Institute of Information and Communication Technologies

We explore lexical choice in Natural Language Generation (NLG) by implementing a model that uses both context and frequency information. Our model chooses a lemma given a WordNet synset in the abstract representations that are the input for generation. In order to find the correct lemma in its context, we map underspecified dependency trees to Hidden Markov Trees that take into account the probability of a lemma given its governing lemma, as well as the probability of a word sense given a lemma. A tree-modified Viterbi algorithm is then utilized to find the most probable hidden tree containing the most appropriate lemmas in the given context. Further processing ensures that the correct morphological realization for the given lemma is produced. We evaluate our model by comparing it to a statistical transfer component in a Machine Translation system for English to Dutch. In this set-up, the word sense of words are determined in English analysis, and then our model is used to select the best Dutch lemma for the given word sense. In terms of BLEU score, our model outperforms a most frequent baseline, in which the most frequent lemma of a given word sense is always chosen. A manual evaluation confirms that our model is able to select the correct lemma when it is given a correct input synset. The majority of errors were caused by incorrect assignment of the word sense in the English analysis phase. Our model does not improve upon a transfer component trained on a parallel corpus. In the original transfer component, there barely are any lemmas that were incorrectly translated in the transfer phase, with the exception of Out of Vocabulary items (OOV’s). In a further experiment we only used our model for OOV’s and obtained a small improvement in BLEU score.