Entity Identification of Patent Information for Carbon Capture Technology Based on the RoBERTa-BiLSTM-CRF Model
Yan Liang, Yun Zhou, Enjiang Zhu, Jianzhi Sun · 2024
Aiming at the problems of unclear entity boundaries, many specialised vocabularies and multiple meanings in patent text data, this paper proposes a named entity recognition algorithm based on RoBERTa-BiLSTM-CRF for carbon capture domain. This algorithm deeply mines the semantic features of text through RoBERTa model, combines with Bi-directional Long Short-Term Memory (BiLSTM) to obtain contextual features, and performs sequence decoding and annotation through Conditional Random Field (CRF) to output the highest-scoring predicted labels, so as to improve the accuracy of model recognition. In this paper, we constructed a patent named entity recognition dataset for carbon capture technology and conducted experiments on this basis, the model has improved the accuracy, recall and F1 value over the mainstream deep learning methods on the carbon capture patent corpus, which can provide support for the evaluation of the value of patents and the analysis and retrieval of patents.