Where Google fails

Abstract

The paper presents ongoing research in contrastive corpus linguistics with envisaged applications in machine translation (MT) and with focus on Google Translate (GT) performance in English-Bulgarian translation. Structural patterns, forms or expressions where automatic translation fails are identified and analysed in view of creating a GT-editing tool providing improved target language output. The paper presents the corpus and the corpus analysis method applied, including the identification of inacceptable string types, their structural analysis and categorization. For each failure type, pre- or post-GT editing transformations are proposed. A first outline is proposed of a GT-editing tool consisting of a pre-GT editor performing string identification, substitution or deletion operations, a post-GT editor with a set of more complex string transformation rules and an additional module transferring structural information.

Keywords

• machine translation
• pre-editing
• post-editing
• google translate
• bitext
• computational linguistics
• corpus linguistics