A Framework for MT and Multilingual NLG Systems Based on Uniform Lexico-Structural Processing

A Framework for MT and Multilingual NLG Systems Based on Uniform Lexico-Structural Processing Benoit Lavoie CoGenTex, Inc. 840 Hanshaw Road Ithaca, NY USA, 14850 benoit@cogentex.com Richard Kittredge CoGenTex, Inc. 840 Hanshaw Road Ithaca, NY USA, 14850 richard @ cogentex.com Tanya Korelsky CoGenTex, Inc. 840 Hanshaw Road Ithaca, NY USA, 14850 tanya @ cogentex.com Owen Rambow * ATT Labs-Research, B233 180 Park Ave, PO Box 971 Florham Park, NJ USA, 07932 rambow @research.att.com Abstract In this paper we describe an implemented framework for developing monolingual or multilingual natural language generation (NLG) applications and machine translation (MT) applications. The framework demonstrates a uniform approach to generation and transfer based on declarative lexico-structural transformations of dependency structures of syntactic or conceptual levels ("uniform lexico-structural processing"). We describe how this framework has been used in practical NLG and MT applications, and report the lessons learned. 1 Introduction In this paper we present a linguistically motivated framework for uniform lexico- structural processing. It has been used for transformations of conceptual and syntactic structures during generation in monolingual and multilingual natural language generation (NLG) and for transfer in machine translation (MT). Our work extends directions taken in systems such as Ariane (Vauquois and Boitet, 1985), FoG (Kittredge and Polgu6re, 1991), JOYCE (Rainbow and Korelsky, 1992), and LFS (Iordanskaja et al., 1992). Although it adopts the general principles found in the above- mentioned systems, the approach presented in this paper is more practical, and we believe, would eventually integrate better with emerging statistics-based approaches to MT. * The work performed on the framework by this co- author was done while at CoGenTex, Inc. The framework consists of a portable Java environment for building NLG or MT applications by defining modules using a core tree transduction engine and single declarative ASCII specification language for conceptual or syntactic dependency tree structures 1 and their transformations. Developers can define new modules, add or remove modules, or modify their connections. Because the processing of the transformation engine is restricted to transduction of trees, it is computationally efficient. Having declarative rules facilitates their reuse when migrating from one programming environment to another; if the rules are based on functions specific to a programming language, the implementation of these functions might no longer be available in a different environment. In addition, having all lexical information and all rules represented declaratively makes it relatively easy to integrate into the framework techniques for generating some of the rules automatically, for example using corpus-based methods. The declarative form of transformations makes it easier to process them, compare them, and cluster them to achieve proper classification and ordering. 1 In this paper, we use the term syntactic dependency (tree) structure as defined in the Meaning-Text Theory (MTT; Mel'cuk, 1988). However, we extrapolate from this theory when we use the term conceptual dependency (tree) structure, which has no equivalent in MTT (and is unrelated to Shank's CD structures proposed in the 1970s). 60 Thus, the framework represents a generalized processing environment that can be reused in different types …

Content truncated to 500 words