An Overview of the SEBAMAT Project

An Overview of the SEBAMAT Project Reinhard Rapp 1ILSP / Athena R.C., 2Magdeburg-Stendal University of Applied Sciences, 3University of Mainz reinhardrapp@gmx.de George Tambouratzis ILSP / Athena R.C., 6 Artemidos & Epidavrou, Maroussi, 15125, Greece giorg_t@ilsp.gr Abstract SEBAMAT (semantics-based MT) is a Marie Curie project intended to contrib- ute to the state of the art in machine translation (MT). Current MT systems typically take the semantics of a text only in so far into account as they are implicit in the underlying text corpora or dictionaries. Occasionally it has been argued that it may be difficult to advance MT quality to the next level as long as the systems do not make more explicit use of semantic knowledge. SEBAMAT aims to evaluate three approaches incorporating such knowledge into MT. 1 The current state of the art in MT SEBAMAT aims to show ways on how to im- prove the translations produced by current MT systems. For several decades the rule-based ap- proach was dominant (Arnold et al., 1994) which focused on grammatical well-formedness. In the statistical approach (SMT, Brown et al. 1990), linguistic rules were replaced by statistical patterns as automatically extracted from large monolingual and parallel text corpora. Recently, the dominance of SMT has been contested by neural MT (NMT), which almost consistently generates better results. Although NMT represents the current state of the art, technical issues and problems have been raised, including NMT’s inferior performance to SMT for limited training data, reduced portability across domains, and sensitivity to semantic divergence in the training data (Koehn & Knowles (2017), Carpuat et al. (2017). Sennrich & Zang (2019) improve NMT for small corpora, but marked im- provements are gained using larger corpora. However, despite considerable advances, the quality of current MT systems is still limited, and the likely reason is that the algorithms used are of a mechanical nature, employing statistical rules or deep learning architectures, without a human-like understanding of the texts to be translated. Amongst others, Kevin Knight point- ed out that MT systems do not sufficiently take into account semantic considerations such as who did what to whom, when where and why. This is also true for NMT, where the translations are ty- pically fluent but often semantically inadequate. SEBAMAT will suggest steps to raise MT quali- ty by taking into account semantics more expli- citly than has usually been done so far. Three main directions will be investigated. 2 Explicit semantic disambiguation Up to now, almost all MT work involving pa- rallel and monolingual corpora has been based on raw texts. However, recently there have been significant advances in word sense induction and disambiguation using corpus-based automatic methods. Inspired by Vintar & Fiser (2016), we suggest to pre-process parallel corpora using word sense disambiguation software, and then apply classical SMT word alignment procedures on the disambiguated rather than the original corpora. That is, word senses rather than words are aligned, and bilingual dictionaries of word senses rather than dictionaries of words are extracted. If the disambiguation can …

Content truncated to 500 words