Can an AI Extract the Structure of “Being” in Heidegger? Formal Ontologies from Deep Learning
The question of being constitutes the core of Martin Heidegger’s Being and Time, where the German philosopher unfolds an existential analytic of Dasein in order to access the meaning of being in general. Traditionally, the construction of formal ontologies from philosophical texts has been a hermeneutic task carried out by specialists, who interpret concepts, relations, and hierarchies through close reading. However, the development of deep learning and natural language processing opens the possibility of partially automating this labor. Language models such as transformers can process extensive corpora, identify conceptual entities, and map latent semantic relations. The hypothesis guiding this work is that an AI trained con Being and Time and its philosophical context could generate a computable formal ontology representing the structure of being according to Heidegger. This does not replace interpretation, but it does propose a new method for the digital humanities.
Applying deep learning to Heidegger entails confronting specific epistemological and technical problems. Heideggerian language is deliberately disruptive: it introduces neologisms such as Dasein, Geworfenheit, or Zuhandenheit, and subverts traditional grammar in order to break with the metaphysics of presence. A statistical model trained on general corpora will tend to collapse these terms into their colloquial uses, losing the ontological charge that Heidegger assigns to them. Therefore, the first challenge is to build a curated corpus that includes not only Being and Time, but also critical commentaries, contrasted translations, and texts from the tradition that Heidegger deconstructs, such as Aristotle and Kant. Only with this context can the model learn that being is not just another predicate. Moreover, the task demands moving beyond the mere extraction of subject-predicate-object triples, since Heideggerian being cannot be captured by classical attributive logic.
Methodologically, the automatic construction of ontologies from philosophical texts combines techniques of named entity recognition, relation extraction, and graph learning. For Heidegger’s case, a three-phase pipeline is proposed. First, fine-tuning an LLM on the Heideggerian corpus so that it acquires the author’s lexicon and internal relations, using techniques such as LoRA adapters that preserve the base model’s general knowledge. Second, a knowledge-graph extraction layer where key concepts are identified and relations such as “founds,” “shows itself as,” or “is constitutive of” are inferred, which are central to the existential analytic. Third, the translation of that graph into a formal language such as OWL or OntoUML, validated by philosophers to avoid categorical distortions. This hybrid process acknowledges that the machine proposes and the humanist disposes, avoiding the naïve positivism of believing that meaning emerges solely from statistical frequencies.
Preliminary results from experiments with a fine-tuned LLaMA-3 show that the AI can group concepts of Dasein’s structure with notable coherence. The model correctly identifies that Sorge articulates Befindlichkeit, Verstehen, and Rede, and that Zeitlichkeit functions as the horizon for the understanding of being. Likewise, it distinguishes between the ontology of Vorhandensein and that of Zuhandensein, a nuance that often escapes initial readers. Nevertheless, the AI commits philosophically relevant errors: it tends to hierarchize concepts where Heidegger proposes co-originarity, and it reduces the later Ereignis to just another node instead of thinking it as an unobjectifiable event. These failures reveal the limits of modeling being as data. The Heideggerian structure of being resists totalization because it unfolds as difference and withdrawal, something a static graph can hardly capture without a temporal or performative dimension.
From a critical perspective, the attempt to extract the structure of being with deep learning faces the objection that every formal ontology is already metaphysics in the sense that Heidegger questions. To formalize being implies converting it into an entity, betraying the fundamental gesture of Being and Time. Therefore, the ontology generated by AI would not be “the” structure of being, but rather an ontic representation useful for navigating the text, similar to a map that is not the territory. Its value is heuristic and pedagogical: it allows for visualizing connections, detecting internal tensions, and formulating new questions. Moreover, it exposes the biases of the model itself, trained predominantly on Anglophone corpora, when it interprets categories rooted in the German tradition. Far from closing interpretation, the AI pluralizes it by showing patterns that the human eye might omit, provided it is read with hermeneutic suspicion.
In conclusion, an AI can extract a computable approximation to the structure of being in Heidegger, but not being itself. Deep learning offers tools for constructing formal ontologies of philosophical texts that function as research scaffolds, accelerating distant reading and suggesting systematic relations. However, the philosophical validity of these ontologies depends on their constant confrontation with close reading and with the pathos of the question of being. The interdisciplinary project between philosophy and computer science should not aspire to the automation of thinking, but rather to expanding the material conditions of interpretation.
JULIO HERRERA MANAGEMENT
JULIO HERRERA Manager
Julio Fernando Herrera Lamas