Home // International Journal On Advances in Software, volume 19, numbers 1 and 2, 2026 // View article

Extracting Entity Relationship Diagrams from Unstructured Text - A Hybrid Approach

Authors:
Vaihunthan Vyramuthu
Gregor Grambow

Keywords: Entity-Relationship Model; Natural Language Processing; Named Entity Recognition; POS-Tagging; SpaCy; LSTM; Transformer; Imbalanced Data

Abstract:
Data models are a crucial part of computer science applications that automate real-world processes. An important model for abstractly capturing data sets and their relations is the Entity-Relationship (ER) model. To create these models, real-world entities and their relations have to be abstracted. However, the transformation from real world descriptions in natural language to standardized ER diagrams can be tedious and error-prone. Recently, Natural Language Processing (NLP) has gained much attention, but this specific area is still mostly handled manually by humans. With this contribution, we aim at automating the time-consuming interpretation of textual database scenarios by describing a hybrid system to capture components of the ER model from German texts using NLP. This hybrid system combines rule-based and model-based approaches. This may lead to a more robust and reliable extraction, as errors in one of the approaches can be compensated by the other. We implemented and tested both approaches, whereas the main extraction is performed by the rule-based variant. The results of the model-based approach are used as a comparison to the rule-based results and can be applied to check the correctness and improve the results. We conducted an evaluation using a set of real-world scenarios describing various entities and their relationships and applied a set of metrics suited to the imbalanced nature of the dataset. Our evaluation shows promising results indicating that a hybrid approach can deliver better results than classical approaches.

Pages: 28 to 39

Copyright: Copyright (c) to authors, 2026. Used with permission.

Publication date: June 30, 2026

Published in: journal

ISSN: 1942-2628