Multi-Level Semantic Fusion Optimization for Few-Shot Relation Classification

Existing few-shot relation classification (RC) studies usually utilize distant supervision (DS) to obtain reliable samples, but this method usually introduces external noise which affects the model performance. Many scholars have optimized the quality of the added samples by adding a semantic denoising module to the distant supervision process. However, these studies usually adopt a single evaluation mechanism to judge the quality of the samples, e.g., only considering the similarity of word and sentence dimensions, which can lead to overfitting or semantic bias in the model. Therefore, to overcome the above challenges, We propose a multi-level semantic fusion optimization method by adding a category semantic recognition module to the denoising part, which fuses features in three dimensions: word, sentence, and category, and allows the model to perform multi-level semantic interactions and associations, which are used as discriminative conditions to screen high-quality external data as additional samples. We conducted extensive experiments on the dataset Fewrel 1.0 and exceeded the state-of-the-art baselines in the field on three metrics: precision, recall, and F1 value.