Few-shot Multi-label Aspect Category Detection Based on Prompt-Enhanced Prototypical Network
Few-shot multi-label aspect category detection is a prominent area of research within fine-grained sentiment analysis.In methods based on prototypical networks,the lack of training data and the presence of irrelevant noise words severely impact the quality of prototype vector generation using attention mechanisms.Addressing this challenge,our study introduces the ProtPrompt model,an innovative prototype network enhanced through prompt learning.Firstly,we align pre-trained tasks with downstream tasks using prompt learning,guiding the model in precise sentence representation to learn more discriminative vectors.This effectively en-hances category distinguishability.Meanwhile,we utilize cosine similarity to calculate the loss function instead of the commonly used Euclidean distance in prototypical networks,mitigating the influence of high-dimensional vector spaces.Secondly,we design an optimized framework to attenuate noise interference on sentence vector representation,thereby fostering the aggregation of sen-tences sharing the same aspect category within the embedding space.Experimental results validate the effectiveness of our proposed ProtPrompt model on three publicly available datasets.The experimental results show that the model improves the F1 score over the state-of-the-art(SOTA)model by 4.35%,8.62%,and 8.39%for the three publicly available datasets,respectively.This substantiates its ability to efficiently detect aspect categories.
国家科技期刊平台
NSTL
万方数据
