Bridging Textual Data and Conceptual Models: A Model-Agnostic Structuring Approach

We introduce an automated method for structuring textual data into a model-agnostic schema, enabling alignment with any database model. It generates both a schema and its instance. Initially, textual data is represented as semantically enriched syntax trees, which are then refined through iterative tree rewriting and grammar extraction, guided by the attribute grammar meta-model \metaG. The applicability of this approach is demonstrated using clinical medical cases as a proof of concept.

💡 Research Summary

The paper presents ArchiTXT, an automated framework that transforms unstructured textual data into a model‑agnostic schema and corresponding data instance, enabling seamless mapping to any underlying database technology (relational, NoSQL, graph, etc.). The approach begins by converting raw text into semantically enriched syntax trees using standard natural‑language processing components such as named‑entity recognizers and parsers. These trees are then subjected to a two‑fold refinement process: a top‑down enrichment phase that injects domain‑specific entities into the tree nodes, and a bottom‑up pruning/aggregation phase that restructures the tree by applying a set of tree‑rewriting rules.

The core theoretical contribution is the definition of a meta‑grammar G, an S‑attribute grammar that encodes global constraints for any target grammar G_T. G specifies the permissible shape of production rules, and a synthesized attribute γ validates each derivation (γ = true for a valid rule, false otherwise). Starting from an initial grammar G₀ that directly reflects the syntactic structure of the input sentences, the system iteratively evolves G₀ into G_T through tree rewriting guided by a similarity measure (e.g., lexical distance or semantic embeddings). This evolution is formally described as a sequence of transformations on both the trees (instances) and the grammars (schemas).

The resulting target grammar G_T is deliberately abstract: it defines only “groups”, “properties”, and “relations” without committing to any concrete data model. For example, from a clinical sentence describing an intravenous urography exam and associated symptoms, ArchiTXT extracts production rules such as

• Grp Exam → Prop examName Prop anat
• Grp Sosy → Prop sosyDesc Prop anat
• Rel ExamSosy → Grp Exam Grp Sosy
  These rules constitute a condensed context‑free grammar that can generate a tree instance I_T containing the essential fragments of the original text. Because G_T adheres to the meta‑grammar G, it is guaranteed to be syntactically and semantically coherent.

In a post‑processing step, G_T can be mapped to a concrete database schema. The authors illustrate two straightforward mappings: (i) a relational schema with tables EXAM, SOSY, and EXAM_SOSY, and (ii) a graph schema where EXAM and SOSY become nodes and EXAM_SOSY becomes an edge. The paper emphasizes that this mapping is not hard‑wired; any model‑specific translation layer can consume G_T because the intermediate representation is model‑agnostic by design.

The contributions are summarized as follows:

1. A novel method for structuring textual data that separates the concerns of information extraction and schema generation.
2. The introduction of a meta‑grammar G that enforces global constraints on generated schemas, ensuring model‑agnosticity.
3. A formalization of the evolution from G₀ to G_T via tree rewriting and similarity‑driven clustering.
4. A prototype implementation (ArchiTXT) evaluated on a set of clinical case reports, demonstrating the feasibility of automatic schema and instance generation.

Related work is positioned in three areas: (a) traditional information extraction and OpenIE, which focus on extracting triples but do not produce structured database instances; (b) data integration frameworks that map existing structured or semi‑structured data to a common conceptual model; and (c) model‑specific schema generation tools that aim to optimize for a particular NoSQL or relational target. ArchiTXT distinguishes itself by operating on raw text, using a hybrid approach that combines syntactic tree transformation with semantic enrichment, and by delivering a truly model‑agnostic intermediate representation.

The evaluation, however, is limited to a small corpus of medical case narratives. Quantitative metrics such as precision, recall, or processing time are not reported, and the scalability of the tree‑rewriting engine remains untested on larger datasets. Moreover, the paper does not provide explicit definitions of the rewriting rules or the similarity functions, which hampers reproducibility. The reliance on external NER and parsing tools also introduces a dependency on their accuracy; errors in entity detection propagate to the final schema.

Future work outlined by the authors includes enriching the meta‑grammar with domain‑specific constraints (e.g., medical coding standards), developing an automatic model‑selection component that chooses the most suitable database technology based on the structural characteristics of G_T, and extending the prototype to handle larger, more heterogeneous corpora. Additional research directions could involve integrating user feedback loops for schema refinement, providing semantic naming for generated groups and properties, and benchmarking against existing ontology‑learning pipelines.

In conclusion, ArchiTXT offers a compelling proof‑of‑concept that bridges the gap between unstructured textual narratives and structured database representations through a formally grounded, model‑agnostic pipeline. By decoupling the extraction of semantic content from the choice of physical data model, it promises greater flexibility, transparency, and interoperability for applications that must ingest and analyze large volumes of text across diverse domains.