Learnable Conjunction with Adjective Enhanced Model for Mandarin Sentiment Analysis of Social Media Text

The advent of research into neural networks has led to the widespread utilisation of deep learning methodologies in the domain of text sentiment analysis, owing to their formidable data processing and pattern recognition capabilities. In recent years, Transformer and its variant Bert have attracted considerable attention, and their performance has been demonstrated to be superior in practice. However, it is important to note that China's unique social media data set is characterised by a variety of languages, distinct cultural backgrounds, and intricate emotional expressions, which poses a significant challenge to the generalisability of models trained on more homogeneous data sets. Consequently, the conventional Transformer model's efficacy in feature extraction is constrained in certain scenarios, impeding the efficiency of its implementation. It is necessary to adapt and optimize the model further to account for the specific attributes of social media data in the task design. Nevertheless, within the framework of Chinese linguistics, the significance of word order and conjunction is inherently apparent. Conversely, the potential of adjectives in emotional expression is frequently disregarded. This study endeavours to investigate an innovative approach, underpinned by a Transformer encoder model, to construct a reinforcement model that can integrate and learn adjective features. The model will be evaluated through experimentation with three publicly available Chinese social media datasets. The experimental data demonstrate that the model can utilise the attention mechanism to not only identify the emotional tendency of keywords but also to effectively combine the positional characteristics of conjunctions and adjectives to obtain local details and contextual meanings of the text. This enhancement of the model's effectiveness in feature extraction is a significant contribution to the field.