SHADO: A Semantics-Preserving Hybrid Framework for Automatic Text Classification Using Domain Ontology

In today's digital era, organizations across industries face unprecedented challenges in managing and extracting insights from vast repositories of unstructured textual content. From customer feedback analysis to regulatory compliance monitoring, the ability to automatically categorize textual documents has become a strategic imperative for competitive advantage and operational efficiency.

Traditional text classification methodologies have predominantly relied on feature extraction techniques such as Bag of Words (BoW) [1], n-grams [2], and TF-IDF [3], combined with machine learning algorithms including support vector machines [4, 5], Naive Bayes [4, 6], decision trees [4], and k-nearest neighbors (KNN) [7, 8]. While these approaches have demonstrated reasonable performance in controlled environments, they exhibit significant shortcomings when confronted with domain-specific terminology, contextual ambiguity, and the nuanced semantics that characterize real-world textual data [9].

The emergence of pre-trained language models and deep learning architectures has partially addressed these limitations, yet a critical gap remains: the lack of explicit domain knowledge integration that could bridge the semantic divide between surface-level textual features and deeper conceptual understanding.

This research addresses this challenge by introducing SHADO (A Semantics-Preserving Hybrid Framework for Automatic Text Classification Using Domain Ontology), a novel hybrid framework that systematically incorporates structured domain knowledge through ontological representations. Unlike existing approaches that treat ontologies as supplementary resources, our method positions domain ontologies as core semantic engines that guide both feature enhancement and classification decision-making processes.

Our contribution is threefold: (1) development of a semantic enrichment pipeline that leverages domain ontologies for contextual feature augmentation, (2) implementation of a dimensionality reduction strategy that preserves semantic relationships while optimizing computational efficiency, and (3) design of a hybrid classification architecture that synergistically combines traditional machine learning models with transformer-based language models like BERT for enhanced accuracy and interpretability.

The structure of this paper follows a systematic progression: we begin with a comprehensive literature analysis positioning our work within the current research landscape, followed by detailed methodology exposition, experimental validation using benchmark datasets, and conclude with performance analysis and future research directions.

Our classification model, called SHADO, introduces a comprehensive, multi-layered strategy for enhancing text classification through the integration of domain ontologies. Its primary objective is to improve the accuracy, robustness, and adaptability of classification systems by leveraging advanced NLP and machine learning techniques. The overall framework is structured into six key phases, as illustrated in Figure 1.

Figure 1. SHADO framework steps

3.1 Data preparation and cleaning

Text preprocessing is the first critical step where textual data is cleaned and prepared for further analysis [44]. This process includes:

• Cleaning: Removing irrelevant elements such as extra spaces, special characters, punctuation, and converting text to lowercase.
• Tokenization: Splitting the text into words, phrases, symbols, or meaningful elements called token [27, 28].
• Lemmatization: Reducing words to their base or root form to ensure uniformity in text representation [22, 29].
• Stop Word Removal: Eliminating words that do not contribute significant se-mantic meaning to focus analysis on relevant content [23].

Our enhanced preprocessing incorporates pattern preservation techniques that maintain important semantic markers such as temporal expressions, email addresses, and compound words. Unlike traditional approaches that indiscriminately remove numeric content, our method intelligently transforms numerical entities into semantic placeholders (<YEAR>, <DECIMAL>, <NUMBER>) while preserving their contextual significance. Additionally, we implement sentence boundary preservation through <SENT_END> markers, enabling downstream semantic analysis to maintain discourse-level context. Algorithm 1 details the steps involved in preprocessing text data:

3.2 Semantic and learning integration

This critical phase transforms preprocessed textual data into semantically enriched representations through strategic ontology integration. The semantic enhancement process bridges the gap between raw textual features and domain-specific knowledge, enabling more contextually aware classification. This phase focuses on enhancing raw text data with domain-specific semantic information to improve the learning and classification process. It is composed of two key components: Domain Ontology Integration and Feature Engineering & Selection.

3.2.1 Domain ontology integration

To provide contextual and domain-aware understanding of the textual data, relevant domain ontologies are integrated into the pipeline. This integration involves three main sub-steps:

(1) Ontology selection

Ontology selection is a critical process, as the quality and relevance of the selected ontologies significantly influence the effectiveness of semantic enrichment. The following criteria guide the selection:

• Domain Relevance: Ontologies must align closely with the subject matter of the texts (e.g., medicine, education, technology, politics) and include relevant concepts and relationships.
• Semantic Richness: Chosen ontologies should be conceptually dense, offering a rich network of terms and relations.
• Accessibility and Maintenance: Preference is given to open-access, well-documented, and actively maintained ontologies, such as those from BioPortal, OBO Foundry, or Linked Open Vocabularies.
• Technical Compatibility: Ontologies must be available in standard formats (e.g., RDF, OWL) to ensure smooth system integration.

These criteria are aligned with best practices outlined in Touza et al. [10], who emphasized the importance of relevance, richness, and format compatibility in selecting ontologies for semantic enrichment.

The number of ontologies used depends on the nature and complexity of the do-main:

• For well-established domains: At least two ontologies are used, even if one is widely recognized. This redundancy helps capture additional semantic nuances and enhances the robustness of the approach.
• For complex or interdisciplinary domains: More than two ontologies may be integrated to cover diverse concepts and ensure comprehensive semantic representation.

(2) Ontology mapping

Ontology mapping links textual terms to structured concepts defined within selected ontologies, enabling a semantic representation of the content. Formally, this process can be modeled by Eq. (3).

$C=f_{{map}}(T, O)$    (3)

where, C is the set of corresponding concepts in the ontology, T the set of terms from the text, O the select ontology and f_map the function that maps each term t ∈ T to one or more concepts c ∈ O.

The process by which textual terms are semantically linked to domain-specific concepts is illustrated in Figure 2, which visualizes the function $f_{m a p}$ by systematically associating terms with ontology concepts based on their semantic relevance and structural alignment. The figure presents a structured workflow that explicitly outlines the conditional logic used for concept matching, including fallback mechanisms when direct mappings are not found.

(3) Semantic enrichment

Once the terms have been associated with relevant concepts in the ontology (as described in Algorithm 2), a further step is needed to deepen their meaning and contextual understanding. This is the role of semantic enrichment, which consists in expanding each term’s representation by incorporating additional knowledge from the ontology—such as hierarchical relations (e.g., parent-child structures) and semantic links (e.g., "part of", "related to").

Figure 2. Flowchart of intelligent ontological mapping

This process allows the initial term set to be enhanced with richer, more structured information, enabling more accurate and meaningful interpretations of the text content. We formalize this enrichment process with the following expression (Eq. (4)):

T_enriched=f_enrich(T,C,R)    (4)

where,

• T is the original set of terms extracted from the text.
• C represents the concepts to which those terms have been mapped.
• R includes the relationships and hierarchies drawn from the ontology.
• f_enrich is the function that performs the enrichment by combining these elements.

The goal is to go beyond simple matching and provide a semantically empowered view of the text—leveraging ontological knowledge to unlock deeper insights and more effective downstream processing. To operationalize this enrichment process, Algorithm 2 outlines how to retrieve and integrate semantic information from the ontology for each mapped term.

3.2.2 Intelligent feature extraction and selection

Building upon the semantic enrichment achieved in the previous phase, this step systematically identifies and selects the most informative features by integrating statistical significance measures with ontological semantic relevance. This hybrid approach ensures that selected features capture both distributional patterns and domain-specific semantic relationships, creating an optimal foundation for subsequent classification models.

Our feature selection methodology combines traditional statistical measures with ontology-derived semantic indicators to create a comprehensive importance scoring system. This dual-criteria approach addresses the limitations of purely statistical methods while preserving computational efficiency.

• Statistical Foundation: The TF-IDF weighting scheme provides the statistical baseline for feature importance:

$T F-I D F(t, d)=T F(t, d) \times log \frac{N}{D F(t)}$    (5)

where, $T F(t, d)$ represents term frequency in document $d, N$ is the total number of documents, and $D F(t)$ is the number of documents containing the term $t$.

• Semantic Enhancement: Ontological centrality measures augment statistical importance by quantifying semantic significance within domain knowledge structures:

$\begin{aligned} & {Total~Importance}=T F-I D F \times { Centrality }_{ {onto }}(c)\end{aligned}$    (6)

where, $\mathrm{Centrality}_{{onto }}(\mathrm{c})$ represents the centrality score of concept $c$ within the ontological graph structure.

The feature extraction process operates through coordinated stages that progressively refine the feature space while maintaining semantic coherence. Algorithm 3a implements this coordinated approach through three sequential phases: statistical filtering, ontological enhancement, and semantic scoring.

The concept centrality quantifies the importance of each ontological concept within its semantic network and is used to weight features in the document representation. This measure combines three complementary aspects of a concept c is defined by Eq. (7):

$\begin{gathered}\text { concept_centrality }(c)=\alpha \cdot{degree}(c)+\beta \cdot {depth}(c)+\gamma \cdot {breadth}(c)\end{gathered}$    (7)

where,

• $\alpha, \beta$ and $\gamma$ are tunable weights reflecting the relative importance of each component.
• Degree centrality ($degree(c)$): the number of direct relationships that c has with other concepts in the ontology.
• Hierarchical depth ($depth(c)$): the position of c within the ontology hierarchy, normalized by the maximum depth. Concepts closer to the root may be more general, while deeper concepts capture domain-specific semantics.
• Semantic breadth ($breadth(c)$): the total number of indirectly connected concepts reachable from c, representing the semantic scope of the concept.

These three components are combined into a normalized score in the range [0,1]:

To address the limitations of single-configuration feature extraction, we propose a multi-modal approach that combines three complementary feature types: (1) word-level TF-IDF with bi-gram extensions capturing semantic relationships, (2) character-level TF-IDF with 3-6 gram patterns capturing morphological information, and (3) ontology-derived semantic features quantifying domain-specific concept density. This hybrid approach increases feature space richness from traditional single-vector representations to multi-dimensional semantic embeddings of 3100+ features. Algorithm 3b describes this process.

3.3 Dimensionality reduction with semantic preservation

Following feature extraction and semantic enrichment, the resulting high-dimensional feature space requires careful dimensionality reduction to maintain computational efficiency while preserving the rich semantic relationships captured through ontological integration. This phase implements a sophisticated reduction strategy that balances performance optimization with semantic fidelity. Our approach employs a multi-stage reduction pipeline that progressively refines the feature space.

The reduction process operates through three coordinated stages, each optimized for specific aspects of the feature space transformation.

(1) Stage 1: Statistical Variance Reduction (PCA)

Principal Component Analysis provides the initial dimensionality reduction by identifying the directions of maximum variance in the feature space is given by Eq. (8). where $X$ represents the original feature matrix, $W_{P C A}$ contains the principal components, and $Y$ is the reduced representation.

$Y=X W_{P C A}$    (8)

(2) Stage 2: Semantic Structure Preservation

This stage applies semantic constraints to ensure that ontologically related features maintain their relationships in the reduced space is given by Eq. (9).

$\begin{aligned} \min _w \| X W-Y & \|_F^2 +\lambda \sum_{(i, j) \in S} w_s(i, j)\left\|w_i-w_j\right\|_2^2\end{aligned}$    (9)

where, $S$ represents semantic similarity pairs, $w_s(i, j)$ denotes semantic weights, and $\lambda$ controls the semantic preservation strength.

(3) Stage 3: Non-linear Manifold Learning (t-SNE)

The final stage employs t-distributed Stochastic Neighbor Embedding to capture non-linear relationships while preserving local semantic neighborhoods is given by Eq. (10).

$P_{j \mid i}=\frac{exp \left(-\left\|x_i-x_j\right\|^2 / 2 \sigma_i^2\right)}{\sum_{k \neq i} {xp}\left(-\left\|x_i-x_k\right\|^2 / 2 \sigma_i^2\right)}$    (10)

where,

• $P_{j \mid i}$ is Conditional probability that point $x_i$ chooses $x_j$ as its neighbor.
• $x_i, x_j$ are data points in the original high-dimensional space.
• $\left\|x_i-x_j\right\|^2$ Squared Euclidean distance between points i and j.
• $\sigma_i^2$ Variance of the Gaussian kernel centered on point i.

The reduction process is implemented through Algorithm 4, which coordinates the three stages while continuously monitoring semantic preservation quality. The algorithm begins with statistical variance reduction via PCA, followed by the application of semantic constraints derived from the ontological graph structure, and concludes with adaptive t-SNE transformation based on dataset characteristics.

The Semantic_Graph G = (V, E) is constructed from all ontology-enriched concepts extracted during feature extraction. Nodes $v \in V$ represent concepts, while edges $e \in E$ represent semantic relationships such as is-a, part-of, or domain-specific connections. Edge weights $w_s(i, j)$ reflect semantic similarity between connected concepts, computed based on ontological proximity or co-occurrence in the corpus.

During dimensionality reduction, semantic constraints are applied to preserve the ontological relationships among features. In Stage 2 of Algorithm 4, these constraints are incorporated into the objective function for linear reduction (Eq. (9)), ensuring that concepts linked in the Semantic_Graph remain close in the reduced space. The regularization parameter λ controls the trade-off between preserving semantic relationships and minimizing reconstruction error.

Quality assessment employs Semantic Neighborhood Preservation (SNP) metrics that measure the retention of k-nearest semantic neighbors and Ontological Relationship Retention (ORR) scores that evaluate the preservation of ontological relationships in the reduced space.

The optimal target dimensionality is determined through cross-validation using classification performance as the primary criterion, balanced with semantic preservation scores and computational efficiency measures. This adaptive approach ensures that the reduced feature space maintains both computational tractability and the rich semantic structure essential for accurate ontology-enhanced classification, providing an optimal foundation for the subsequent hybrid classification phase.

3.4 Hybrid classification model

The final training phase leverages the semantically enriched and dimensionally optimized feature representations to construct a sophisticated ensemble classifier that combines the complementary strengths of traditional machine learning algorithms with modern transformer architectures. This hybrid approach integrates six distinct learning paradigms: Support Vector Machines excel at finding optimal decision boundaries in high-dimensional spaces, Random Forest provides robust feature importance estimates, Gradient Boosting provides a powerful trade-off between bias and variance, making it ideal for capturing complex patterns in structured data, KNN captures local neighborhood patterns enhanced by semantic similarity, Multi-Layer Perceptron learns complex non-linear mappings, and BERT contributes contextualized understanding of textual semantics.

The ensemble employs a two-tier architecture where base models generate individual predictions that are subsequently combined through an intelligent voting mechanism. Unlike simple majority voting, our approach implements weighted voting that assigns different importance to each model based on their reliability and past performance on similar semantic contexts. Algorithm 5 coordinates this multi-algorithm approach through systematic training, prediction aggregation, and consensus formation.

To enhance the ensemble’s robustness, each base model is assigned a model reliability score, which reflects its predictive performance on a validation set. In our implementation, the reliability score $w_i$ for classifier i is proportional to its cross-validation accuracy and normalized over all base models, as given in Eq. (11):

$w_i=\frac{C V_{\text {score_i }}}{\sum C V_{\text {scores }}}$    (11)

During ensemble prediction, each classifier contributes to the final decision based on its weight $w_i$ and its predicted probability $p_i$ for each class. The final prediction is determined by aggregating the weighted probabilities and selecting the class with the highest total score, as shown in Eq. (12):

$Final~Prediction =arg~max \left(\sum w_i \times P_i\right)$    (12)

This performance-based weighted voting ensures that models demonstrating higher reliability have greater influence on the ensemble decision, while models with lower accuracy contribute less. Consequently, the approach leverages the collective intelligence of multiple algorithms, improving overall classification accuracy and providing more stable and interpretable confidence scores.

3.5 Optimization and validation

This critical phase implements comprehensive evaluation protocols to ensure optimal performance and robust generalization of the trained ensemble model. The optimization process operates through three coordinated activities: cross-validation for generalization assessment, hyperparameter tuning for performance maximization, and result analysis for model selection and validation.

Cross-validation evaluates generalization capability through systematic data partitioning into training and test sets, preventing overfitting while enabling robust performance assessment. Hyperparameter tuning employs GridSearchCV to optimize each base model's configuration, systematically exploring parameter spaces for all ensemble components including traditional machine learning models and BERT fine-tuning parameters. Performance analysis evaluates models across multiple dimensions using precision, recall, F1-score, and accuracy metrics, while confusion matrices reveal detailed classification patterns and semantic coherence assessment measures alignment with ontological relationships.

The validation process compares weighted voting, majority voting, and stacking approaches to identify optimal ensemble configurations. Final model selection balances classification accuracy with semantic consistency and computational efficiency, ensuring practical applicability while maintaining the ontology-enhanced performance advantages that distinguish our approach from traditional classification approaches.

3.6 Prediction phase

After thorough optimization and validation, the SHADO framework moves into operational deployment for real-world document classification. In this phase (Figure 3), a streamlined prediction pipeline is activated to process new, unseen documents by applying a predefined sequence of steps established during training.

Figure 3. Prediction phase workflow for new document classification

Unlike the training phase, supervised ontology-based enrichment is not applied here, since the document’s class is not yet known. The system instead performs class-independent operations, including text preprocessing, feature extraction (TF-IDF and general semantic weighting if applicable), dimensionality reduction using pre-trained transformation matrices, and final classification through the optimized ensemble model.

The system delivers comprehensive classification results including definitive class assignments representing the ensemble's highest confidence predictions and detailed confidence scores indicating prediction reliability. These confidence metrics prove valuable for applications requiring threshold-based decision making or scenarios where multiple potential classifications need consideration. Additional outputs include individual model predictions for transparency. This comprehensive output supports both automated decision-making and human interpretation, ensuring practical applicability across diverse domain-specific classification tasks while maintaining the semantic richness that distinguishes ontology-enhanced classification from traditional approaches.

In this section, we present a comprehensive analysis of SHADO’s experimental results across multiple domains, discuss comparative performance metrics, and highlight the framework’s impact on advancing semantic-driven text classification.

5.1 Results

This section provides a detailed experimental evaluation of the proposed SHADO framework. It includes a domain-wise performance assessment with semantic coherence metrics across five thematic areas, results obtained using baseline machine learning methods, an analysis of the confusion matrix generated by SHADO, and a comparative evaluation against recent state-of-the-art ontology-enhanced classification approaches published between 2023 and 2025.

5.1.1 SHADO performance evaluation

A comprehensive evaluation shows that SHADO performs better than traditional classification methods in all domains. The ontology-based approach achieves high accuracy and preserves semantic meaning. Table 7 summarizes the results, including both standard metrics and measures of semantic coherence across the five domains.

Table 7. SHADO performance summary across domains

The results in Table 7 indicate consistently strong performance across all five domains, with F1-scores above 97% and only minor variations between domains. The Medical and Technology categories exhibit slightly higher accuracy and F1-scores, which can be attributed to the richness and structural completeness of their ontological resources.

Notably, SNP and ORR scores remain high (average ≈ 0.89 and 0.91, respectively), confirming that the learned embeddings preserve both semantic proximity (SNP) and ontological relationships (ORR). A cross-domain correlation analysis shows that domains with higher SNP/ORR values tend to achieve higher F1-scores (Pearson r = 0.82 for SNP, r = 0.79 for ORR), indicating that maintaining semantic coherence in the representation space directly enhances classification reliability.

5.1.2 Performance results using baseline methods

SHADO’s performance is presented alongside results obtained from standard classification algorithms applied to the same preprocessed datasets. These baseline models represent traditional approaches with semantic enrichment and ontology integration. Table 8 reports the outcomes of these methods, allowing the effectiveness of SHADO’s ontology-driven framework to be contextualized in relation to established techniques.

Table 8. Comparative performance analysis

To ensure a fair comparison, we further evaluated the baseline classifiers on the same datasets without semantic or ontological enrichment, using only lexical features after standard preprocessing (Algorithm 1). This complementary analysis isolates the contribution of the ontology-based enrichment introduced in SHADO.

Table 9. Comparative performance of baseline models without semantic enrichment

The results in Table 9 reveal a notable improvement of +4.6% in average accuracy when semantic enrichment and ontology integration are applied (Table 8). This demonstrates that the observed performance gain of SHADO does not merely result from model architecture or hyperparameter tuning, but from the semantic reinforcement provided by ontological grounding and contextual feature expansion. By comparing both settings (with and without enrichment), we confirm that SHADO’s ontology-driven representation significantly enhances document understanding, leading to higher stability (CV Score) and better generalization across domains.

5.1.3 Ablation study: Component-wise contribution analysis

To further quantify the contribution of each component within the SHADO framework, we conducted an ablation study. This analysis isolates the effect of three major modules: (i) ontology-based enrichment, (ii) semantics-preserving dimensionality reduction, and (iii) hybrid ensemble aggregation. Each variant was tested under identical experimental settings using the same preprocessed corpus.

Table 10. Ablation analysis of key ATCIADO components

Results in Table 10 clearly show that each module contributes progressively to the overall performance of SHADO. Ontology enrichment alone yields a +2.7% accuracy improvement by embedding domain semantics and resolving lexical ambiguity. The semantics-preserving reduction adds a further +1.1% gain, demonstrating the benefit of structure-aware compression. Finally, integrating the hybrid ensemble boosts both accuracy and robustness (+0.9%), validating the synergy between diverse learners.

These findings confirm that SHADO effectiveness emerges from the cumulative interaction of its components, rather than from a single module.

5.1.4 Evaluation against recent state-of-the-art approaches

To show that SHADO competes well with recent methods, we compared it with the latest ontology-enhanced and transformer-based approaches from 2023 to 2025. This highlights where SHADO fits in current research and what makes it stand out. Table 11 provides a detailed comparison with recent state-of-the-art models, showing SHADO's strong performance across various domains.

Table 11. Comparative discussion with recent ontology-enhanced approaches

Beyond quantitative metrics, Table 11 outlines the qualitative distinctions that set SHADO apart from prior ontology-enhanced systems.

Unlike earlier classifiers such as Bouchiha et al. [17] and Yelmen et al. [24], which employed fixed ontology mappings, SHADO dynamically reconfigures semantic relations through contextual graph propagation and semantic centrality weighting. This adaptability deepens semantic reasoning while preserving computational efficiency (O(n log n) for ontology updates).

Compared with Li et al. [37] and Ali et al. [40], SHADO achieves a superior balance between lexical precision and conceptual abstraction, enabled by its hybrid lexical–semantic scoring (α = 0.7).

Although the model of Ngo et al. [32] attained competitive accuracy, its static ontology and higher processing cost limit its interpretability.

In contrast, SHADO integrates explainable ontology-driven constraints, offering transparent decision paths and stronger domain alignment—qualities essential for real-world expert systems.

5.1.5 Confusion matrix of the proposed SHADO approach

To illustrate SHADO’s classification performance, Figure 7 presents the confusion matrix across five target domains. It highlights correct classifications, residual errors, and provides a concise diagnostic view of the model’s precision and consistency.

Figure 7. Confusion matrix

5.2 Discussions

The experimental results confirm the effectiveness of SHADO in leveraging ontologies to enhance text classification. Achieving F1-scores above 97% across all domains demonstrates the robustness of the framework, supported by rich ontological resources. High scores in SNP (0.884–0.901) and ORR (0.907–0.923) validate the core hypothesis that semantic structures can be preserved throughout the classification pipeline. This highlights not only strong predictive accuracy but also semantic coherence essential for interpretability in domain-specific applications. The framework demonstrates exceptional semantic preservation capabilities, with medical domain achieving the highest SNP (0.901) and ORR (0.923) scores, reflecting the rich ontological resources available in healthcare.

Cross-validation analysis reveals strong model stability, with CV scores consistently above 0.90 across all ensemble components, indicating robust generalization capabilities and minimal overfitting. The alignment between CV scores and test performance validates the framework's reliability for real-world deployment. Notably, the correlation between semantic preservation metrics and classification performance suggests that domains with richer ontological structures benefit more from the SHADO approach, while domains with sparse or ambiguous semantic relationships present greater challenges for ontology-enhanced classification.

Compared to recent ontology-based approaches, SHADO stands out for its consistent high performance across multiple domains—thanks to its adaptive ontology selection and semantic preservation strategies—while many other methods focus on single-domain optimization.

Two key innovations support this performance: (1) semantically constrained dimensionality reduction, which preserves ontological relationships during feature transformation, and (2) a hybrid ensemble architecture that combines traditional algorithms with transformer models, using ontology-informed weighting.

Finally, the SNP and ORR metrics offer valuable indicators for real-world deployment, especially in domains requiring explainable decisions. SHADO thus combines accuracy, semantic integrity, and practical applicability, marking a promising direction for hybrid approaches in text classification.