A semantic modeling-based NLP approach to sentiment analysis of film and television reviews
TL;DR
Data experiments and kernel density distribution analysis confirm that this architecture maintains low computational inference latency while ensuring effective feature extraction, providing a feasible engineering algorithm reference architecture for sentiment analysis of unstructured user review texts.
Abstract
To address the challenges of semantic decay and implicit sentiment polarity classification in film and television review texts, this paper proposes a multi-dimensional semantic feature fusion network model (MDSF-Net). This model integrates pre-trained encoding layers and one-dimensional local convolutional branches to extract local and global features from text sequences. It calculates the sentiment alignment weights of evaluation entities through a cross-layer attention mechanism and optimizes the decision boundary using a joint loss function incorporating a label smoothing strategy. In benchmark comparison experiments, MDSF-Net achieves a macro-average F1 score of 93.36% and a classification accuracy of 93.42%, representing improvements of 3.25% and 3.27% respectively compared to the baseline. Data experiments and kernel density distribution analysis confirm that this architecture maintains low computational inference latency while ensuring effective feature extraction, providing a feasible engineering algorithm reference architecture for sentiment analysis of unstructured user review texts.