Soft-Aspect ABSA: A Probabilistic Framework with Cluster Stability and Class-Imbalance Diagnostics

Aspect-based sentiment analysis (ABSA) is increasingly the granularity at which customer feedback is consumed, and recent work has pushed the field rapidly toward transformer- and graph-based architectures [3,5–7]. However, most modern ABSA approaches assume either a closed manually curated aspect taxonomy or a fully supervised aspect extractor trained on benchmark corpora such as SemEval. Neither assumption holds in low-resource emerging-market settings, where aspects must be discovered from the corpus itself, annotation budgets are negligible, and class distributions can be unexpectedly skewed. This article introduces Soft-Aspect ABSA, a probabilistic, topic-model-agnostic framework that promotes unsupervised topic-model output to first-class aspects via a temperature-controlled softmax over topic-membership posteriors. We instantiate the framework with a spectral-clustering plus non-negative matrix factorisation (NMF) substrate on a corpus of 292 Google Play Store reviews of a Ghanaian retail-bank mobile application (April–September 2024). The corpus exhibits an inverted class imbalance (30.8% positive / 69.2% negative under a keyword-bootstrap rule) and a four-cluster topic decomposition. A baseline TF-IDF embedding head trained with binary cross-entropy collapses to the majority class on the held-out test set: accuracy 0.6949, minority-class F1 0.000, Matthews correlation 0.000, despite a ROC-AUC of 0.934 that indicates well-ranked probabilities. The framework licenses two closed-form remediations — class-weighted cross-entropy and focal loss [28] — that we evaluate empirically on the same head. Focal loss with γ = 1 lifts minority-class F1 from 0.000 to 0.818, Matthews correlation from 0.000 to 0.746, and ROC-AUC to 0.986, demonstrating that the framework correction is not merely formal but is recoverable on the case-study data. We also run a bootstrap stability protocol for cluster-count selection (B = 50) that flags the silhouette-max k* = 4 as only moderately stable (I_stab = 0.64). The contribution is methodological: a reusable scaffold for low-resource ABSA pipelines in which the aspect set is not given a priori.