Using clustering to improve Italian-English machine translation

 

One of the problems affecting machine translation from Italian into English regards adjective-noun pairs. The problem arises when a single Italian adjective has more than one possible English translation, such as It. alto – English tall, high, and the adjective choice depends on the noun co-occurring with it (cf. a tall man, a high price). It can also occur if a given noun collocates with different adjectives in the two languages, i.e. adjectives which are not translations of each other: for example, a nightmare is un brutto sogno (‘ugly dream’) in Italian, but a bad dream (It. ‘cattivo sogno’) in English. Machine translation systems often output incorrect translations when faced with this type of problem.

My proposed solution offers a translation model based on the clustering of English nouns. Since it is shown that words occurring in similar contexts tend to exhibit similar behaviour, the intuition is that, by clustering nouns on the basis of their lexical environment, we will have groups of nouns which behave homogeneously with respect to adjective selection. This model is tested on a simple translation task and its performance is compared to that of baseline operating on the principle of always choosing the most frequent English translation of the Italian adjective. The clustering-based model’s superior performance leads to the conclusion that ‘enhanced’ translation models offer a significant advantage over simple statistical models and such enhancements to the generation component of translation models are to be advocated.

 

 


last updated February 2006