In this survey, we provide a comprehensive description of recent neural entity linking (EL) systems. We distill their generic architecture that includes candidate generation, entity ranking, and unlinkable mention prediction components. For each of them, we summarize the prominent methods and models, including approaches to mention encoding based on the self-attention architecture. Since many EL models take advantage of entity embeddings to improve their generalization capabilities, we provide an overview of the widely-used entity embedding techniques. We group the variety of EL approaches by several common research directions: joint entity recognition and linking, models for global EL, domain-independent techniques including zero-shot and distant supervision methods, and cross-lingual approaches. We also discuss the novel application of EL for enhancing word representation models like BERT. We systemize the critical design features of EL systems and provide their reported evaluation results.
The paper introduces methods of adaptation of multilingual masked language models for a specific language. Pre-trained bidirectional language models show state-of-the-art performance on a wide range of tasks including reading comprehension, natural language inference, and sentiment analysis. At the moment there are two alternative approaches to train such models: monolingual and multilingual. While language specific models show superior performance, multilingual models allow to perform a transfer from one language to another and solve tasks for different languages simultaneously. This work shows that transfer learning from a multilingual model to monolingual model results in significant growth of performance on such tasks as reading comprehension, paraphrase detection, and sentiment analysis. Furthermore, multilingual initialization of monolingual model substantially reduces training time. Pre-trained models for the Russian language are open sourced.