<p>Unsupervised aspect category detection aims to identify the underlying aspect categories discussed in a given sentence without any annotated labels. Recent studies typically generate pseudo-labels from review corpora and subsequently train models in a supervised manner. However, during pseudo-label generation, existing methods either fail to capture aspect discriminability within sentences or rely solely on pre-trained models derived from general corpora, which can mislead the training process and degrade model performance. To mitigate the limitations, we propose a novel framework (SRM-CSR) that integrates lexical-level aspect-relevant information and sentence-level contextual representations for generating high-quality pseudo labels. Specifically, in the lexical-level stage, we extract Aspect-Relevant Terms (ARTs) based on two properties: (1) domain specificity, measured by word frequency divergence between the review corpus and a general corpus; and (2) semantic stability, reflected by semantic consistency across different sentences. We further introduce an entropy-driven discriminability mechanism that can assign higher weights to aspect terms within sentences based on their similarity distributions with the extracted ARTs. In the sentence-level stage, we leverage Sentence-BERT to encode sentences into contextual representations and compute the similarity between seed sentences and sentences from the review corpus. Pseudo-labels are generated independently in both stages. We construct the training set based on the consistent results produced by the two stages, which is subsequently used to train a neural classifier based on a post-trained Domain Knowledge BERT. Extensive experiments on three real-world datasets demonstrate the effectiveness of the pseudo-labeling strategy in SRM-CSR, with the proposed method achieving an average improvement of 2.9 percentage points in macro-F1 over the strongest baselines.</p>