Article Preview
TopIntroduction
With the rapid development of social platforms and e-commerce websites, people are now free to express their opinions and emotions on topics of interest. This sentiment-driven information is widely disseminated in various forms, including text, images, videos, and audio. Analyzing these sentiment-imbued opinions can facilitate a better understanding of user behavior, discern users' product preferences, identify user needs, customize marketing strategies, enhance public sentiment analysis, and detect fake news (Sahoo & Gupta, 2021; Sarkissian & Tekli, 2021;Zhang et al., 2021). Accurately analyzing this vast amount of sentiment information has become a research hotspot in the field of natural language processing (NLP).
Text Sentiment Analysis, also referred to as Opinion Mining, is a process aided by computers to swiftly capture and organize the subjective evaluation information on the internet. It adeptly mines and assesses individuals’ viewpoints, emotions, reviews, stances, and the sentimental inclinations of texts about entities like products, services, organizations, events, and topics, followed by inductive reasoning and inference of this mined data (Do et al., 2022; Hajjar & Tekli, 2022; Wang & Yang, 2021).
Currently, researchers both domestically and internationally have shifted their focus from traditional research based on dictionaries and machine learning to sentiment classification methods grounded in deep learning (Barbosa et al., 2022; Daou et al., 2021). For instance, Abdi et al. (2019) acknowledged the strengths of CNN (convolutional neural network) and RNN, proposing an LSTM (long short-term memory) model that elevates the accuracy of sentiment classification in reviews by over 5% through multi-feature fusion. Bin et al. (2022) embarked on sentiment tendency analysis on Weibo using Baidu's ERNIE and a dual attention mechanism (Att-M). Bin et al. performs dynamic Dynamic feature representation of text through ERNIE semantic representation capability, which combines combining sentiment resources and attention mechanism to solve the problem of different meanings for the same word in traditional word vectors, but their approach lacks the use of linguistic features. Xingjie and Yunze (2020) introduced a method based on convolutional neural networks and attention models for text sentiment analysis and personalized recommendations. The dimensions in the user preference vector indicate the user's preference for the corresponding dimension of the product, but there is no effective fusion of ratings and text. Dandan et al. (2021) utilized the BERT model to pre-train balanced short texts, representing language models with feature vectors, subsequently executing Chinese short text categorization; but it does not incorporate location information such as emoticons and symbols.
These deep learning-based methodologies outperform those built on manually crafted features, yet their capability in capturing deeper semantic information remains somewhat deficient (Barr et al., 2022; Tekli et al., 2021). Moreover, there is limited research both domestically and internationally on aspect-based sentiment classification of Chinese short texts. The majority of these studies directly apply attention mechanisms, overlooking the importance of syntactic relationships. This results in an inability to fully utilize the contextual semantic information of aspect nodes and represent polysemy within contexts (Ismail et al., 2022; Sarivougioukas & Vagelatos, 2022). Concurrently, the untapped potential of prior sentiment resources within neural networks leads to subpar sentiment classification outcomes.