them with our data. The contribution is divided into three main parts. In Section 2, we describe some basic aspects of the Praguian approach to the syntactic structure (tectogrammatics); criteria according to which some relations from the tectogrammatics are considered to be discourse relations are introduced in Section 3; and in Section 4 a comparison of intra-sentential and inter-sen-