Sentiment Analysis of Elderly in Mobile Social Networks: A Two-Stage Framework from Dataset Construction to Model Optimization

Accurate identification of the emotional polarity of elderly individuals' text is of great significance for delivering more effective care and enhancing the well-being of seniors at both societal and familial levels. A rich corpus of text data and an appropriate model are prerequisites for accurately identifying sentiment polarity. However, the current scarcity of sentiment corpora for the Chinese elderly population has led to insufficient accuracy and effectiveness in sentiment recognition models for elderly texts. Therefore, from the perspective of elderly emotional psychology, we have designed a two-stage framework to accurately identify the sentiment polarity in texts from the elderly population. The framework consists of two main components: 1) the construction of a sentiment corpus for the elderly and 2) the development of a sentiment recognition model. In the first stage, we significantly improved the efficiency and quality of data annotation through semi-automated means, addressing the issue of corpus scarcity for the elderly population. In the second stage, we proposed the ERNIE-BiLSTM-MultiCNN sentiment polarity recognition model, which combines ERNIE, BiLSTM, and multi-scale convolutional neural networks (MultiCNN). This model effectively improves the accuracy of current sentiment recognition models for the elderly. On the constructed sentiment corpus, the ERNIE-BiLSTM-MultiCNN model achieved an accuracy of 95.08%, a precision of 95.11%, a recall of 95.08%, and an F1 score of 95.08%.