Kiwifruit Planting Entity Recognition Based on Character and Word Information Fusion

Aiming at the problem of high complexity of real words and low recognition accuracy in the named entity recognition task of kiwifmit planting field, a entity recognition method of kiwifruit planting integrating character and word information was proposed. Based on BiGRU - CRF model, word level and character level information were fused. At the word level, by introducing word set information and using multiple self-attention mechanisms ( MHA) to adjust the weights of different words in the word set. At the same time, attention mechanism was used to ignore the unreliable word sets and focus on the important word sets to improve the entity recognition effect. At the character level, the unsupervised bidirectional encoder representations form transformers (BERT) pre-training model was introduced to enhance the semantic representation of words. Experiments were conducted on a homemade corpus in the kiwifruit cultivation domain containing 12 477 annotated samples and seven categories of entities, and the results showed that the Fl value of the model was improved by 1.58 percentage points compared with the SoftLexicon model. In addition, the experimental comparison of the model ResumeNER with Lattice - LSTM, WC - LSTM and other models in the open data set ResumeNER was carried out, and the best recognition effect was achieved. The Fl value reached 96. 17%, indicating that the method proposed had certain generalization ability. © 2022 Chinese Society of Agricultural Machinery. All rights reserved.