Dictionary Alignment for Context-sensitive Word Glossing

Dictionary Alignment for Context-sensitive Word Glossing Willy Yap and Timothy Baldwin Department of CSSE University of Melbourne {willy,tim}@csse.unimelb.edu.au Abstract This paper proposes a method for automati- cally sense-to-sense aligning dictionaries in different languages (focusing on Japanese and English), based on structural data in the respective dictionaries. The basis of the proposed method is sentence similarity of the sense definition sentences, using a bilingual Japanese-to-English dictionary as a pivot during the alignment process. We experiment with various embellishments to the basic method, including term weighting, stemming/lemmatisation, and ontology ex- pansion. 1 Introduction In a multi-lingual environment such as the Internet, users often stumble across webpages authored in an unfamiliar language which potentially contain infor- mation of interest. While users can consult dictio- naries to help them understand the content of the webpages, the process of looking up words in un- familiar languages is at best time-consuming, and at worst impossible due to a range of reasons. First, the writing system of the language may be unfamil- iar to the user, e.g. the Cyrillic alphabet for a mono- lingual English speaker. Second, the user may not be familiar with the non-segmenting nature of lan- guages such as Chinese and Japanese, and hence be incapable of delimiting the words to look up in the dictionary in the first place. Third, the user may be unable to lemmatise the word to determine the form in which it is listed in a dictionary. There are several alternatives to help decipher webpages in unfamiliar languages. The first one is to use an online machine translation system such as Altavista’s Babel Fish1 or Google Translate.2 1http://babelfish.altavista.com/ 2http://www.google.com/translate t Figure 1: Multiple translations for the Japanese word 上げる [ageru] produced by rikai.com. The correct translation in this context is “to raise”. While web-based machine translation services occa- sionally produce good translations for linguistically- similar languages such as English and French, they do not perform very well in translating languages which are removed from one another (Koehn, 2005). The second alternative is a pop-up glossing ap- plication. The application takes raw text or a URL, parses the words, and returns the pop-up translation of each word as the mouse hovers over it. Some example pop-up glossing applications for Japanese source text and English glosses are Rikai3 and POPjisyo.4 With the aid of these pop-up transla- tions, the manual effort of segmenting words (if nec- essary) and looking up each can be avoided. This application is also useful as an educational aid for learners of that language. The drawback with these applications is they dis- play all possible translations of a given word irre- spective of context. Faced with the task of deter- mining the correct translation themselves, users fre- quently misinterpret words. An illustration of this situation is given in Figure 1. 3http://www.rikai.com/perl/Home.pl 4http://www.popjisyo.com Proceedings of the Australasian Language Technology Workshop 2007, pages 125-133 125 We propose a context-sensitive dictionary gloss- ing application to enhance the utility of on-line glossing applications by sensitising the presented glosses to the context …

Content truncated to 500 words