Defining DTG derivations to get semantic graphs

Tue aim of this paper is to find a fonnalism of the TAG family, where the derivation controller can be interpreted as a semantic dependency graph, in the sense of Meaning-Text Theory (ZM67; M88). In a previous paper (CK98), we study tliis interpretation of the derivation tree (DT) in the case of standard TAG. We prove that, in the general case, if the predieate-argument cooccurence principle1 [= P ACP] holds and if elementary trees correspond to a semantic unit (A91), substitution arcs can be read as · semantic dependencies where the dependent is the anchor of the substituted tree, and adjunction arcs of any typecan be read as semantic dependencies in the opposite direction. Yet we also characterized cases where the DT shows wrong (semantic) dependencies (cf also (RVW95)). A problem may occur when, in the same sentence, clausal complementation is handled both with substitution of an embedded clause and with adjunction of a main verb.2 Further, there are well-known cases that TAG cannot handle if the PACP holds (e.g. clitic climbing in Romance (B98), Kashmiri wh-extraction (RVW95), extraction out of NP in French (A98). Finally, in some cases, the argumental positions in a tree are not filled by the right arguments, and thus the derivation tree does not show the right semantic dependencies (pied-piping (CK98)).