A dual-branch transformer architecture via asymmetric keyword prior and parallel attention pooling for fine-grained sentiment analysis

Fine-grained sentiment classification in highly specialized domains poses fundamental algorithmic challenges, as general-purpose pre-trained language models assign uniform prior weights to all input tokens and lack structured mechanisms for disentangling aspect-level representations. To resolve these limitations, this paper proposes a novel domain-adaptive Transformer architecture specifically designed for Chinese cultural heritage tourism review analysis, featuring two core algorithmic components. First, we introduce an asymmetric attention mechanism that injects a domain indicator bias directly into the pre-softmax score matrix. By augmenting the standard scaled dot-product $$QK^{\top }/\sqrt{d_{k}}$$ with an additive term $$\alpha \cdot \text {diag}(m) \cdot \textbf{1}^{\top }$$, this mechanism breaks the token-symmetric attention prior in a principled manner without altering encoder depth or parameter count. Second, we design a parallel attention-pooling architecture that utilizes K independent learnable query vectors to compute soft alignment weights, producing disentangled aspect representation vectors. These modules are seamlessly integrated and jointly optimized under a multi-task cross-entropy objective, simultaneously supervising global polarity and K-way aspect-level predictions. The six aspect dimensions—historical authenticity, cultural experience, facility accessibility, service quality, price perception, and environmental atmosphere—are grounded in the experience economy framework and validated through preliminary data analysis of the constructed corpus. Theoretical and empirical analyses demonstrate that the proposed asymmetric bias provides an interpretable relevance signal, and the joint architecture significantly outperforms strong baselines across multiple evaluation metrics, proving the necessity of each structural design decision.