A Tool for Automated Revision of Grammars for NLP Systems

A Tool for Automated Revision of Grammars for NLP Systems Nanda Kambhatla and Wlodek Zadrozny IBM T.J. Watson Research Center 30 Saw Mill River Road, Hawthorne, NY, 10532 {nanda, wlodz} @us.ibm.com Abstract We present an algorithm and a tool for automatically revising grammars for natural language processing (NLP) systems to disallow specifically identified sentences or sets of sentences. We also outline an approach for automatically revising attribute value grammars using counter-examples. Developing grammars for NLP systems that are both general enough to accept most sentences about a domain, but constrained enough to disallow other sentences is very tedious. Our approach of revising grammars automatically using counter-examples greatly simplifies the development and revision of tightly constrained grammars. We have successfully used our tool to constrain over-generalizing grammars of speech understanding systems and obtained higher recognition accuracy. 1 Introduction Natural language processing systems often constrain the set of "utterances" from a user (spoken, typed in, etc.) to narrow down the possible syntactic and semantic resolutions of the utterance and reduce the number of misrecognitions and/or misunderstandings by the system. Such constraints on the allowed syntax and the inferred semantics are often expressed in the form of a "grammar "l, a set of Throughout this document, by using the word "grammar", we refer to a Context-Free Grammar that consists of a finite set of non-terminals, a finite set of terminals, a unique non-terminal called the start symbol, and a set of production rules of the form A-> a, where A is a non-terminal and a is a string of terminal or non-terminal symbols. The 'language' rules specifying the set of allowed utterances and possibly also specifying the semantics associated with these utterances. For instance, grammars are commonly used in speech understanding systems to specify both the set of allowed sentences and to specify "tags" to extract semantic entities (e.g. the "amount" of money). Constraining the number of sentences accepted by a grammar is essential for reducing misinterpretations of user queries by an NLP system. For instance, for speech understanding systems, if the grammar accepts a large number of sentences, then the likelihood of recognizing uttered sentences as random, irrelevant, or undesirable sentences is increased. For transaction processing systems, misrecognized words can lead to unintended transactions being processed. An effective constraining grammar can reduce transactional errors by limiting the number of sentence level errors. The problem of over-generalization of speech grammars and related issues is well discussed by Seneff (1992). Thus, speech grammars must often balance the conflicting requirements of • accepting a wide variety of sentences to increase flexibility, and • accepting a small number of sentences to increase system accuracy and robustness. Developing tight grammars which trade-off these conflicting constraints is a tedious and accepted by a grammar is the set of all terminal strings that can be generated from the start symbol by successive application of the production rules. The grammar may optionally have semantic interpretation rules associated with each production rule (e.g. see (Allen 95)). 210 difficult process. Typically, grammars …

Content truncated to 500 words