Data augmentation is a fundamental technique in machine learning to enhance model generalization by artificially expanding training datasets. However, conventional augmentation approaches often rely on heuristic transformations that may not fully capture domain-specific knowledge. This position paper advocates a data-centric AI perspective on data augmentation, emphasizing the integration of semantic technologies, particularly domain ontologies, to guide augmentation strategies. The use of techniques from Symbolic AI for data augmentation has been dealt with only in a few recent papers. Our goal is to explore further this idea, based on the consideration that an explicit representation of the domain may be helpful in two key tasks: optimizing the generation of new data, and validating the generated data, both fundamental steps for all data augmentation strategies. We aim at developing novel approaches that combine ontologies and data augmentation techniques to address these two tasks, in particular by relying on automated reasoning. We argue that leveraging knowledge representation and symbolic reasoning enables more principled and context-aware data augmentation, leading to improved model robustness and fairness.
In recent years, machine learning (ML) has emerged as a transformative technology across a wide
range of domains, from healthcare and finance to autonomous systems and natural language
processing. A critical factor in the success of ML models is their ability to generalize well to
unseen data, which is heavily influenced by the quality and diversity of the training datasets. In
this framework, data augmentation (DA) has become a fundamental technique to enhance model
generalization by artificially expanding training datasets through data transformations [
This position paper advocates for a novel data-centric AI perspective on data augmentation,
emphasizing the integration of semantic technologies, particularly domain ontologies, to guide
augmentation strategies. While research on data augmentation is far from new, the coupling
with semantic technologies remains largely unexplored. Although some studies employ
ontologies in the data augmentation process [
The integration of domain ontologies into data augmentation processes ofers several
advantages. First, ontologies provide a structured and formal representation of domain knowledge,
enabling the generation of data that is semantically consistent with the underlying domain.
Second, ontologies can facilitate the validation of augmented data by providing a framework for
automated reasoning, ensuring that the generated data adheres to domain-specific constraints
and rules. Moreover, the use of symbolic reasoning in conjunction with data augmentation opens
up new possibilities for addressing challenges related to model interpretability and fairness [
This work is part of a wider project, called ENDURANCE, which is a builds upon seminal
work in data provenance [
The rest of the paper is organized as follows. In Section 2 we review the main DA methods proposed in the literature. In Section 3, we introduce ontologies and describe the role they play in data management. Then, in Section 4, we depict the approaches we aim at exploring in order to exploit ontologies for DA, while in Section 5 we conclude the paper by discussing future work.
DA methods can be broadly categorized into three main approaches, though it is important to
acknowledge that many techniques draw inspiration from multiple domains, often blending
ideas across these categories [
Heuristic or rule-based augmentation methods rely on manually defined transformations, making them relatively simple to implement and computationally inexpensive. However, their efectiveness is often dependent on the expertise of domain specialists who tailor the transformations to the specific dataset. While this approach can yield useful variations of data, it may struggle to introduce the level of diversity needed to significantly enhance model performance.
A more dynamic class of techniques is mixing-based augmentation, which generates new
training samples by interpolating between two or more existing examples. Originally designed
for image processing, these methods have since been adapted for NLP, time series, and tabular
data. However, the discrete nature of language tokens and structured data poses challenges,
as direct linear interpolation is often not feasible. To circumvent this issue, mixing-based
approaches typically operate within a continuous embedding space [
Recent methods involve deep models explicitly trained to generate new data. Techniques
such as variational autoencoders (VAEs) [
Each of these approaches ofers distinct advantages and trade-ofs. While rule-based methods provide control and interpretability, they may lack the richness of data needed for more complex learning tasks. Mixing-based methods ofer an eficient way to increase sample diversity but require careful adaptation for discrete data types. Deep generative models, on the other hand, represent the cutting edge of DA, capable of synthesizing high-quality data, though they require significant computational resources and robust safeguards to maintain reliability. As research in this area continues to evolve, the interplay between these approaches is likely to shape the next generation of DA strategies.
DA in computer vision (CV) aims to enrich image datasets with diverse, label-consistent, variations enabling exposure to previously unseen visual patterns.
Rule-based Transformations. Simple geometric or photometric operations (e.g., random
lfipping, cropping, rotation, color jittering) preserve semantic labels yet diversify the input
distribution. These lightweight edits are ubiquitous in CNN training pipelines, ofering consistent
improvements with minimal overhead [
Noise and Erasing. Injecting noise into the image matrix (e.g., Gaussian noise injection [16])
or masking random regions, similar to dropout regularization (e.g., Cutout [17]) encourages
models to develop more robust feature representations. Variants of these techniques incorporate
content-aware erasing, allowing for more controlled manipulation of hidden information.
Mixing Strategies. Inspired by Mixup [18], several methods overlay or interpolate pairs of
images. This smooths the decision boundary by forcing the model to learn from “mixed” samples
(e.g., partial dog, partial cat). Such mixing may reduce overfitting and improve adversarial
robustness. Variants like CutMix [19] combine random patches from diferent images, efectively
blending object context while encouraging the network to leverage global cues.
Deep Generative Models. These methods utilize powerful generative models, such as GANs
and difusion-based approaches, to create realistic synthetic images while preserving essential
features like class labels [
DA for natural language focuses on modifying text corpora by altering syntax while maintaining semantic integrity.
Rule-based Transformations. Simple lexical edits like synonym substitution or random swaps expand a dataset with minimal overhead. For example, EDA [20] replaces or deletes words to yield new training samples while preserving semantics—often enough to boost text classification performance under low-resource conditions.
Embedding-based Mixup. Adapted from computer vision’s Mixup [18], methods like MixText
[
Backtranslation. A popular sequence-to-sequence approach, backtranslation [21] passes text through two or more translation steps (e.g., English→Italian→English). The resulting paraphrases have altered surface forms yet retain original meaning, yielding label-consistent diversity valuable in low-resource or domain-shift scenarios.
Language Models Generative architectures such as GPT [22] create synthetic text samples by fine-tuning on specific labels or topics, expanding the linguistic diversity of training data. Similarly, sequence-to-sequence models like T5 [23] facilitate augmentation through tasks such as paraphrasing and controlled text transformations. Meanwhile, masked language models (e.g., BERT [24]) contribute to DA by replacing tokens with contextually plausible alternatives, improving coverage of rare patterns and linguistic styles. More recently, large language models (LLMs) such as GPT-3.5 and LLaMA have demonstrated advanced capabilities in generating lfuent and coherent text with minimal prompting, making them valuable tools for NLP DA. Their ability to generate highly context-aware transformations introduces new possibilities for enriching training corpora, though challenges such as computational costs and potential biases in generated content must be carefully managed.
Tabular data are widely used across various domains, typically found as web tables, databases,
or spreadsheets, containing both numerical and categorical features. Due to the discrete and
heterogeneous nature of tabular data, any augmentation procedure must adhere to
domainspecific constraints, such as column types, positivity constraints, and permissible value ranges
[
Some research has proposed end-to-end pipelines for tabular DA (TDA), incorporating
multiple stages to ensure high-quality augmentation [
Row-level Augmentation (Record Generation): New rows can be generated using models
that sample from learned data distributions, such as SMOTE [25], Gaussian Mixture Models, or
GANs. These methods help mitigate class imbalance or data scarcity by generating additional
training examples that preserve the statistical properties of the original data [
In Computer Science, "ontology" is a specific term denoting an artifact that is designed for enabling the modeling of knowledge about some domain. In Artificial Intelligence, an ontology is a symbolic representation of a domain interest, a formal tool used to capture and reason over an explicit specification of the domain knowledge [26].
There is actually a long history of this notion. The early "semantic networks" [27] and "frame systems" [28] proposed and studied in the ’70s are the first examples of ontology formalisms developed with the rise of symbolic artificial intelligence (AI) and structured knowledge representation languages.
In the ’80s, the research on object-oriented languages and conceptual data modeling borrowed several principles from AI and adapt them towards the goal of devising new methodologies for software development and database design. The common aspect of these formalisms are constituted by representational primitives with which to model a domain, namely classes (or concepts), attributes (or properties), and relationships (or relations among class members).
In the ’90s, the research on Description Logics [29] provided solid logical foundations of ontologies, continuing the work started by Brachman and Levesque in 1986 [30], and carrying out a large body of research with the goal of understanding how the expressiveness of the class definition mechanisms used in a given Description Logic influences the feasibility and the efectiveness of automated reasoning algorithms. After more than three decades, we have now a global picture of the expressiveness/complexity trade-of in Description Logics.
In recent years ontologies are advocated as tools to provide a semantic layer for data integration, knowledge graphs and machine learning, with the goal of improving efectiveness, quality and explainability of several AI tasks.
Ontology-Based Data Management (OBDM) [31] is a well-founded paradigm aiming at enriching data with semantics, so as to rely on ontology reasoning for data management. Specifically, an OBDM system is a three-layered architecture comprising an ontology, a set of autonomous data sources and a so-called mapping establishing the relationship between the two. Given that within this setting data are typically very large, literature on OBDM has mainly focused on the study of forms of reasoning over the data that are tractable with respect to the data.
While traditional reasoning over ontologies focused on inferring schema-level properties, OBDM introduces the issue of using the conceptual level to access and manipulate data. For this reason, the most studied form of reasoning in this setting is query answering, i.e. the problem of logically deriving answers to queries, based on the knowledge expressed in the ontology and the mappings. Notably, query answering is the basis for other relevant tasks in OBDM, such as data quality checking [32, 33] and privacy preserving query processing.
In general, OBDM resorts to symbolic techniques to infer (new) insights from data, as opposed to ML, that induces insights from data through sub-symbolic techniques.
As discussed in section 2, current DA techniques can be categorized into two broad classes: rule-based or sub-symbolic, i.e., based on sub-symbolic AI approaches. As each approach shows strengths that could help mitigating the other’s weaknesses, we argue that a more robust DA framework should mix together techniques from both approaches. Specifically, taking into account the semantics of the application domain defined by ontological axioms may improve two fundamental steps performed via purely sub-symbolic DA techniques: optimizing the generation, and strengthening the validation of synthetic data.
In what follows, we propose several novel approaches to combine ontologies and sub-symbolic DA techniques to address these two tasks. A crucial assumption for all these techniques is that ontological axioms and training data are compatible, i.e., we can perform standard reasoning tasks over the specification defined by the data and the axioms. For this purpose, we envision the use of specifically crafted mappings, able to reconcile ML models and domain knowledge in the ontology.
Additionally, we assume that ontological axioms are always able to define correct and meaningful information on the domain of interest. Due to the specific use cases we are considering, this may require the use of expressive ontological languages that can handle probabilistic rules. Generation The first task we aim to tackle is the optimization of the data generation process. In this context, we discuss two diferent approaches in which an ontology is used to influence the data generation process. For this purpose, let the generative process be denoted by a function (·), called generator, which outputs new unseen data samples.
• Ontology-guided generation: The ontology may be used to process the output of (·) and improve it according to its axioms. In particular, one could use an ontology to verify the consistency of the generated samples against the domain knowledge possessed by experts and xfi possible inconsistencies. A similar approach is proposed by [ 34] where logical constraints are incorporated in the output layer of a neural network. However, such constraints are expressed purely in terms of formulae over the attributes of the data samples. In the approach we propose, on the contrary, we want to formalize these constraints using a higher level of abstraction, i.e. using the symbols defined by an ontology. This will allow to perform reasoning over the domain, possibly inferring new knowledge concerning the generated sample, and therefore further augmenting the data. • Ontology as the generator: Another possible approach is to use the ontology as a set of rules that define the behavior of the generator (·) itself. In this context, the ontology could be used to generate data samples that are either already consistent with the domain knowledge or even partially consistent. In the latter case, we could also exploit the ontology to measure the inconsistencies and eventually address them.
Validation Performing a validation step on the output of a generative technique is a common practice that has been shown to improve the quality of the output. Usually, given a generator (·) , its validation involves training a binary classifier called discriminative model, or discriminator, (·) . This discriminator is trained to distinguish between real samples, i.e., those close enough to the original dataset, and synthetic samples, i.e., those that that lack of some defining features. Combining this technique with ontologies could lead to improved validation results. • Ontology-enhanced validation: this approach employs the ontology to improve the performance of a given discriminator (·) , i.e., providing a more accurate classification. To this end, we propose two alternative strategies: ontology-first and discriminator-first. The ontology-first strategy consists in applying the ontology rules to the input of (·) , before the classification is computed. The intuition behind this approach is that ontological axioms could be used to infer additional knowledge on the data samples. This knowledge could enhance the input of (·) thus making the classification more precise. Adding this information to the data samples, however, may require the modification of the the architecture of (·) to ensure compatibility. In contrast, the discriminator-first approach applies the ontological axioms to the output of (·) , verifying its consistency with the domain knowledge, without interfering with the model architecture. For example, a sample misclassified as real by (·) , could get the correct label synthetic after verifying the violation of some domain rules. • Ontology as the discriminator: another novel approach we propose is the use of an ontology as the discriminator (·) itself. Instead of training a model to recognize synthetic data, the ontology could be used to directly validate the output of a generator (·) checking consistency directly against domain knowledge.
Combining generation and validation in an online approach In the previous paragraphs we described techniques to employ ontologies in an ofline fashion, i.e., given and . A natural, and particularly interesting, extension of our study is to investigate online approaches where ontologies are integrated in the training phase of generative and discriminative machine learning models. In particular, we propose to use this approach in conjunction with GAN models, in which ontologies could be used to guide the learning process of both and .
Some of the most notable approaches in the literature are based on compiling the semantic constraints into the loss function of the training algorithm [35, 36], or directly into the model structure [37, 38]. However, none of these approach resort to an ontology to model the domain knowledge, which is at the core of our proposal. We argue that leveraging domain ontologies in the training phase of both the generative model (·) and the discriminative model (·) of a GAN leads to a context-aware generation of data, and therefore to improve DA.
This paper advocates for integrating domain ontologies into DA to address the limitations of conventional heuristic approaches in knowledge-intensive ML applications. By formalizing domain knowledge through semantic technologies, ontology-driven DA enhances both data generation and validation processes.
Three key findings emerge: (1) ontologies enable semantically consistent training data expansion, overcoming the domain-agnostic nature of traditional transformations like rotations or noise injection; (2) automated reasoning via ontological constraints ensures augmented data validity, preventing semantic inconsistencies that could degrade model performance, and (3) explicit knowledge representation supports fairness and interpretability in high-risks domains like healthcare by mitigating biases inherent in purely statistical DA methods.
The proposed framework bridges symbolic AI and data-centric paradigms, ofering systematic
strategies to align synthetic data with domain-specific rules. To validate the efectiveness of
the proposed approach, the project will conduct experimental analysis across multidisciplinary
scenarios encompassing three distinct use cases:
Entity Resolution: The first use case will involve entity resolution tasks on structured and
semi-structured data [
NLP Scenario: In collaboration with the Senate of the Italian Republic, the second use case will focus on clustering textual data (law amendments) [41], exploring how ontologies can support generative approaches for text generation and NLP tasks.
Hand Writer Identification: The third use case will address the writer identification problem for medieval manuscripts [42]. This use case in the digital humanities will test the applicability of DA techniques to image-based tasks. Tasks involving images may benefit from an online approach to augmentation as we discussed in the previous sections.
Future work should explore: (1) neuro-symbolic integration combining ontological reasoning with neural data generation, (2) constraint-based augmentation using ontological rules, and (3) automated validation pipelines verifying semantic consistency across augmentation cycles.
[15] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative adversarial networks, Communications of the ACM 63 (2020) 139–144. [16] F. J. Moreno-Barea, F. Strazzera, J. M. Jerez, D. Urda, L. Franco, Forward noise adjustment scheme for data augmentation, in: 2018 IEEE symposium series on computational intelligence (SSCI), IEEE, 2018, pp. 728–734. [17] T. DeVries, G. W. Taylor, Improved regularization of convolutional neural networks with cutout, arXiv preprint arXiv:1708.04552 (2017). [18] H. Zhang, M. Cisse, Y. N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization, arXiv preprint arXiv:1710.09412 (2017). [19] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, Y. Yoo, Cutmix: Regularization strategy to train strong classifiers with localizable features, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6023–6032. [20] J. Wei, K. Zou, Eda: Easy data augmentation techniques for boosting performance on text classification tasks, arXiv preprint arXiv:1901.11196 (2019). [21] R. Sennrich, B. Haddow, A. Birch, Improving neural machine translation models with monolingual data, arXiv preprint arXiv:1511.06709 (2015). [22] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al., Language models are unsupervised multitask learners, OpenAI blog 1 (2019) 9. [23] C. Rafel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, P. J. Liu, Exploring the limits of transfer learning with a unified text-to-text transformer, Journal of machine learning research 21 (2020) 1–67. [24] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 2019, pp. 4171–4186. [25] N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, Smote: synthetic minority over-sampling technique, Journal of artificial intelligence research 16 (2002) 321–357. [26] N. Guarino, D. Oberle, S. Staab, What is an ontology?, Handbook on ontologies (2009) 1–17. [27] W. A. Woods, What’s in a link: Foundations for semantic networks, in: Representation and understanding, Elsevier, 1975, pp. 35–82. [28] M. Minsky, A framework for representing knowledge, 1974. [29] F. Baader, W. Nutt, Basic description logics, in: The description logic handbook: theory, implementation, and applications, 2003, pp. 43–95. [30] R. J. Brachman, H. J. Levesque, Readings in knowledge representation, Technical Report,
AT and T Bell Labs., 1985. [31] A. Poggi, D. Lembo, D. Calvanese, G. De Giacomo, M. Lenzerini, R. Rosati, Linking data to ontologies, in: Journal on data semantics X, Springer, 2008, pp. 133–173. [32] C. Daraio, M. Lenzerini, C. Leporelli, P. Naggar, A. Bonaccorsi, A. Bartolucci, The advantages of an ontology-based data management approach: openness, interoperability and data quality, Scientometrics 108 (2016) 441–455. [33] M. Console, M. Lenzerini, Data quality in ontology-based data access: The case of consistency, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 28, 2014. [34] N. Hoernle, R. M. Karampatsis, V. Belle, K. Gal, Multiplexnet: Towards fully satisifed logical constraints in neural networks, 2021. URL: https://arxiv.org/abs/2111.01564. arXiv:2111.01564. [35] J. Xu, Z. Zhang, T. Friedman, Y. Liang, G. V. den Broeck, A semantic loss function for deep learning with symbolic knowledge, 2018. URL: https://arxiv.org/abs/1711.11157. arXiv:1711.11157. [36] M. Fischer, M. Balunovic, D. Drachsler-Cohen, T. Gehr, C. Zhang, M. Vechev, DL2: Training and querying neural networks with logic, in: K. Chaudhuri, R. Salakhutdinov (Eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, PMLR, 2019, pp. 1931–1941. URL: https://proceedings.mlr.press/v97/fischer19a.html. [37] G. G. Towell, J. W. Shavlik, M. O. Noordewier, Refinement of approximate domain theories by knowledge-based neural networks, in: Proceedings of the Eighth National Conference on Artificial Intelligence - Volume 2, AAAI’90, AAAI Press, 1990, p. 861–866. [38] A. Daniele, L. Serafini, Neural networks enhancement with logical knowledge, 2021. URL: https://arxiv.org/abs/2009.06087. arXiv:2009.06087. [39] V. Di Cicco, D. Firmani, N. Koudas, P. Merialdo, D. Srivastava, Interpreting deep learning models for entity resolution: an experience report using lime, in: Proceedings of the second international workshop on exploiting artificial intelligence techniques for data management, 2019, pp. 1–4. [40] R. Fagin, P. G. Kolaitis, D. Lembo, L. Popa, F. Scafoglieri, A Framework for Combining Entity Resolution and Query Answering in Knowledge Bases, in: Proceedings of the 20th International Conference on Principles of Knowledge Representation and Reasoning, 2023, pp. 229–239. [41] A. Sajeva, S. Iannucci, C. Marchetti, P. Merialdo, R. Torlone, Clustering amendments with semantic embeddings, SEBD 2024 (2024). [42] L. Lastilla, S. Ammirati, D. Firmani, N. Komodakis, P. Merialdo, S. Scardapane, Selfsupervised learning for medieval handwriting identification: A case study from the vatican apostolic library, Information Processing & Management 59 (2022) 102875.