SRL Modeling for Russian

Semantic role labeling (SRL), or shallow semantic parsing, involves assigning semantic labels to the participants of predicate-argument structure. SRL and neural SRL tasks in particular have received much attention during the past decade. Results so far have been very promising, yet only several authors investigated this field on the Russian language data. Thus we carried out the SRL research based on semantically annotated resource FrameBank (Lyashevskaya, Kashkin 2015) by means of neural networks. Our approach includes preprocessing stage with cleaning and reorganizing the data, the comparison of different pretrained embeddings benefits, building stage with multiple layers stacked and architecture variants observed, and the overall manual and statistical analysis of achieved results at the final stage. As a result, we propose the multi-layered neural network with bidirectional long short-term memory (BiLSTM) applied for semantic parsing of Russian sentences. Keywords: natural language processing, semantic role labeling, semantic parsing, the Russian language, neural networks.