Disambiguating translation strategies in MT using automatic named entity recognition

18 Selecting Translation Strategies in MT using Automatic Named Entity Recognition Bogdan Babych1,2 and Anthony Hartley1 1 Centre for Translation Studies,University of Leeds,UK 2 Department of Computer Science, University of Sheffield,UK [bogdan@comp.leeds.ac.uk; a.hartley@leeds.ac.uk] Abstract We report on the results of an experiment aimed at enabling a machine translation system to select the appropriate strategy for dealing with words and phrases which have different translations depending on whether they are used as proper names or common nouns in the source text. We used the ANNIE named entity recognition system to identify named entities in the source text and pass them to MT systems in the form of "do-not-translate" lists. A consistent gain of about 20% in translation accuracy was achieved for all tested systems. The results suggest that successful translation strategy selection is dependent on accurate segmentation and disambiguation of the source text – aspects which could be significantly improved by named entity recognition. We further suggest an automatic method for distinguishing and lexical differences in MT output that could have applications in automated MT evaluation for morphologically rich languages. 1. Introduction Language communities develop certain acceptable practices and norms for translating different types of concepts, expressions and texts from other languages and cultures. These practices are described as translation methods, translation strategies and translation procedures. (Vinay and Darbelnet, 1958, 1995). Translation methods relate to whole texts, while strategies and (finer-grained) procedures relate to sentences and smaller units (Newmark, 1988:81). The choice of a translation strategy often depends on the type of a translated unit. For example, for certain types of proper names the optimal translation strategy is transference, i.e., a “do- not-translate” or “transliterate” strategy, while the majority of common nouns are translated with other strategies: literal translation, transposition, modulation, etc. (Newmark, 1988: 81-88). This implies that recognising different types of units in the source text is a necessary condition for optimising the choice of translation strategy and, ultimately, for improving the quality of the target text. The problem of selecting translation strategies for words that may be used as proper names or common nouns in the source language is related to a more general problem of word sense disambiguation (WSD) – one of the most serious problems for Machine Translation technology. Dealing with “proper vs common disambiguation” (PCD) often requires combining different knowledge sources, in a similar way to WSD (Stevenson and Wilks, 2001). But the cross-level nature of this problem also suggests that improvement in MT quality could be achieved through improving related aspects of the source-text analysis, such as Named Entity (NE) recognition (Babych and Hartley, 2003; Somers, 2003:524). For the purposes of this discussion, we assimilate proper nouns to NEs and investigate NE recognition as a possible solution to the PCD problem insofar as it might enable the selection of the correct strategy. Accurate NE recognition is important for the general quality of MT for the following reasons: 1. The translation of the same token may be different depending on whether the token is a common …

Content truncated to 500 words