Sentences that are paraphrases of each other convey essentially the same meaning, often in essentially the same words, but in different constructions. A major problem is to provide for automatic recognition of sameness of mean¬ing in spite of structural differences so that when a re¬quest for factual information is processed, any of the possible paraphrases of an answer will be retrieved from a natural language text. How this problem could be resolved by automated parsing grammars has not been clear because such grammars are primarily concerned with making structural differences explicit, and also because the grammars in current use are greatly in need of refinement.nIn an experimental way, the RAND parsing grammar has been extended and refined so that identical labels are attached to comparable elements of structurally different paraphrases, such as the active-passive pair, "X commands the third fleet" and "The third fleet is commanded by X." The sentences are first parsed automatically;then the output of the PARSE routine is fed into a subroutine called FLAG, which attaches labels that make the sameness of meaning explicit. The labels are attached by using coded information derived from the rules applied in the course of parsing the sentences. A sample of the output from the PARSE and FLAG routines (Fig. 2, p. 26) shows how the differences in structure are resolved automatically and how the labels are attached that make possible the auto-matic retrieval of both sentences in response to a single relevant request. It is shown in detail how interrogative sentences also may be recognized as paraphrases of a single request and how the output from their processing can be used as input to a retrieval search.nMaking the RAND grammar more adequate for fact re¬trieval has also improved it for its primary task of parsing, since establishing parastructural relationships in order to be able to recognize paraphrases resulted in more numerous and more precise syntactic classes for dif¬ferentiating among structures. A plan is set forth for further extension of the grammar, using a code-matching technique for ease in handling the numerous, complexly interrelated classes that appear to be necessary. An extended grammar, using the PARSE and FLAG routines, would label a complicated sentence so that it could be retrieved as a response to many different questions for which it is a relevant answer. The fact that a single sentence may answer many different questions indicates that natural language sentences are often highly efficient information storage devices.
展开▼