Using Patterns for Machine Translation

Using patterns for Machine Translation (MT) Stella Markantonatou, Sokratis Sofianopoulos, Vassiliki Spilioti, George Tambouratzis, Marina Vassiliou, Olga Yannoutsou∗ Institute for Language and Speech Processing, Artemidos 6; 151 25 Athens (Greece) {marks | s_sofian | vspiliot | giorg_t | mvas | olga}@ilsp.gr Abstract. In this paper an innovative approach is presented for MT, which is based on pat- tern matching techniques, relies on extensive target language monolingual corpora and em- ploys a series of similarity weights between the source and the target language. Our system is based on the notion of ‘patterns’, which are viewed as ‘models’ of target language strings, whose final form is defined by the corpus. 1. Introduction With this work, we further explore the ideas tested within the METIS-I1 system (Dologlou et al. 2003) which proved the feasibility of the innovative idea that sound translations could be received with hybrid MT that relied on monolingual corpora – rather than parallel ones – and flat bilingual lexica. This is the main difference between METIS systems and cor- pus-based approaches (EBMT, SMT) which rely on bilingual corpora. For corpus-based MT approaches, which have taken the lead from rule-based ones (Hutchins 1995), the basic resources, i.e. parallel corpora, are scarce. Such corpora are rare and available for the very widely spoken languages only. In addition, they quite often represent a certain register or sublanguage. Efforts to face the problem have focused on reducing the size of the required parallel corpus (Al Onaizan (2000), Brown (2003)). By resorting to mono- lingual corpora only, the METIS projects pur- sue a radically different solution to the prob- lem of scarcity of resources. However, METIS-I too faced a serious problem of sparseness of data as it could manipulate only 1METIS was funded by EU under the FET Open Scheme (METIS-I, IST-2001-32775), while METIS-II, the continuation of METIS, is being funded under the FET-STREP scheme of FP6 (METIS-II, IST-FP6- 003768). The assessment project METIS ended in Febru- ary 2003, while the second phase started in October 2004 and has a 36 month duration. sentences as units. In METIS-II, the frame- work of the present work (Markantonatou et al. 2005), material at sub-sentential level, namely chunks, is exploited to generate translations. The great promise with corpus-based ap- proaches lies in that ‘hard-to-manipulate’ lin- guistic information can be induced from the corpus rather than being explicitly represented with a constantly growing collection of rules. The syntactic and semantic preferences of words (one of the reasons why the number of rules tends to explode in both hand-crafted and tree-bank induced grammars (Gaizauskas 1995)) constitute a large part of the implicit information provided by the corpus. A similar argument can be made about word order. Thus, work on (various approaches to) corpus-based MT aimed at making do without resorting to any expensive linguistic resources such as (rich) computational lexica and grammars (Nagao 1984, Brown 1990). However, it has become evident that some amount of linguistic knowledge is necessary (see, for instance Pop- owich (2005) for the case of SMT …

Content truncated to 500 words