Few-Shot NER in Marine Ecology Using Deep Learning

Abstract

In the field of marine ecological named entity recognition (NER), challenges arise due to limited domain-specific text, weak semantic representations of input vectors and the neglect of local features. To address these challenges of NER in a low-resource environment, a deep learning-based few-shot NER model was proposed. Firstly, Sequence Generative Adversarial Nets (SeqGAN) was utilized to train on the original text and generated new text, thereby expanding the original corpus. Subsequently, BERT-IDCNN-BiLSTM-CRF was introduced for extracting marine ecological entities. BERT (Bidirectional Encoder Representation from Transformers) was pre-trained on the expanded corpus. The embeddings produced by BERT were then fed into Iterative Dilation Convolutional Networks (IDCNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM) to facilitate feature extraction. Finally, Conditional Random Fields (CRF) was employed to enforce label sequence constraints and yielded the final results. For the proposed few-shot NER method based on deep learning, comparative experiments were conducted horizontally and vertically against BiLSTM-CRF, IDCNN-CRF, BERT-IDCNN-CRF and BERT-BiLSTM-CRF models on both the original and expanded corpora. The results show that BERT-IDCNN-BiLSTM-CRF outperforms BERT-BiLSTM-CRF by 2.48 percentage points in F1-score on the original corpus. On the expanded corpus, BERT-IDCNN-BiLSTM-CRF achieves a F1-score 2.65 percentage points higher than that on the original corpus. This approach effectively enhances entity extraction in the domain of marine ecology, laying a foundation for downstream tasks such as constructing marine ecological knowledge graphs and ecological governance.