Development of a Hashtag Segmentation System in Russian

The primary aim of this work is to create a system for the hashtag segmentation in Russian which is suitable for managing spelling and grammar mistakes, slang and fictional words. To achieve this goal we develop two different systems: the first applies dictionary search and the second one relies on recurrent neural networks. The dictionary search approach which uses OpenCorpora reached 71.2% accuracy rate on the real hashtag dataset and 75.9% accuracy rate on the artificial data from fiction. The results of the char-LSTM are 64,2% and 73,3% on the same datasets. During the investigation, it was revealed that the first technique is more effective for hashtags which consist of regular dictionary words but the accuracy is getting lower when fictitious words or grammar mistakes appear in a string. The second approach handles out-of-dictionary words more precisely, however, it works significantly worse with names and confuses functional words with prefixes. The dictionary search-based method provides better results, so it forms the final system which is the console application. The volume of this work is 53 pages including 10 pictures and 2 tables.