Combinatory Examples Extraction for Machine Translation

Combinatory Examples Extraction for Machine Translation0 Alberto Sim˜oes and Jos´e Jo˜ao Almeida Universidade do Minho, Campus de Gualtar, 4710-057 Braga Abstract One of the bottlenecks of example-based machine translation (EBMT) is to be able to amass automatically quantities of good examples. In our work in EBMT, we are investigating how far one can go by performing example extraction from parallel corpora using Probabilistic Translation Dictionaries to obtain example segmentation points. In fact, the success of EBMT highly depends on examples quality and quantity, but also in their length. Thus, we give special importance on methods to extract different size examples from the same translation unit. With this article we show that it is possible to extract quantities for examples from parallel corpora just using probabilistic translation dictionaries extracted from the same corpora. 1 Introduction Recent research on Machine Translation (MT) have been focused on corpus based translation, and two specific trends: Statis- tic Based Machine Translation (SBMT) and Example Based Machine Translation (EBMT — Somers, 1999; Hutchins, 2005). In this second approach to MT, parallel corpora is used as a base of examples of pre- viously done translations. The translation process is defined as: translate : sentence × tmdb −→ sentence translate(s, db) def = let l1 = split(s) l2 = <match(db, x) | x ∈ l1> in recombine(l2) Basically, we will first split the sentence to be translated in segments, try to translate them using a database (db) of translations, and then recombine the translations in a fi- nal sentence. To split into segments is im- portant because it is highly improbable we find the full sentence we want to translate in the examples database. 0This work has been partially funded by Fun- da¸c˜ao para a Ciˆencia e Tecnologia of Portu- gal through grant POSI/PLP/43931/2001, and co- financed by POSI, within Linguateca. This approach to translation is being one of the more promising, mainly because it just depends on corpora. Traditional trans- lation approaches used language specific grammars that were difficult to write and error prone1. So, EBMT quality highly depends on the used corpora quality and the number and length of the examples available2. In fact, the granularity of the examples is very im- portant during EBMT. While a full-sentence example if very liable, it is not flexible (not easy to reuse). In the other hand, a sin- gle word example is very flexible, but not liable for translation (because it lacks con- text). So, we need examples with different sizes between these two extremes. This paper describes an algorithm to ex- tract segments from parallel corpora, using different example granularities and combi- nations. This algorithm is based on a trans- lation matrix where translation probabilities between words are set, and the relationship between words and segments are extracted. 1In this paper we will not discuss the differences between SBMT and EBMT systems. For some dis- cussion on the subject please refer to Hutchins, 2005 2EBMT depends as well on the type of corpora used, as it should …

Content truncated to 500 words