Example-based controlled translation

73 Example-Based Controlled Translation Nano Gough and Andy Way School of Computing Dublin City University Dublin 9, Ireland [{ngough, away}@computing.dcu.ie] Abstract. The first research on integrating controlled language data in an Example-Based Machine Translation (EBMT) system was published in [Gough & Way, 2003]. We improve on their sub-sentential alignment algorithm to populate the system’s databases with more than six times as many potentially useful fragments. Together with two simple novel improvements—correcting mistranslations in the lexicon, and allowing multiple translations in the lexicon—translation quality improves considerably when target language translations are constrained. We also develop the first EBMT system which attempts to filter the source language data using controlled language specifications. We provide detailed automatic and human evaluations of a number of experiments carried out to test the quality of the system. We observe that our system outperforms Logomedia in a number of tests. Finally, despite conflicting results from different automatic evaluation metrics, we observe a preference for controlling the source data rather than the target translations. 1. Introduction Research in Machine Translation (MT) has explored many different methods over the years, including rule-based, statistical and example-based models as well as hybrid and multi-engine approaches. Certain MT systems have been developed for particular sublanguage domains. Furthermore, since 1996, there has been a growing interest in controlled languages and their application in MT as demonstrated by the series of CLAW workshops on controlled language applications. These have sparked the development of both monolingual and multilingual guidelines and applications using controlled language (CL) for many languages. Natural language grammars can be restricted in such a way that ambiguity and complexity is lessened or eliminated completely. Controlled languages are subsets of natural languages whose grammars and dictionaries have been restricted for this purpose. As well as aiding human comprehension of texts, CLs can also be used for improving the computational processing of text and potential benefits have been claimed for the integration of controlled languages with translation tools. Until quite recently, however, the area of Controlled Translation has been largely ignored. Only a limited number of rule-based MT (RBMT) systems have been used to translate controlled language documentation, including Caterpillar’s CTE and CMU’s KANT system [Mitamura & Nyberg, 1995], and General Motors CASL and LantMark [Means & Godden, 1996]. However, such systems can be very complex and expensive to develop for controlled translation, as it is difficult to fine-tune such general-purpose systems to derive specific, restricted applications. It is widely recognised that the use of traditional RBMT systems can lead to the well known ‘knowledge- acquisition bottleneck’. It is also acknowledged that the use of corpus-based MT technology can overcome this problem. It is difficult, therefore, to comprehend why more work has not been done in the development of Example- Based MT (EBMT) systems for controlled language applications, especially when one considers that the quality of EBMT systems depends heavily on the quality of the reference translations in the system database—the more these are controlled, the better the expected quality of translation output by …

Content truncated to 500 words