ELMo
A 2018 model from the Allen Institute for AI that produced contextual word embeddings from a two-way recurrent network.
Also: Embeddings from Language Models
ELMo, from the Allen Institute for AI and the University of Washington, was published in early 2018. It trained two stacked LSTM language models, one reading left to right and one right to left, and combined their internal states into a vector for each word in context. Adding ELMo vectors to existing systems improved results across question answering, sentiment and other benchmarks. It was one of the first widely used pre-trained language models, and was quickly overtaken by transformer models such as GPT-1 and BERT later the same year.