Learning-to-Translate Based on the S-SSTC Annotation Schema

Learning-to-Translate Based on the S-SSTC Annotation Schema Tang Enya Kong Zaharin Yusoff Christian Boitet Enyakong1@gmail.com University Sarawak Malaysia, Sarawak, Malaysia zarinby@gmail.com Université Joseph Fourier, Grenoble, France Christian.Boitet@imag.fr Abstract We present the S-SSTC framework for machine translation (MT), introduced in 2002 and developed since as a set of working MT systems (SiSTeC-ebmt). Our approach is example-based, but differs from other EBMT approaches in that it uses alignments of string-tree alignments, and in that supervised learning is an integral part of the approach. Our model directly deals with three main difficulties in the traditional treatment of MT that stem from its separation from the "translation task" (the 'world'). First, by allowing the system to learn from real translation examples directly, we avoid the need to indefinitely pursue the elusive goal of writing grammars to exactly describe intermediate syntactico- semantic monolingual representations and their correspondences. Second, we make explicit the dependence of the MT system performance on the input from the environment. That is possible only because the learning process uses feedback from the real translation knowledge when constructing its knowledge representation. Third, such MT systems using an inductively learned knowledge base yield a desirable non-regressive behavior by using translation mistakes to improve their knowledge base. 1. Introduction The S-SSTC-based framework for the construction of MT systems has been introduced in 2002 [Mosleh et. al. 2002] and developed since to an operational state (SiSTeC-ebmt for English-Malay and English-Chinese). In this article, we would like to stress a particular aspect, namely that this approach is better capable of modeling the translation knowledge of human translators than other example-based approaches. Because the translation knowledge is represented as alignments (synchronizations) between string-tree alignments (SSTCs, or structured string-tree correspondences), it is more natural to translators (and post-editors) than direct word-word, string-string or chunk-chunk correspondences used in classical SMT and EBMT models. It is also totally static, hence more understandable than procedural knowledge embedded in almost all RBMT approaches. The learning process which is an integral part of the development of SiSTeC-ebmt MT systems can in fact be viewed as a special case of the study of reasoning reported in [Khardon&Roth 94], because it combines the interfaces to the 'world' used by known learning models with the reasoning task and a performance criterion suitable for it. In such a framework, the intelligent agent is given access to its learning interface, and is also given a grace period in which it can interact with this interface and construct its representation Knowledge Base (KB) of the 'world'. Its reasoning performance is measured only after this period, when it is presented with 'queries' from some query language, relevant to the 'world', and has to answer whether such 'queries' are implied by the learned 'world' model. In our case, the 'world' is the 'translation task' captured in terms of the parallel texts produced by human translators and enriched by their S-SSTCs, and the 'queries' are simply modeled by a predicate Translate(ST,TT) where ST is the source language text and TT is a …

Content truncated to 500 words