This study addresses the challenges of semantic mixing, limited interpretability, and complex feature structures in fine-grained sentiment and opinion classification by proposing an interpretable feature disentanglement framework built on the latent space of large language models. The framework constructs multi-component latent representations that separate emotional polarity, opinion direction, target attributes, and pragmatic cues during encoding, thus overcoming the limitations of traditional methods that merge diverse semantic factors into a single representation. During representation learning, the model first uses a large model encoder to generate basic semantic features and then builds multiple independent subspaces through learnable projections. A covariance constraint is introduced to reduce coupling across semantic components and to create clear boundaries in the latent space. To preserve the essential information of the original text, a reconstruction consistency mechanism integrates features from all subspaces to rebuild the global representation and enhance semantic completeness. The framework also incorporates semantic anchors to align latent components with interpretable semantic dimensions, giving each subspace a clear emotional or opinion-related meaning and improving transparency at the mechanism level. Experimental results show that the framework outperforms existing methods across multiple metrics and handles complex syntax, implicit semantics, and coexisting emotions with greater stability. It achieves high accuracy and interpretability in fine-grained sentiment and opinion analysis. Overall, the proposed disentanglement framework provides an effective approach for building structured, multidimensional, and interpretable representations of textual emotions and opinions and holds significant value for complex semantic understanding tasks.

This paper addresses the problem of text classification models being vulnerable and lacking robustness under adversarial perturbations by proposing a robust text classification method based on large language model calibration. The method builds on a pretrained language model and constructs a multi-stage framework for semantic representation and confidence regulation. It achieves stable optimization of classification results through semantic embedding extraction, calibration adjustment, and consistency constraints. First, the model uses a pretrained encoder to generate context-aware semantic features and applies an attention aggregation mechanism to obtain global semantic representations. Second, a temperature calibration mechanism is introduced to smooth the output probability distribution, reducing the model's sensitivity to local perturbations. Third, adversarial consistency constraints are applied to maintain feature alignment between original and perturbed samples in semantic space, ensuring dynamic preservation of semantic robustness. The method adopts a joint loss function to balance three optimization objectives: classification accuracy, robustness, and confidence. To verify its effectiveness, sensitivity experiments on hyperparameters, environments, and data distributions are conducted. The results show that the model maintains high performance and stability under conditions such as word substitution, noise injection, and class imbalance, significantly outperforming several mainstream baseline models. This study achieves the integration of semantic-level robustness optimization and calibration learning, providing a new approach for building highly reliable text classification systems.

The internal structure and operation mechanism of large-scale language models are analyzed theoretically, especially how Transformer and its derivative architectures can restrict computing efficiency while capturing long-term dependencies. Further, we dig deep into the efficiency bottleneck of the training phase, and evaluate in detail the contribution of adaptive optimization algorithms (such as AdamW), massively parallel computing techniques, and mixed precision training strategies to accelerate convergence and reduce memory footprint. By analyzing the mathematical principles and implementation details of these algorithms, we reveal how they effectively improve training efficiency in practice. In terms of model deployment and inference optimization, this paper systematically reviews the latest advances in model compression techniques, focusing on strategies such as quantification, pruning, and knowledge distillation. By comparing the theoretical frameworks of these techniques and their effects in different application scenarios, we demonstrate their ability to significantly reduce model size and inference delay while maintaining model prediction accuracy. In addition, this paper critically examines the limitations of current efficiency optimization methods, such as the increased risk of overfitting, the control of performance loss after compression, and the problem of algorithm generality, and proposes some prospects for future research. In conclusion, this study provides a comprehensive theoretical framework for understanding the efficiency optimization of large-scale language models.