Neural language models: training text representation you’ve always been waiting for
The speech is a gentle overview of the most promising continuous space language models, which obtain new state-of-art results on a variety of natural language processing tasks. It covers both performance of statistical language models and neural language models together with a detailed comparison of these methodologies pros and cons. During the talk, Galina will also provide technical details of existing implementations of the algorithms and complexity of their maintenance in a real-life production environment. Special attention will be paid to the architectures of BERT, ELMo, GPT-2, and ULMFiT.