A method for automatic detection and manual localization of content-based translation errors and shifts

Abstract

In this paper, we describe a method for automatic detection and manual localization of four compositional translation errors and shifts between automatically aligned segments in source and target languages. The automatic detection of errors and shifts is based on a content word precision algorithm which measures the equality of information content between source and target segments. The manual localization of errors and shifts within the segments is based on the compositionality principle. The method allows for the detection and localization of two potential errors; omission and addition, as well as two translation shifts required to avoid a translation error such as over-translation (when a correct translation results in more content words than in the source segment) and under-translation (when a correct translation results in less content words than in the source segment). Because of manual localization within bilingual pairs of segments, the method is not intended for automatic error detection but for human-assisted revision of translations. The analysis, described with the method and the algorithm is applied to real translation examples culled from a state-of-the-art translation corpus sampled for various translation errors. The algorithm and the localization method have implications for the development of more content-oriented natural language processing as well as for the training of professional translators; they can also be useful for formal and systematic description of content-based translation errors.