Section: New Results

Modelling discourse-level information

Participants : Laurence Danlos, Timothée Bernard.

We have continued our work on the formalisation of discourse-level information. First, we have proposed in [24] a new model in STAG syntax and semantics for subordinate conjunctions (SubConjs) and attributing phrases —attitude/reporting verbs (AVs; believe, say) and attributing prepositional phrase (APPs; according to). This discourse-oriented model is based on the observation that SubConjs and AVs are not homogeneous categories. Indeed, previous work has shown that SubConjs can be divided into two classes according to their syntactic and semantic properties. Similarly, AVs have two different uses in discourse: evidential and intentional. While evidential AVs and APPs have strong semantic similarities, they do not appear in the same contexts when SubConjs are at play. Our proposition aims at representing these distinctions and capturing these various discourse-related interactions.

We have also investigated how sentential and discourse TAG-based grammars can be interfaced, in collaboration with Aleksandre Maskharashvili and Sylvain Pogodalla (LORIA). Tree-Adjoining Grammars (TAG) have been used both for syntactic parsing, with sentential grammars, and for discourse parsing, with discourse grammars (see for example our D-STAG model or the D-LTAG model). Yet the modelling of discourse connectives (coordinate conjunctions, subordinate conjunctions, adverbs...) in TAG-based formalisms for discourse differ from their modelling in sentential grammars. Because of this mismatch, an intermediate processing step is required between the sentential and the discourse processes, both in parsing and in generation [27]. We have developed a method to smoothly interface sentential and discourse TAG grammars, without using such an intermediate processing step. This method, based on Abstract Categorial Grammars (ACG), allows for building D-STAG discourse structures that are direct acyclic graphs (DAG) and not only trees.