A Pattern-Based Machine Translation System - Yakushite Net MT Engine

A Pattern-Based Machine Translation System — Yakushite Net MT Engine Miki Sasaki and Toshiki Murata Oki Electric Industry Co., Ltd. 2-5-7 Hon-machi, Chuo-ku, Osaka 541-0053, JAPAN {sasaki234, murata656}@oki.com Abstract This paper introduces our pattern-based machine translation system. Our system enables us to register translation patterns including more complicated sentence structures rather than just simple words and limited types of expressions. Having registered in the Japanese to English section of IWSLT05, we aimed to input as many expressions as possible to our system in order to achieve successful translation of travel expression into English within a limited period of time. 1. Introduction Our machine translation in our system is pattern-based. With this system translation knowledge description is more flexible as the knowledge is written in the form of “translation pattern”. We challenged IWSLT05, the Japanese-English task, manual transcription and C-STAR track with this machine translation system of ours. To tackle this task, we focused on manual registration of translation rules which we extracted from the training corpus. Although the score we got in the evaluation campaign was not what we can be proud of, we proved that the translation patterns in our system are sufficiently effective. The next section presents our system description. Section 3 reports our training methods planned. Section 4 reports our results. Finally, Section 5 presents our conclusions. 2. System Description Our system (Yakushite Net MT Engine) uses a pattern-based machine translation method [1][2][3][4][5]. In this method, all grammars, word dictionaries, idioms, expressions, and sentences which are needed for translation are treated as translation patterns, and translation is done by preparing massive amount of patterns and using some of those patterns in appropriate combinations. Because grammars and word dictionaries are treated as translation patterns, they both can be registered in the same way to our system. Likewise, we can register translation patterns including more complicated sentence structures rather than just simple words and limited types of expressions. Idiomatic phrases that are characteristic of travel conversations, which are usually difficult to describe in a general dictionary, can also be registered as translation patterns to the system and used in the machine translation. 2.1. Architecture of Our Pattern-Based Machine Translation System Figure 1 shows the architecture of our system. Straight arrows show the flow of the translation and dotted arrows show the sequence for dictionary reference. First, the source sentence is analyzed morphologically, normalizing words and decorating them with morphological features by the morphological analyzer. This decorated sequence of words is then passed to the parser. The sentence is parsed by using the translation patterns in the appropriate dictionaries. Our system has the user and system dictionaries. When the sentence is parsed successfully, the parse tree is translated by top-down generation of the parse tree of the target language. Then, the post generator handles the generated tree to produce refined sentences. Lastly, the morphological synthesizer adjusts inflection and conjugation, and the translated sentence is outputted. Figure 1: the Architecture of our System system dictionary morphological analyzer source sentences parser/generator post …

Content truncated to 500 words