Example-based machine translation of part-of-speech tagged sentences by recursive division

MT Summit VII Sept.1999 Example-Based Machine Translation of Part-Of-Speech Tagged Sentences by Recursive Division Tantely Andriamanankasina, Kenji Araki and Koji Tochinai Graduate School of Engineering, Hokkaido University Kita 13 Nishi 8, Kita-ku, Sapporo 060-8628, Japan Email: {tantely,araki,tochinai}@media.eng.hokudai.ac.jp Abstract Example-Based Machine Translation can be applied to languages whose resources like dictionaries, reliable syntactic analyz- ers are hardly available because it can learn from new translation examples. However, difficulties still remain in translation of sen- tences which are not fully covered by the matching sentence. To solve that prob- lem, we present in this paper a translation method which recursively divides a sen- tence and translates each part separately. In addition, we evaluate an analogy-based word-level alignment method which pre- dicts word correspondences between source and translation sentences of new transla- tion examples. The translation method was implemented in a French-Japanese ma- chine translation system and spoken lan- guage text were used as examples. Promis- ing translation results were earned and the effectiveness of the alignment method in the translation was confirmed. 1 Introduction In the traditional Rule-Based Machine Translation (RBMT), huge amount of rules and dictionaries need to be prepared and maintained [1]. To avoid that hard and time-consuming task, Example-Based Ma- chine Translation (EBMT) was proposed [2], The ba- sic idea of EBMT is to extract, among a collection of translation examples, a number of translation exam- ples whose source sentence is similar to the sentence to be translated and achieve the translation task by imitating these translation examples. Rules are not required. Instead, the system learns from translation examples. Various EBMT models have been proposed [3. 4. 5] and different issues were discussed [6]. Current EBMT models use syntactic analyzer re- sults. Utilization of syntactic and semantic analyzers is expected to produce accurate translation results. However, these tools themselves are not perfect and still not available in languages where research has not been sufficiently carried out. On the other hand, re- cent lexical analyzers [7, 8] are extremely precise. We therefore, have proposed an EBMT method not de- pending on syntactic or semantic analyzers [9]. Lexi- cal analyzer is merely used together with the parallel and aligned corpus. The study is limited to Part Of Speech (POS) tags. However, the translation model is considered to be adaptable to any additional tags which may raise the precision of the translation. If tagged text are available, or the recently proposed Global Document Annotation1 becomes widespread, highly accurate translation system will be expected. Our proposed method is implemented in a French- Japanese EBMT system. However, it is generally de- signed for languages whose resources are hardly avail- able. The translation is possible and a correct result can be earned when the input sentence is almost covered by the matching sentence. However, when such exam- ple cannot be discovered, it is hard to translate sen- tences correctly. Even for relatively short sentences, there are cases where uncovered segments, when they exist, cause errors. Besides, there are errors from mis- matches appearing during the …

Content truncated to 500 words