Multi-Modal Named Entity Recognition Analysis of Mongolian Language Based on Knowledge Enhanced Pre-Trained Models

With the emergence of massive multimodal data in the information age, research in the field of multimodal named entity recognition (MNER) has made significant progress in mainstream languages such as Chinese and English. However, due to the lack of efficient multimodal characterization methods, research in low-resource languages such as Mongolian is still relatively limited. Therefore, this paper proposes a MNER method for Mongolian based on a knowledge-enhanced pretraining model. First, this paper utilizes the CLIP pretraining model to construct a unified feature representation space, which improves the cross-modal knowledge information alignment capability and alleviates the cross-modal Mongolian semantic divide problem. Second, this paper proposes a multimodal interaction strategy combining cross-modal cross-attention mechanism and gating mechanism, which effectively captures the key features and improves the Mongolian semantic relevance among different modalities. Finally, the conditional random field is used for sequence annotation, which realizes the accurate recognition of Mongolian named entities. The experimental results show that the F1 value of the Mongolian MNER method proposed in this paper reaches 76.81% in the experiments, which is significantly better than the baseline model, and provides new research ideas and methods for the field of Mongolian MNER.