Recurrent Neural Networks and Morphological Features in Language Modeling for Serbian

This paper describes the current state-of-the-art language model for the Serbian language, and also a specific way of dealing with one of the issues that is present in Serbian automatic speech recognition systems. This issues arises from the fact that Serbian is a highly inflective language, which leads to recognition mistakes where the basic form of the word is guessed correctly, but the word ending is not. The mistaken word ending can indicate, for example, a wrong word case, grammatical gender or grammatical number. A way to solve this problem was to use additional lexical and morphological word features as neural network input while assessing the likelihoods of predicted words. The experiments have shown significant reduction in word error rates when employing certain combinations of these additional features.