Few-shot named entity recognition based on triplet loss

With the development of small-sample learning, small-sample named entity recognition has become an emerging research field in recent years and has achieved more research results, but there is still a large research space. For example, small-sample named entity recognition also has inaccurate detection of entity boundaries, and the constructed prototype vectors cannot well summarize the meaning of the category in the vector space and are difficult to distinguish from the prototype vectors of other categories, resulting in poor model performance. In order to address these two problems, this paper proposes a new small sample-based named entity recognition model, which firstly uses the category name vectors as triplet loss function anchor vectors, and the accuracy of the constructed category prototypes in the spatial vectors can be effectively improved by using this method. Second, an improved adaptive sampling is used to extract positive and negative sample examples, which solves the problem of insufficient model training. Finally, to alleviate the problem of inaccurate detection of entity boundaries, a span correction module is proposed in the inference stage of the model as a way to correct the problem. The experimental results show that the method achieves an average F1 value of 46.3% on the CONLL2003, WNUT17, and GUM datasets, which is an improvement of 3.1 percentage points compared to the TadNER model based on contrast learning. This indicates that the model achieves some success on the small sample named entity recognition task. In addition, the effectiveness of each method is further verified in the paper by ablation experiments.