Improving Machine Translation Quality

Improving MT Quality: Towards a Hybrid MT Architecture in the linguatec 'Personal Translator' Gregor THURMAIR linguatec Gottfried-Keller-Str. 12 81245 Munich, Germany g.thurmair@linguatec.de Abstract This paper reports on measures to improve the quality of MT systems, by using a hybrid system architecture which adds corpus-based and statistical components to an existing rule- based system backbone. The focus is on improving the accuracy of the dictionary resources. 1 Baseline Although there have been significant improvements in the last period, the quality of machine translation is still an issue: The acceptance of the technology would improve significantly if the quality of translation would not prevent users from larger scale usage. Research has not really focused on this issue; instead there were many attempts to start anew, hoping that a change in technology would lead to improved system quality; however, up to now, this has not proven to be the case. This also holds for the latest developments in statistical machine translation (SMT); it is a development which tries to apply learning techniques on existing bilingual corpus material, and uses already existing translation material to compute translation equivalents for phrases, and complete sentences, on this basis (Knight and Koehn, 2003). The question how the results compare to existing approaches has recently been studied in a bit more detail (Thurmair, 2005, with a comparison of • German-to-English MT), with two main results: The overall quality of SMT is outperformed by existing rule-based MT systems. • The overall quality of both approaches is not yet sufficient, as between 20 and 30% of the evaluated sentences were ranked as being unacceptable. A closer look at the results shows that the main sources of errors in SMT (about 60%) are related to phenomena like Satzklammer and split verb constructions, non-standard constituent ordering, gapping etc., all of which could be rather easily described in a rule-based framework, while the main sources of errors in rule-based systems (again about 60%) consist of lexical issues, and wrong selection of lexical material, which in turn a corpus-based approach can easily avoid1. Examples are a. Wrong VP treatment and constituent order in SMT: (1) Beim Anlegen einer Tabelle wird ein Schlüssel nach Gesichtspunkten vergeben2. programmtechnischen when you create a table is a key after a a technical assign . (2) Die Parameter der Datenbanksystemprüfung können Sie in der R/3-Tabelle DBCHECKORA konfigurieren3. the parameters of the database system check in R /3 configure table DBCHECKORA . b. Wrong lexical selection in rule-based MT: (1) Der zweite Teilbaum beschreibt die Struktur des zu lesenden Datenbestandes4. The second partial tree describes the structure of the data stock to be read. (2) Für die Verknüpfung mit Organisation- sobjekten müssen Sie kein HR im Einsatz haben5. You don't have to have any HR for the bonding with organization objects in use As a consequences, a combination of linguistic and statistical techniques into a hybrid system 1 The rest consists mainly in grammatical mistakes; here the problem is less that grammatical structures are not covered but rather that the …

Content truncated to 500 words