An alternative description of extractions in TAG

The aim of the paper is to propose a new description of extraction in plain TAG. Contrary to Kroch 1987’s analysis, our description is based on the fact that the power of a relative clause to adjoin on a noun can be attached to the wh-word rather than to a verb. This analysis solves some problems of the previous analysis, notably by giving the right semantic dependency in case of pied-piping. We are thankful to our two reviewers for many valuable comments. Introduction The only description of extractions in TAG we know has been developed by Kroch & Joshi (1986), Kroch (1987) and implemented in the developed grammars of English (XTAG 1995) and French (Abeillé 1991, Candito 1999). This implementation solves the unboundedness of extractions with predicative adjoining, but the pied-piping is solved using a special feature. We think that this solution of pied-piping is not absolutely convenient, because some edge of the derivation tree cannot be interpreted as semantic dependency (Candito & Kahane 1998). Our assumption is based on the fact that a TAG derivation tree can be interpreted as a semantic graph, that is a predicate-argument structure. Moreover this implementation fails to describe some cases of extraction, such as some French dont-relatives. We propose a new description of extraction in TAG which solve most of these problems. Nevertheless, our study must rather be appreciated as an investigation of the limits of the TAG formalism, because we think that TAG is not the most appropriate framework for the implementation of our description of extractions. The same analysis is more suitably implemented in GAG/DTG (Candito & Kahane 1998). 1. Semantic dependencies The meaning of a sentence comes from the combination of the meaning of the lexical units of the sentence. A lexical meaning or semanteme can be considered as a semantic functor or predicate. For instance, consider: (1) Peter often saw black cats. In (1), the meaning ‘see’ is a binary functor whose argument are ‘Peter’ and ‘cat’, whereas ‘often’ and ‘black’ are unary functors with respectively ‘see’ and ‘cat’ as arguments. This predicateargument structure can be represented by a graph (Fig. 1), called a semantic graph (Zolkovski & Mel’çuk 1967, Mel’çuk 1988). An edge of such a graph is called a semantic dependency. The two extremities of a semantic dependency are called the semantic governor and the semantic argument. A semantic graph can be converted into a logical formula by reification : for each semanteme a variable is introduced as first argument of the predicate; this variable is used by other predicates pointing on it in the semantic graph. The semantic graph of Fig. 1 is thus converted in the formula: ‘Peter’(x) & ‘cat’(y) & ‘black’(p,y) & ‘see’(e,x,y) & ‘often‘(q,e) 116 Sylvain Kahane, Marie-Hélène Candito, Yannick de Kercadio