Text-embedding Enhanced Emotion Recognition in News for Low-latency Inference on Edge Devices

With the rapid advancement of artificial intelligence, emotion recognition based on textual data has become a prominent research frontier. This study presents an efficient framework for news emotion recognition that leverages text-embedding techniques and is optimized for low-latency inference on edge computing devices. The approach integrates various text embedding models with classical machine learning algorithms using a dataset of annotated news comments for training and evaluation. Experimental results demonstrate that with the increase in sample size, the recognition accuracy increases, but the growth curve tends to be flat; in terms of emotion recognition, the recognition performance of text2vec model is slightly better than that of moka massive mixed embedding (M3E), and both models have strong compatibility with emoticons. The K-nearest neighbor (KNN) algorithm performs better than the decision tree (DT) and random forest (RF) algorithms. The maximum recognition accuracy of the model is slightly higher than that of the [Formula: see text]. At the same time, the smaller model volume will help identify the deployment of models in edge computing devices and provide lower latency services.