A Meta-Rule Treatment for Englich Wh-Constructions

This paper describes a general metarule treatment of English whconstructions (relative clauses, and questions ) in the context of a broad -coverage logic grammar that also includes an extensive metarule treatment of co-ordinate conjunction . ' Vh-constructions pose difficulties for parsing , due to their introduction of a dependency between the wh-word (e.g., which ) and a corresponding gap in the following clause: This is the book which I thought you told me to refer to ( ) . The gap can be arbitrarily far away from the wh-word , but it must occur within the clause, or tIle sentence is not well -formed , as in *The book whicll I read it . A meta -rule treatment has several advantages over an Extraposition Grammar -style treatment : a natural delimitation of the gap scope, the ability to translate / compile the grammar rules , and ease of integration with conjunction . \ Vh-constructions are handled by annotating those grammar rules that license a gap or realize a gap . These annotations are converted , via the metarule component , into parameterized rules . A set of paired input / output parameters pass the need for a gap from parent to child and left sibling to right sibling until tIle gap is realized ; once the gap is realized , the parameter takes on a no_gap value , preventing further gaps from being realized . Tilis 'change of state ' in the paired parameters ensures that each gap is filled exactly once. The conjunction metarule operates on the parameterized wh-rules to link gaps within conjoined structures by unification , so tllat any gap within a conjoined structure is treated identically for all conjuncts . I This work has been supported in part by DARPA under contract NOOO14 -85-C-OO12 , administered by the Office of Naval Research; and in part by internal U nisys funding . Wh -constructions are one of the classically difficult parsing problems , because a correct treatment requires interaction of non-adjacent constituents , namely the wh-word, which introduces a constituent in clause-initial position , and the following construction which is missing a constituent (the gap). The gap can be arbitrarily far from the introducing wh-word (an unbounded dependency ) ; in particular , it can appear within deeply embedded constructions, such as the person that [1 had hoped [Jane would tell [( ) to get the books])] , where there are three levels of embedded structure . It is possible , in principle , to write a rule for each case where a gap can appear . I Iowever , since the number of constructions which can accommodate a gap is very large (e.g., most complement types), this is both extremely labor -intensive and unmaintainable from the grammar writer 's point of view . It is also possible to write general rules for gap-realization , e.g., a noun phrase can be realized as a gap . If this approach is taken , then these rules must be carefully constrained to accept gaps only when inside a wh-construction ; in addition , the wh-construction must contain exactly one gap . These restrictions involve complex and expensive search up and down the parse tree , to determine whether a gap is occurring inside a wh-construction . In many ways, the wh-problem parallels the problem of coordinate conjunction that has also been a major obstacle for natural language systems . Both constructions involve gaps, both affect large portions of the grammar , and both require a major modification to the grammar and/ or to the parsing mechanism to handle the linguistic phenomena . There have been two basic approach es to conjunction and wh-constructions in the computational linguistics literature : modification of the parser (interpreter ) and meta -rules . Of these, the first approach has been far more common . For conjunction , a number of variants on the ' interrupt ' driven approach have been presented, both in conventional natural language processing systems [13, 12, 14] , and in the context of logic grammars [4] . The same is true for logic grammar implementations of wh-constructions : the most generally used treatment is the interpreter-based treatment of Extraposition Grammar (XG) [10] . Meta -rules offer an appealing alternative to interpreter -based approach es, both for conjunction and for wh-expressions . Metarules are particularly well -suited to phenomena that range over a variety of syntactic structures , where the linguistic description would otherwise require regular changes to a large set of grammar rules . The use of metarules turns out to be efficient computationally . It also preserves compactness of the underlying grammar , so that the grammar is still maintainable from the point of view of the grammar -writer . Finally , the meta-rule approach avoids additional interpretive overhead and permits translation / compilation of grammar rules for efficient execution [5] . For conjunction , the meta -rule approach forms the basis for a comprehensive Hirschman 2