Deep Ontology Alignment Using a Natural Language Processing Approach for Automatic M2M Translation in IIoT

Saleha Javed, Muhammad Usman, Fredrik Sandin, Marcus Liwicki, Hamam MokayedSensors 2023 · DOI

TL;DR

This paper introduces a novel deep learning approach for automatically aligning ontologies of heterogeneous Industrial Internet of Things (IIoT) devices. By leveraging Natural Language Processing (NLP) techniques and knowledge graphs, the model enables seamless machine-to-machine (M2M) translation, paving the way for more robust and scalable IIoT systems.

What Problem Does This Paper Address?

The Industrial Internet of Things (IIoT) is rapidly expanding, connecting a vast array of devices with diverse underlying structures and communication protocols. This heterogeneity creates a significant challenge for interoperability, particularly in achieving seamless machine-to-machine (M2M) communication and translation. Traditional solutions often rely on costly, manual engineering of translators for each pair of ontologies, which is not scalable for dynamic IIoT environments. The paper addresses the critical need for an automated, self-learning approach to align these disparate ontologies, enabling devices to understand and communicate with each other without human intervention or specialized hardware.

Methods

The core of the proposed solution is a deep learning model that combines language and structural information from IIoT device ontologies. The approach involves:

  • Knowledge Graph (KG) Construction: Ontologies (specifically SSN and SOSA in this work) are represented as knowledge graphs, where entities (classes, instances) and relations form the graph structure.
  • Language Encoder: A modified BERT encoder is used to process the metadata, labels, and relations of the ontology as language. This generates language-based embedding vectors for each component.
  • Structural Encoder: This component takes the KG triplets (head, relation, tail) and encodes the structural information of the ontology. It extracts features like path length and connectivity to represent the graph's architecture.
  • Interaction Model: This model combines the language and structural embeddings to learn the alignment between entities from different ontologies. The paper proposes both supervised and unsupervised learning schemes for this interaction.
  • BERT-INT Modification: The model builds upon the BERT-INT approach but enhances it by incorporating structural information more effectively and modifying the input arrangement for the BERT encoder to better leverage its capabilities.

Datasets

  • SSN and SOSA Ontologies: These are the primary ontologies used to construct the knowledge graphs. They are W3C-standardized ontologies relevant to IoT sensor devices.
  • Smart Building System Dataset (Kaggle): This dataset, collected from 51 rooms in a smart building, was used to generate ontology instances for training and evaluation. It contains time-series sensor data (temperature, humidity, luminosity, motion).
  • DBP15K Dataset: This dataset, commonly used for cross-lingual entity alignment, was also utilized for evaluating the model's performance in a more general context.

Key Insights and Ideas

  • Bridging the NLP-IIoT Gap: The paper highlights a significant gap in research at the intersection of NLP and IIoT, particularly in ontology alignment. This work is a pioneering effort to bridge this gap.
  • Synergy of Language and Structure: The key insight is that combining both the linguistic (metadata, labels) and structural (graph topology) information of ontologies leads to more robust and accurate alignment.
  • Self-Learning for Scalability: The proposed self-learning model eliminates the need for manual engineering of translators, making it highly scalable for the dynamic and ever-expanding IIoT landscape.
  • Ontology as a Device's "Language": The paper frames ontologies as the "language" of devices, making NLP techniques a natural fit for understanding and translating between them.

Key Results

  • Improved Performance: The proposed model consistently outperformed the baseline BERT-INT model by approximately 1.2-2.7% in Hits@1, Hits@10, and MRR scores on the DBP15K dataset.
  • Ablation Study Effectiveness: An ablation study demonstrated that using both the language and structural encoders yielded the best results, confirming the importance of both information sources.
  • Qualitative Alignment Visualization: t-SNE plots and heatmaps were used to visualize the alignment process, showing how entities with similar semantic and structural properties were correctly mapped.

Limitations and Future Work

  • Dataset Scarcity: A significant limitation is the lack of publicly available, large-scale, real-world IIoT ontology datasets. The authors had to synthesize datasets, and the SSN and SOSA ontologies are relatively concise compared to language ontologies.
  • Subjective Annotations: For validation, human experts annotated the alignments, which introduces subjectivity. The paper calls for the creation of benchmark datasets for IIoT ontology alignment.
  • Generalization to Broader IIoT: The current work focuses on sensor devices. Future work could explore generalizing the approach to a wider range of IoT devices and ontologies.
  • Translation Module Extension: The proposed ontology alignment model is seen as a foundational step. Future work could extend it into a full M2M translation module that interprets messages directly from device ontologies.

Summary updated

BibTeX
@article{javed2023deep,
  author = {Saleha Javed and Muhammad Usman and Fredrik Sandin and Marcus Liwicki and Hamam Mokayed},
  title = {{Deep Ontology Alignment Using a Natural Language Processing Approach for Automatic M2M Translation in IIoT}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23208427},
}
Figure