Using terms and informal definitions to classify domain entities into top-level ontology concepts: An approach based on language models

Abstract

The classification of domain entities into top-level ontology concepts remains an activity performed manually by an ontology engineer. Although some works focus on automating this task by applying machine-learning approaches using textual sentences as input, they require the existence of the domain entities in external knowledge resources, such as pre-trained embedding models. In this context, this work proposes an approach that combines the term representing the domain entity and its informal definition into a single text sentence without requiring external knowledge resources. Thus, we use this sentence as the input of a deep neural network that contains a language model as a layer. Also, we present a methodology used to extract two novel datasets from the OntoWordNet ontology based on Dolce-Lite and Dolce-Lite-Plus top-level ontologies. Our experiments show that by using the transformer-based language models, we achieve promising results in classifying domain entities into 82 top-level ontology concepts, with 94% regarding micro F1-score.

Introduction

The development of domain ontologies based on top-level ontology concepts has proved valuable in many domains, such as geology [1], [2] and biomedicine [3], [4], [5]. One reason is that different domain ontologies developed under the same top-level ontology become semantically interoperable because their domain entities specialize in the same top-level concepts. However, the task of identifying which top-level concept a domain entity specializes is laborious and time-consuming because it is usually performed manually and requires a high level of expertise in both the target domain and ontology engineering [6].

In this work, we proposed an approach for classifying domain entities into top-level concepts using the combination of terms representing the domain entities and their informal definitions. Our approach uses terms and informal definitions because they provide the intended meaning of the domain entities in a particular domain and are provided early in the ontology development process. Thus, we combined both into a single text sentence. From this, we fed a deep-neural architecture that includes a pre-trained language model as a layer. After that, we output the predicted top-level concept. In addition, this work proposes a methodology to extract two novel datasets from the OntoWordNet ontology and the Dolce-Lite and Dolce-Lite-Plus top-level ontologies. Each resulting datasets contain 120,489 instances, where each instance has the term representing the OntoWordNet entity, its informal definition, and its respective Dolce-Lite or Dolce-Lite-Plus top-level concept as its target class.

In our experiments, we evaluated eight free available transformer-based language models in the proposed deep-neural architecture: BERT tiny, mini, small, medium, and base [7]; RoBERTa base [8]; ALBERT base [9]; and ELECTRA small [10]. We fine-tuned each model for our classification task during the training stage using an unbalanced and stratified sample. The experimental results suggest that using the combination of the term representing the domain entity and its informal definition with language models has promising results. In addition, the experiments show that the BERT-Base model achieves the best overall scores, with 94% and 87% in micro F1-score using the Dolce-Lite-Plus and Dolce-Lite datasets, respectively.

The paper is organized as follows. In Section 2, we present the background notions that support this proposal, revisiting the OntoWordNet ontology, the top-level ontologies of Dolce-Lite-Plus and Dolce-Lite, and the main proposals on language models and ontology learning. In Section 3, we describe the methodology for extracting datasets from existing domain ontologies and the proposed approach for classifying domain entities into top-level concepts using terms, informal definitions and language models. In Section 4, we show the experiments performed using the two extracted datasets and the obtained results. Finally, in Section 5, we present the concluding remarks of our work.