Machine translation: statistical approach with additional linguistic knowledge

In this thesis, three possible aspects of using linguistic (i.e. morpho-syntactic) knowledge for statistical machine translation are described: the treatment of syntactic differences between source and target language using source POS tags, statistical machine translation with a small amount of bilingual training data, and automatic error analysis of translation output. Reorderings in the source language based on the POS tags are systematically investigated: local reorderings of nouns and adjectives for the Spanish–English language pair and long-range reorderings of verbs for the German–English language pair. Both types of reorderings result in better performance of the translation system, local reordering being more important for the scarce training corpora. For such corpora, strategies for achieving an acceptable translation quality by applying appropriate morpho-syntactic transformations are exploited for three language pairs: Spanish– English, German–English and Serbian–English. Very scarce task-specific corpora as well as conventional dictionaries are used as bilingual training material. In addition to conventional dictionaries, the use of phrasal lexica is proposed and investigated. A framework for automatic analysis and classification of actual errors in translation output based on combining existing automatic evaluation measures with linguistic information is presented. Experiments on different types of corpora and various language pairs show that the results of automatic error analysis correlate very well with the results of human evaluation. The new metrics based on analysed error categories are used for comparison of different translation systems trained on various sizes of texts with and without morpho-syntactic transformations. For improving the quality of a statistical machine translation system by the use of morphosyntactic information, the choice of the method and the significance of improvements strongly depend on the language pair, the translation direction and the nature of the corpus. Error analysis of the translation output is important in order to define weak points of the system and apply methods for improvement in the optimal way.