(Based on work done with Gosse Bouma [Rijksuniversiteit Groningen] and Rob Malouf [Stanford University])

In the tradition of G/HPSG, extraction has been treated via inheritance of values for the feature SLASH from gap, up along the 'extraction path' to the filler. This analysis is attractive because it makes information about extractions accessible for local selection. One could view it as a prediction of such an analysis that there are languages registering extraction through verb morphology (Chamorro, Palauan), the possibility of verb-inversion (French, Icelandic), tone downstep suppression (Kikuyu) etc. However, under past proposals, not all verbs in an extraction domain are 'slashed'. Gazdar (81) and Pollard and Sag (94) agree that verbs whose subject has been extracted are unslashed, and in P&S-94 adv-initial clauses are unslashed as well. This distribution of SLASH specifications is inconsistent with the results of Hukari and Levine (95), who show that languages registering extraction effects also show them in subject and adverb extraction.

This paper develops a new extraction analysis that avoids this inadequacy. Lexemes specify a verb's dependency structure, but not its valence or SLASH value. Every dependent is realized as a SUBJ, COMPS, and/or SLASH member. A word's SLASH value must also be the union of the SLASH values of its dependents. Thus a verb's object can either be a nongap element on its COMPS list or else a gap in the verb's SLASH value. In either case, the verb passes its SLASH value (empty in the former case) to its mother. Gaps may be both in SLASH and the SUBJ value, allowing us to correctly predict the behavior of the languages discussed by H&L. Likewise, since extractable adjuncts are also included in the DEPENDENTS list, it follows that a verb will be [SLASH {ADV}] just in case an adverbial modifying it is extracted. Hence we predict H&L's observation that adverb extraction is lexically registered.

Our account not only provides better cross-linguistic coverage than earlier G/HPSG treatments, it also accounts for subject, complement, and adverbial extraction via a single constraint on the relation between lexemes and words. In this presentation, I will summarize a number of further predictions, including a number that follow directly from the elimination of WH-trace, the 'subinding' that occurs in both English and French, and the that-trace alternation and its analogue (que-qui alternations) in French.