Authorship Attribution is a well-established task in computational linguistics, involving the identification of the likely author of a given text based on linguistic and stylistic features. This work explores a hybrid approach to authorship attribution by constructing a feature vector that integrates both syntactic and semantic information from scientific texts. The experimental setup focuses on evaluating the effectiveness of these features in distinguishing authorship across of three curated corpora of scientific documents. We employ computationally lightweight classification models (as logistic regression) in order to assess the discriminative power of the features themselves, rather than relying on the complexity of the classifier. Our results indicate that the enriched feature representation achieves accuracy levels ranging from 0.85 to 0.96, aligning with or exceeding benchmarks reported in the literature. These findings underscore the viability of combining manual feature engineering with pre-trained language models and traditional machine learning algorithms for authorship attribution. Furthermore, the approach offers a compelling alternative to deep learning-based pipelines, especially in scenarios where computational resources are limited.