Apertium-IceNLP: A rule-based Icelandic to English machine translation system

Apertium-IceNLP: A rule-based Icelandic to English machine translation system Martha Dís Brandt, Hrafn Loftsson, Hlynur Sigurþórsson School of Computer Science Reykjavik University IS-101 Reykjavik, Iceland {marthab08,hrafn,hlynurs06}@ru.is Francis M. Tyers Dept. Lleng. i. Sist. Inform. Universitat d’Alacant E-03071 Alacant, Spain ftyers@dlsi.ua.es Abstract We describe the development of a pro- totype of an open source rule-based Icelandic→English MT system, based on the Apertium MT framework and IceNLP, a natural language processing toolkit for Icelandic. Our system, Apertium-IceNLP, is the first system in which the whole morphological and tagging component of Apertium is replaced by modules from an external system. Evaluation shows that the word error rate and the position- independent word error rate for our pro- totype is 50.6% and 40.8%, respectively. As expected, this is higher than the corre- sponding error rates in two publicly avail- able MT systems that we used for com- parison. Contrary to our expectations, the error rates of our prototype is also higher than the error rates of a comparable system based solely on Apertium modules. Based on error analysis, we conclude that bet- ter translation quality may be achieved by replacing only the tagging component of Apertium with the corresponding module in IceNLP, but leaving morphological anal- ysis to Apertium. 1 Introduction Over the last decade or two, statistical machine translation (SMT) has gained significant momen- tum and success, both in academia and industry. SMT uses large parallel corpora, texts that are translations of each other, during training to derive a statistical translation model which is then used to c⃝ 2011 European Association for Machine Translation. translate between the source language (SL) and the target language (TL). SMT has many advantages, e.g. it is data-driven, language independent, does not need linguistic ex- perts, and prototypes of new systems can by built quickly and at a low cost. On the other hand, the need for parallel corpora as training data in SMT is also its main disadvantage, because such corpora are not available for a myriad of languages, espe- cially the so-called less-resourced languages, i.e. languages for which few, if any, natural language processing (NLP) resources are available. When there is a lack of parallel corpora, other machine translation (MT) methods, such as rule-based MT, e.g. Apertium (Forcada et al., 2009), may be used to create MT systems. In this paper, we describe the development of a prototype of an open source rule-based Icelandic→English (is-en) MT system based on Apertium and IceNLP, an NLP toolkit for process- ing and analysing Icelandic texts (Loftsson and Rögnvaldsson, 2007b). A decade ago, the Ice- landic language could have been categorised as a less-resourced language. The current situation, however, is much better thanks to the development of IceNLP and various linguistic resources (Rögn- valdsson et al., 2009). On the other hand, no large parallel corpus, in which Icelandic is one of the languages, is freely available. This is the main rea- son why the work described here was initiated. Our system, Apertium-IceNLP, is the first sys- tem in which the …

Content truncated to 500 words