A General-Purpose Material Entity Extraction Method from Large Compound Corpora Using Fine Tuning of Character Features

By automatically identifying chemical names in scientific publications, it is possible to enhance information management of this crucial entity category in various materials science domains. This enables improved retrieval and linking of relevant concepts, providing data support for associating new research findings with a vast corpus of previously published literature in materials design. Current general-domain pretrained language models primarily focus on powerful contextual word embedding features, while the contextual features of special characters and uppercase characters in chemical entities have not been explicitly considered, which can easily lead to challenges in entity nesting and recognition. Therefore, we propose a method that is more suitable for the chemical domain. Specifically, we enhance the relationships between chemical entity character embeddings using the prediction paradigm of pre-trained language models. Subsequently, we optimize the word embeddings of chemical entities within each context. Additionally, we also explore the incorporation of textual dependency syntactic features to enhance the inference performance. Experimental results on several publicly available chemical datasets confirm the effectiveness of our proposed model (Our codes are publicly available at https://github.com/fansyangz/charT ).