Martine Smets

Linguistically informed statistical models of constituent structure for ordering in sentence realization

We present several statistical models of syntactic constituent order for sentence realization. We compare several models, including simple joint models inspired by existing statistical parsing models, and several novel conditional models. The conditional models leverage a large set of linguistic features without manual feature selection. We apply and evaluate the models in sentence realization for French and German and find that a particular conditional model outperforms all others. We employ a version of that model in an evaluation on unordered trees from the Penn TreeBank. We offer this result on standard data as a reference-point for evaluations of ordering in sentence realization.

Modular MT with a learned bilingual dictionary: rapid deployment of a new language pair

The MT system described in this paper combines hand-built analysis and generation components with automatically learned example-based transfer patterns. Up to now, the transfer component used a traditional bilingual dictionary to seed the transfer pattern learning process and to provide fallback translations at runtime. This paper describes an improvement to the system by which the bilingual dictionary used for these purposes is instead learned automatically from aligned bilingual corpora, making the systems transfer knowledge entirely derivable from corpora. We show that this system with a fully automated transfer process performs better than the system with a hand-crafted bilingual dictionary. More importantly, this has enabled us to create in less than one day a new language pair, French-Spanish, which, for a technical domain, surpasses the quality bar of the commercial system chosen for comparison.

The adaptation of a machine-learned sentence realization system to French

We describe the adaptation to French of a machine-learned sentence realization system called Amalgam that was originally developed to be as language independent as possible and was first implemented for German. We discuss the development of the French implementation with particular attention to the degree to which the original system could be reused, and we present the results of a human evaluation of the quality of sentence realization using the new French system.

Machine translation as a testbed for multilingual analysis

We propose that machine translation (MT) is a useful application for evaluating and deriving the development of NL components, especially in a wide-coverage analysis system. Given the architecture of our MT system, which is a transfer system based on linguistic modules, correct analysis is expected to be a prerequisite for correct translation, suggesting a correlation between the two, given relatively mature transfer and generation components. We show through error analysis that there is indeed a strong correlation between the quality of the translated output and the subjectively determined goodness of the analysis. We use this correlation as a guide for development of a coordinated parallel analysis effort in 7 languages.