This paper introduces contextual string embeddings, a novel type of word embedding derived from character-level language models. These embeddings are trained without explicit word notions, modeling words as character sequences and being contextualized by surrounding text, leading to different embeddings for the same word in different contexts. The approach leverages the internal states of a trained character language model to generate embeddings that are highly effective for sequence labeling tasks. The embeddings are evaluated against previous methods and show significant improvements in tasks such as named entity recognition (NER) for English and German, as well as in phrase chunking and part-of-speech (PoS) tagging. The proposed embeddings outperform existing methods, achieving new state-of-the-art results on the CoNLL03 NER task for both English and German. The paper also demonstrates that the embeddings can subsume previous embedding types, enabling simplified sequence labeling architectures. The character-level language model is compact and efficient to train, allowing for easy adaptation to new languages or domains. The authors release all code and pre-trained language models in a simple-to-use framework to facilitate research and application. The approach is based on a sequence tagging architecture that uses a bidirectional character-level neural language model to generate contextual string embeddings, which are then used in a BiLSTM-CRF sequence labeling module. The results show that the proposed embeddings are highly useful for sequence labeling tasks, with significant improvements in NER and slight improvements in PoS and chunking. The paper also discusses the inherent semantics of the embeddings, showing that they capture syntactic-semantic features and can disambiguate words in context. The approach is compared to prior works, and it is found that the proposed method outperforms previous approaches in most tasks. The paper concludes that the proposed contextual string embeddings are a promising approach for sequence labeling tasks and that further research is needed to apply them to additional sentence-level tasks.This paper introduces contextual string embeddings, a novel type of word embedding derived from character-level language models. These embeddings are trained without explicit word notions, modeling words as character sequences and being contextualized by surrounding text, leading to different embeddings for the same word in different contexts. The approach leverages the internal states of a trained character language model to generate embeddings that are highly effective for sequence labeling tasks. The embeddings are evaluated against previous methods and show significant improvements in tasks such as named entity recognition (NER) for English and German, as well as in phrase chunking and part-of-speech (PoS) tagging. The proposed embeddings outperform existing methods, achieving new state-of-the-art results on the CoNLL03 NER task for both English and German. The paper also demonstrates that the embeddings can subsume previous embedding types, enabling simplified sequence labeling architectures. The character-level language model is compact and efficient to train, allowing for easy adaptation to new languages or domains. The authors release all code and pre-trained language models in a simple-to-use framework to facilitate research and application. The approach is based on a sequence tagging architecture that uses a bidirectional character-level neural language model to generate contextual string embeddings, which are then used in a BiLSTM-CRF sequence labeling module. The results show that the proposed embeddings are highly useful for sequence labeling tasks, with significant improvements in NER and slight improvements in PoS and chunking. The paper also discusses the inherent semantics of the embeddings, showing that they capture syntactic-semantic features and can disambiguate words in context. The approach is compared to prior works, and it is found that the proposed method outperforms previous approaches in most tasks. The paper concludes that the proposed contextual string embeddings are a promising approach for sequence labeling tasks and that further research is needed to apply them to additional sentence-level tasks.