Example-Based Translation on Massivcly Parallel Computers

本研究的主要成果如下。我们收集了日文和英文的计算机科学术语，建立了包含7000个双语术语(翻译实例)的翻译数据库。研究了基于动态规划的最佳匹配检索方法。该方法检索了几个与检索输入类似的翻译实例。我们研究了并行最佳匹配检索方法，并在大型并行机nCUBE2.4上实现了该方法。我们使用并行最佳匹配检索程序在nCUBE2上构建了一个技术术语翻译系统。该系统是商用大规模并行计算机上的第一个机器翻译系统。该系统表现出较高的性能和良好的可扩展性：性能与处理器的数量成正比，并且可以扩展到具有数百万个翻译实例的系统。我们在大规模并行计算机CM-5上构建了另一个翻译系统。该算法具有较高的性能和良好的可扩展性：在SS2上比顺序系统快约15倍。通过本研究，我们得出结论：基于实例的并行翻译具有良好的性能和可扩展性。

The Major results of this research are as follows.1. We collected technical terms of Computer Science in Japanese and in English, and constructed translate database with 7,000 bilingual technical terms (translation examples).2. We investigated the best match retrival method using Dynamic Programming. This method retrieves several translation examples similar to the retrieval input.3. We investigated the parallel best match retriecval mcthod and implemented it on a massively parallel computer, nCUBE2.4. We constructed a translation system of technical terms on an nCUBE2 using the parallel best match retrieval program. This system is the first machine translation system on a commercial massively parallel computer. This system shows high performance and good scalability : the performance is proportional to the number of processors, and it can be scaled up to the system with some million translation examples.5. We constructed another translation system on a massively parallel computer, CM-5. It also shows high performance and good scalability : It is about fifteen times faster than the sequential system on a SS2.From this research, we concluded that parallel Example-Based Translation is promising : it shows high performance and good scalability.