Lexical choice in Abstract Dependency Trees

In this work lexical choice in generation for Machine Translation is explored using lexical semantics. We address this problem by replacing lemmas with synonyms in the abstract representations that are used as input for generation, given a WordNet synset. In order to find the correct lemma for each node we propose to map dependency trees to Hidden Markov Trees that describe the probability of a node given its parent node. A tree-modified Viterbi algorithm is then utilized to find the most probable hidden tree containing the correct lemmas given their context. The model is implemented in a Machine Translation system for English to Dutch. The output sentences, generated from the modified dependency structures, contained a lot of erroneous substituted words. This is mainly due to the fact that a large amount of synsets, used as input for the model, are incorrect. The input to the model now contains the synset that is most frequent given the lemma in general, not the optimal synset given the domain of the sentences. We therefore propose to implement a domain specific WSD-system in our pipeline in future work.