GF Wide-coverage English-Finnish MT system for WMT 2015

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 141–144, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. GF Wide-coverage English-Finnish MT system for WMT 2015 Prasanth Kolachina Computing Science Chalmers University of Technology prasanth.kolachina@cse.gu.se Aarne Ranta Computing Science Chalmers University of Technology aarne.ranta@cse.gu.se Abstract This paper describes the GF Wide- coverage MT system submitted to WMT 2015 for translation from English to Finnish. Our system uses a interlingua based approach, in which the interlingua is a shared formal representation, that ab- stracts syntactic structures over multiple languages. Our final submission is a re- ranked system in which we combine this baseline MT system with a factored LM model. 1 Introduction Interlingual translation is an old idea that has been suggested numerous times and refuted almost as many times. A typical criticism is that the very idea is utopistic: that one can never build an inter- lingua that faithfully represents meaning in all lan- guages of the world. However, as the focus in ma- chine translation has shifted from the perfect ren- dering of meaning to less modest goals, the idea of an interlingua can be reconsidered. In the current paper, we describe our sys- tem submission to the WMT shared task in the English-Finnish track. Our system is an interlingua-based system, the interlingua based on an abstract syntax in the sense of Grammatical Framework (GF) (Ranta, 2011). GF has been pre- viously shown to work for domain-specific MT outperforming state-of-art systems using semantic interlinguas (Ranta et al., 2011). Departing from this, the GF wide-coverage Translator is an at- tempt following the current mainstream in the field of MT: we are content with browsing quality in the output of the MT systems, while achieving the low cost of interlingual MT systems. As such, the shared abstract syntax is mapped to different “sur- face” languages representing an abstraction of the deep syntactic structure for each of the languages. The abstraction from word order, morphology and certain deep syntactic phenomena, allows the in- terlingua to cope with unrelated languages. At the same time, these systems are scalable beyond toy examples, into wide-coverage systems. We submit this system as our baseline over the English Finnish language pair for the WMT shared task. In addition, we also submitted a “re-ranked” variant of the same system as our primary submis- sion, using statistical language models to re-score the translations from the baseline. Automatic eval- uation metrics have shown small improvements from re-ranking our baseline system 1. The paper is organized as follows: we describe our baseline system in Section 2 and the re-ranked variant in Section 3. We present our experiments and relevant discussion in Section 4. 2 GF Wide-coverage Translator The GF Translator pipeline has three main phases: • Parsing converts the source sentence into a forest of abstract syntax trees (AST), i.e. in- terlingual representations. • Disambiguation selects the most probable AST. • Linearization converts the AST into a sen- tence in each of the target languages. Disambiguation is for …

Content truncated to 500 words