Toward High Performance Machine Translation: Preliminary Results from Massively Parallel Memory-Based Translation on SNAP

[MT Summit III, Washington, DC, USA, July 1991] Toward High Performance Machine Translation: Preliminary Results from Massively Parallel Memory-Based Translation on SNAP* Hiroaki Kitano, Dan Moldovan, and Seungho Cha Center for Machine Translation Parallel Knowledge Processing Laboratory Carnegie Mellon University Department of Electrical Engineering-Systems Pittsburgh, PA 15213 U.S.A. University of Southern California hiroaki@cs.cmu.edu Los Angeles, CA 90089-1115 Abstract This paper describes a memory-based machine translation system developed for the Semantic Net- work Array Processor (SNAP). The goal of our work is to develop a scalable and high-performance memory-based machine translation system which utilizes the high degree of parallelism provided by the SNAP machine. We have implemented an ex- perimental machine translation system DMSNAP as a central part of a real-time speech-to-speech dia- logue translation system. It is a SNAP version of the ΦDMDIALOG speech-to-speech translation sys- tem. Memory-based natural language processing and syntactic constraint network model has been incorporated using parallel marker-passing which is directly supported from hardware level. Exper- imental results demonstrate that the parsing of a sentence is done in the order of milliseconds. 1 Introduction In this paper, we will demonstrate that machine translation speeds in the order of milliseconds is attainable by using a marker-propagation algorithm and a specialized parallel hard- ware. The significance of the high-performance (or real-time) nat- ural language processing is well known. Parsing sentences at the milliseconds speeds enables the realization of a speech recognition module capable of real-time speech understand- ing which eventually leads to the real-time simultaneous in- terpretation system. Also, the millisecond order performance enables the system to parse hundreds of sentences in a sec- ond, or over 3 million sentences per hour. This in turn makes possible bulk processing of text such as full-text retrieval, summarization, classification, translation, indexing and tag- ging. In order to accomplish the high-performance natural lan- guage processing, we have designed a highly parallel ma- chine called Semantic Network Array Processor (SNAP) [Moldovan and Lee, 1990] [Lee and Moldovan, 1990], and implemented an experimental machine translation system *This research is funded by the National Science Foundation Grant No. MIP-9009111 and MIP-9009109. A version of this paper will appear in the International Joint Conference on Artificial Intelligence (IJCAI-91). called DMSNAP using a parallel marker-passing scheme. DM- SNAP is a SNAP implementation of the ΦDMDIALOG speech- to-speech dialogue translation system [Kitano, 1990a] [Ki- tano, 1991a], but with some modifications to meet hardware constraints. Despite its high performance, our system carries out sound syntactic and semantic analysis including lexical ambiguity, structural ambiguity, pronoun reference, control, unbounded dependency, and others. In the next section, we describe briefly the SNAP architec- ture, then, describe design philosophy behind the DMSNAP followed by descriptions on implementation and linguistic processing. Finally, performance are presented. 2 SNAP Architecture The Semantic Network Array Processor (SNAP) is a highly parallel array processor fully optimized for semantic network processing with marker-passing mechanism. In order to facil- itate efficient propagation of markers and to ease development of applications, a set of marker propagation instructions has been microcoded. …

Content truncated to 500 words