We present and evaluate neural network models for semantic role labeling of texts in Russian. The benchmark for evaluation and training was prepared on the basis of the FrameBank corpus. The paper addresses different aspects of learning a neural network model for semantic role labeling on different feature sets including syntactic features acquired with the help of SyntaxNet. In this work, we rely on architecture engineering and atomic features instead of commonly used feature engineering. We investigate the ability of learning a model for labeling arguments of "unknown" predicates that are not present in a training set using word embeddings as features for the replacement of predicate lemmas. We publish the prepared benchmark and the models. The experimental results can be used as a baseline for further research in semantic role labeling of texts in Russian.
|Журнал||Komp'juternaja Lingvistika i Intellektual'nye Tehnologii|
|Состояние||Опубликовано - 2017|
|Опубликовано для внешнего пользования||Да|