A bilingual study on the prediction of morph-based improvement.
Balázs Tarján, Tibor Fegyó, Péter Mihajlik · 2014
Morph-based language modeling has been efficiently applied in improving the accuracy of Large-Vocabulary Continuous Speech Recognition (LVCSR) systems – especially in morphologically rich languages. However, the rate of improvements varies greatly and the underlying principles have been only superficially studied. Having a method that can predict the expected improvement prior to experimentations would be largely useful. In this paper, we introduce language-independent factors affecting morphbased improvement and show how they can be utilized in estimating the effectiveness of statistical morph-based language modeling. The task was broadcast news transcription in two less investigated languages, Hungarian and Romanian. It was found that in case of under-resourced conditions morph-based models can bring significant improvement – even for a morphologically less rich language like Romanian. In addition, it was shown that noninitial morph tagging can constantly outperform explicit modeling of word-boundaries both in terms of letter and word accuracies.