Achieving morphological agreement with Concorde

Neural conversational models are widely used in applications like personal assistants and chat bots. These models seem to give better performance when operating on word level. However, for fusion languages like French, Russian and Polish vocabulary size sometimes become infeasible since most of the words have lots of word forms. We propose a neural network architecture for transforming normalized text into a grammatically correct one. Our model efficiently employs correspondence between normalized and target words and significantly outperforms character-level models while being 2x faster in training and 20\% faster at evaluation. We also propose a new pipeline for building conversational models: first generate a normalized answer and then transform it into a grammatically correct one using our network. The proposed pipeline gives better performance than character-level conversational models according to assessor testing.