An Adaptation of Lexical Conceptual Structures to Multilingual Processing in an Existing Text Understanding System

An Adaptation of Lexical Conceptual Structure to Multilingual Processing in an Existing Text Understanding System Bonnie Glover Stalls Robert S. Belvin Alfredo R. Arnaiz Christine A. Montgomery Robert E. Stumberger Language Systems, Inc. {glover,robin,arnaiz,chris,res}@lsi.com Abstract The extension of an existing text understanding system to multilingual translation using an interlingual approach prompted the search for a language-independent means of expressing predicates and predicate relations. The approach adopted was Lexical Conceptual Structure (LCS), as put forth by [2, 3] and [1]. LCS incorporates notions of event structure, argument relations, and core meaning into a concise language-independent formalism that preserves struc- ture at the same time that it allows for divergences among languages. The task of installing LCS proved simple and straightforward, given the text understanding system's explicit representa- tion of sentential content and capability of indexing the analyzed constituents of the sentence to the appropriate slots in the LCS. 1 Introduction Language Systems, Inc. (LSI)'s core automated text understanding technology has been applied over the last three years to multilingual processing in a voice-to-voice translation system that we are developing for the Air Force. Originally designed as an application-independent language analyzer, LSI's Data Base Generation (DBG) system is a flexible, modular system that produces a knowledge representation of the events and entities in a text, incorporating syntactic, semantic, discourse, and other relevant information. The DBG system has been adapted to a number of different applications, including data base update for space event reports [4] and for air activities messages [14], message fusion for radiotelephone traffic [6], data base generation for reports of terrorism in Latin America (MUC-3 and MUC-4)[5, 8] and for the transfer of microelectronics technology (MUC-5) [7], and most recently, the project for which we have developed the multilingual processing capability that we are describing here, machine-aided voice translation [9,10, 11]. 2 Multilingual Processing and LSI's DBG System The Machine-Aided Voice Translation (MAVT) project, now in its second phase of prototype de- velopment, is being designed to assist English-speaking Air Force personnel in interacting with speakers of Spanish, Arabic, and Russian. The MAVT testbed consists of three subsystems: speech recognition, language processing, and speech generation. Like the voice-to-voice English → Span- ish → English system developed in the first phase of the MAVT project, the system currently under development is a speaker-independent continuous speech translation system, processing query- response interactions in a military domain. As with the previous system, the language processing 106 functions—understanding, translation, and generation—are performed by LSI's DBG natural lan- guage processing system. Much of the extension of the DBG system for the MAVT project has necessarily focused on multilingual capabilities. In the first phase of the project, the DBG system already had in place a multilingual syntactic parser that was used for Spanish and English. This parser will be used to parse Arabic and Russian as well. DBG produces, as output of the understanding phase of process- ing, a knowledge representation of the sentence. This knowledge representation is an application- independent data structure of related event …

Content truncated to 500 words