English-Latvian Toponym Processing: Translation Strategies and Linguistic Patterns

The paper presents a study of a challenging task in machine translation and crosslanguage information retrieval – translation of toponyms. Due to their linguistic and extra-linguistic nature, toponyms deserve a special treatment. The overall translation process includes two stages of processing: dictionary-based and out-ofvocabulary toponym translation. The latter is divided into three steps: source string normalisation, translation, and target string normalisation. The translation process implies an application of translation strategies and linguistic toponym translation patterns. Possible translation strategies, including transliteration and translation per se along with combined strategies, and linguistic toponym translation patterns, including multi-word patterns as well, were investigated and implemented for English-Latvian machine translation. 10,000 The UK-related toponyms from Geonames were selected for a development set. The evaluation of output quality on basis of a test set has showed 67% accuracy in out-ofvocabulary translation: 58% on a set containing one-word toponymic units and 81% on a multi-word test set.