A comprehensive understanding of a geological fault is crucial for the oil and gas industry, as it directly influences reservoir quality, potentially leading to leaks or maintaining a seal [1]. In addition to the oil and gas industry, faults also play an essential role in mining, geothermal, and construction industries [2]. However, the geological data required for interpretation is often scattered across various sources and managed by different disciplines, presenting a complex challenge for geological information retrieval [3]. Furthermore, the geological knowledge derived from such dispersed and sparse data is often fraught with ambiguity [4].

The term 'fault' can represent a spatial arrangement structure, an abstract 2D plane, or a 3D deformed volume [5]. However, in textual documents, all these concepts are expressed simply as 'fault. ' This ambiguity in geological fault knowledge and terminology significantly hinders the efficiency of geological information retrieval from complex databases. Consequently, there is a growing demand within the oil and gas industry for integrated geological data and information models capable of enhancing the retrieval process. To address this demand, one key solution is the formalization of geological knowledge.

Building formal geological ontologies stands out as a promising solution for domain-specific data integration and retrieval in the oil and gas industry. In semantic technologies, an ontology is a formal (machine-readable), explicit specification of a conceptualization (abstract of the world that we want to represent in the knowledge) that is shared and agreed upon by the domain community [6]. These ontologies establish a semantic foundation that enhances search engines' capability to recognize and interpret domain-specific terminology and relationships.

Within the geological community, various ontologies have been proposed, ranging from core-ontology for geology [7], fracture ontology [8], plastic rock deformation ontology [9], geological map ontology [10], geological time ontology [11], fault ontology [5] ,deep-marine deposits [12], to risks associated with the petroleum reservoir [13]. A notable industrial case is Petrobras, the Brazilian oil company, which is developing a specialized search engine for geoscientific technical reports empowered by ontology and knowledge graphs [14,15].

The development of an ontology requires the involvement of ontologists, domain experts and users to define the purpose, scope, requirements, etc [16]. During the conceptualization and formalization stages, domain experts bear the primary responsibility for selecting and providing knowledge and terminology resources. Yet, assessing the comprehensiveness of the selected terms and concepts, particularly in a domain as semantically ambiguous as geology, poses challenges. Furthermore, there is also a need to convince non-domain users that the selected terms have good coverage of domain knowledge, which is generally accepted within the geology community. To address this challenge during ontology development, approaches such as term frequency analysis and information extraction from domain documents are recommended to employ [17]. This method has been applied in various geological information and knowledge modeling tasks, such as the subsurface energy [18], mineral exploration [19,20], and geological natural hazard [21]. However, the essential yet ambiguous concept of 'fault'

has not yet undergone a term frequency analysis.

In this paper, to support the knowledge modeling of geological fault for the energy industry, we conduct a term frequency study (Sect. 2) of the 'fault' concept and its related terms from academic paper abstracts. Compared to the verbose full paper, the abstract contains the most important concepts of the research objectives. Cao et al. [22] compared topics extracted from academic papers abstracts and full text and found that the similarity between results is higher when more documents are analyzed. To balance the focus and extension of fault concept, the selected papers range from domain-specific, domain-related, and industrial-related to general domain journals. We listed the renowned geoscience journals with good impact factors and citation scores, and then specialists chose the most relevant for each domain. The extracted terms are subsequently presented to geologists and ontologists to assess their alignment with the domain's understanding (Sect. 3). The evaluation shows promising results. The entire corpus for this study is publicly available.

In our pursuit of identifying terms relevant to the Geological Fault Domain, we adapted the methodology from Garcia et al. [18]. Our approach consisted of the following steps: (i) the selection of scientific journals within the domain of interest; (ii) the compilation of a comprehensive corpus comprising abstracts; (iii) the application of TF-IDF analysis to determine the primary keywords for the Geological Fault Domain; and (iv) a final evaluation of these keywords by domain experts against a pre-established ontology (Figure 1). TF-IDF (Term Frequency -Inverse Document Frequency) is a well-established method in information retrieval for evaluating the importance of keywords within a corpus. Essentially, it identifies terms that are frequent within a specific document but relatively rare across the entire corpus. This approach aids in highlighting words and expressions that best describe a particular document. Differently from Garcia et al. [18], we included in the analysis sets of documents with several levels of domain focuses, including general academic papers. If only one domain is analyzed, important expressions could seem common, when they are relatively rare compared to documents of different subjects. Using several degrees of focus allows us to highlight important terms and expressions for every domain.

The TF-IDF score for each term 𝑡 in a document 𝑑 is calculated as the product of two main components: Term Frequency (𝑡𝑓) and Inverse Document Frequency (𝑖𝑑𝑓). The term frequency, denoted as 𝑡𝑓 (𝑡, 𝑑), is the sum of term 𝑡 occurrences in document 𝑑. In contrast, the inverse document frequency is given by the formula:

] + 1

Where:

• 𝑛 represents the total number of documents in the document set.

• 𝑑𝑓 (𝑡) is the number of documents in the document set that contain the term 𝑡.

Ultimately, the TF-IDF score for a term 𝑡 in document 𝑑 is computed as:

For our TF-IDF calculation, we utilized abstracts of academic papers as our documents. Abstracts offer distinct advantages, as they condense essential information and vocabulary into concise sentences and are accessible for a wide range of papers, including those not available as open-access. Our initial step involved the selection of relevant scientific papers. The TF-IDF method compares the vocabulary of a specific set of texts against the broader corpus. To ensure a diverse vocabulary and to balance the focus and scope of fault-related terms, we initially chose scientific journals within the Geological Fault Domain. Gradually, we expanded our selection to encompass papers from broader knowledge areas. Figure 2 illustrates the distribution of papers across different knowledge areas and highlights our chosen journals. Once we compiled the corpus containing all abstracts, we preprocessed it by removing stop words, applying stemming (a technique that reduces words to their root or base form), and converting all text to lowercase. Subsequently, we calculated the TF-IDF scores for all words and expressions across all documents, considering single words (1-gram) as well as two (2-gram) and three-word (3-gram) expressions. Summing the TF-IDF values across documents within our domain of interest resulted a list of the most significant keywords.

Results. In our investigation, we compiled lists of the most important terms and expressions for all subdomains of Figure 2; Tables 1 and 2 show respectively the keywords for 'Geological fault domain' and 'Geological fault with tectonic domain'. We calculate the TF-IDF value for each word of the vocabulary (and for the 2-gram and 3-gram expressions) for every paper. Then, we summed and sorted all TF-IDF values of the documents of the same domain.

Besides the lists of important keywords for the geological fault domain, this paper presents two noteworthy outcomes:

1. We have assembled a corpus comprising 4,879 scientific papers that focus on various geoscience domains. List of the 60 most relevant terms after calculating and summing TF-IDF for all abstracts of "Geological Fault Domain" journals. They are split in single words term (1-gram), and expressions with two (2-gram) and three (3-gram) words (domain experts makes the less relevant terms in red color).

2. We have demonstrated a methodology for compiling documents and extracting important keywords from them.

Evaluations. In collaboration with geologists, we analysed the alignment between the highfrequency terms in our results and the domain knowledge. For the top-20 high-frequency terms in the Geological Fault Domain (table 1), 49 of 60 terms are closely related to the knowledge of fault; in the results of Geological Fault with tectonics Domain (table 2), only 40 of 60 terms are closely related to the knowledge of fault. It's worth noting that geologists identified some terms as "noisy, " as they are used to describe faults or are related to specific study areas.

In addition to relevance checks, there are some interesting analysis results from the discussion between geologists and ontologists. In Tables 1 and 2, the terms shear zone and fault zone, in geologists' view, are interchangeable in the context of brittle deformation. The terms damage zone and fault core are two components of fault zone. The term fault rock shares a certain level of similarity with fault core, but not necessarily the same. The pull apart basin is the result of normal fault, and thrust fault is a type of low angle fault. These distinctions, while clear to geologists in an academic context, highlight potential semantic ambiguities in everyday usage. Such distinctions prove invaluable when geologists seek specific data and information from databases. Additionally, we also noticed that our data sources are academically biased, which contributes to the presence of certain noisy terms.

This paper proposes an approach with TF-IDF to quantify the term frequency of geological faults during the conceptualization phase of ontology development. The experiment has yielded interesting results for both geologists and ontologists. Our experiment contributes to developing basic terminology for creating knowledge models of geological faults, which will facilitate the retrieval of geological information and data from various sources. The experiment results also provide a basis with good knowledge coverage for ontological analysis to help geologists and ontologists deconstruct the semantically overloaded term 'fault' and its various hidden meanings. In future research, we plan to 1. refine the identified terms for improving the quality of the vocabulary list; 2. incorporate more industrial and technical documents for a more comprehensive analysis; 3. conduct ontological analyses of the terms and implement the proposed corpus to support the development of Petrobras' in-house search engine.