This paper describes a methodology used within the framework of the dual dimension of Terminology for the creation of an ontology-based multilingual terminology resource on an institutionalised domain, namely the balance of payments (BOP). Modelling is based on preliminary knowledge, interactions with experts, and corpus analysis. The terminology resource is operationalised and made interoperable, and shall be reusable by translators. In this paper, we go through the different operations performed for modelling and present some challenges faced.
Terminology is a discipline studying systematised concepts, which have an expressive side that is
most of the time linguistic [
This paper describes the steps involved in the creation of an ontology-based multilingual
terminology resource on an institutionalised domain called the balance of payments (BOP) [
Modelling of terminological information is based on three sources: knowledge acquired through translation and terminology management in our professional activity as a translator and a terminologist in a central bank, interactions with domain experts, and corpus analysis. Basing terminological modelling equally on these three sources makes the originality of our methodology: we are relying neither solely on linguistic analysis, nor exclusively on inputs by experts. A specialised multilingual corpus was built for this research and used both to attest the existence of known terms and for heuristic purposes.
We present hereafter our assumptions (Section 2), before describing the domain of study (Section 3). Section 4 goes through the methodology and its implementation, and Section 5 discusses the results and mentions challenges that were faced, before we conclude (Section 6).
As designations used in specialised domains to represent concepts by linguistic means [
As for the conceptual dimension of Terminology, a concept is, according to [
Concerning operationalisation – i.e. computational representation of the concept system [
We study the domain of the balance of payments (BOP) [
The BOP is an institutionalised domain. Tasks pertaining to its creation (data collection,
compilation, presentation and dissemination) are performed by statisticians at central banks or
statistics offices. The domain of the BOP is standardised at the international level. Statisticians have
to follow special recommendations and accounting principles mainly set out by the International
Monetary Fund (IMF). The current reference manuals, BPM6 [
The BOP is at the intersection of at least three disciplines, namely macroeconomics, statistics, and accounting. Statisticians compile data on macroeconomic phenomena that either occur between entities that are resident in their economy and non-residents (e.g. exports and imports of goods and services, income flows, financial flows), or result from these relationships (financial positions), and they record and present this data in dedicated accounts in the BOP. Furthermore, 1 “Optical mouse: computer mouse in which movements are detected by light sensors” and “mechanical mouse: computer mouse in which movements are detected by rollers and a ball”. 2 That location can be a country, an economic union or a currency union, and is called “an economy” in the context of the BOP. 3 See https://www.imf.org/en/News/Articles/2025/03/20/pr25072-imf-and-statistical-community-release-new-globalstandards-for-macroeconomic-stats. concepts that are essential for the BOP are shared with a neighbouring statistic, namely national accounts, and are often described more precisely in the corresponding reference manual [14].
We have acquired knowledge of the BOP by translating documents on that topic and by managing a terminology database in the context of our professional activity at the Swiss central bank4. Our knowledge is mostly text-based: our acquaintance with BOP concepts and terms is determined – and limited – by the texts we have translated or read in order to understand texts we had to translate. Furthermore, we have widened our knowledge of the BOP by interacting with experts (statisticians responsible for the establishment of the BOP). We are taking part in the text creation process in French, which has a direct influence on the methodology used: our relationship to the BOP is not entirely an external one, we are a kind of “initiate” [15] as we have been in contact with the linguistic means used to talk about the BOP in German, in French, and in English since 2012. We would not be able to produce the BOP terminology based on that experience only, and we still rely on texts to check whether an expression is actually in use, and on specialists to ascertain that it is a term. To sum up, our knowledge of the domain is limited because it is mainly text-based, and we complemented it by consulting experts and with corpus analysis.
For the creation of our terminology resource, we built a specialised corpus on the BOP. Our preliminary knowledge let us make hypotheses about the information to search in the corpus in order to extract term candidates5. The work is both corpus-based and corpus-driven: results obtained with first queries led to the elaboration of additional queries and gave insights on aspects and elements, which had not been anticipated.
As for the conceptual dimension, the concept system was modelled as an ontology, based on our knowledge, based on corpus attestations, and based on interactions with experts.
A specialised corpus was created for this research and organised in five text types that were defined according to the communicative settings they reflect [17], [18] and to the knowledge necessary to understand them (see Figure 1). 4 www.snb.ch/en. 5 We use the term “term candidate” instead of “candidate term”, which is the preferred term according to [16]. This decision is motivated by the following: the thing that has to be named is a “string of characters that has been collected by means of term extraction but has not yet been selected (…) to be considered for inclusion in a terminological data collection” ([16]). In other words, at the time we are considering it, we are not sure that the string of characters in question is actually a term. We would consequently interpret “candidate term” as an elliptic expression denoting elements of the lexicon that are “candidate for the status of term”. Some of them will ultimately not be included in the resource. Based on this interpretation, the head of a noun phrase being its last element in English, we consider it preferable not to use an elliptic expression, and consequently we decided to place “candidate” as head, and thus in last position. At the same time, we would be very interested in a discussion of the reasons that have led to select “candidate term” as the preferred term in [16].
The five types are Regulatory documents (REs, regulatory issues and laws for establishing the BOP), Reference documents, manuals, methodology (RMs, principles of the domain, i.e. statistical and methodological topics like data sampling, compiling, computing, and conducting of surveys), Research papers (PPs), Press releases (PRs, publication of data on the BOP at regular intervals), and General presentations for non-specialists (GPs).
The corpus encompasses 656 documents, with about 29 million characters and 5 million tokens. Texts were published by a central bank, a statistics office or an international organisation like the IMF, in English (48%), in French (38%) or in German (14%), between 2009 and 2024.6 Central banks, statistics offices, and international organisations publishing documents on the BOP are institutions acknowledged for their expertise [15] in that domain. As these documents have been published, they correspond to authentic communicative situations. In all settings, the authors are experts. There is a predominance of RMs: they account for about 10% of the number of documents but for more than half the size of the corpus. RMs have the aim of standardising the domain.
We collected all texts on the websites of authoring institutions and checked with the latter whether we had missed relevant documents. We can thus assume that the corpus covers the whole range of communicative situations that exist in the specialised domain being studied [19]. Its representativeness is thus qualitative, based on typicality and specialisation [20].7
We focussed on RMs in English because these documents correspond to a setting “expert to expert” or “expert to intiate” and aim at standardising the domain, and because English is the language, in which this standardisation takes place.8
Morphosyntactic patterns in English for queries in AntConc 4.2.49 [21] were defined based on
known terms. These patterns are shown in Table 1.
6 The time frame 2009-2024 is determined by the period of validity of the current BOP reference manual [
_JJ _NN _NN _NN net acquisition of financial assets other changes in volume account _JJ _NN _IN _JJ _NN _JJ _NN _IN _NN _NN
Each query returned a list of term candidates. Successively extending and shrinking the cluster size allowed capturing additional term candidates, and terms were inferred based on the researcher’s knowledge.10
All term candidates were then submitted to a first selection and validation process based on our knowledge. Firstly, we rejected pleonasms (like *financial liabilities, all liabilities having the essential characteristic of being “financial” in the context of accounting, as explained in[14]) and usage variants that could be confusing (like *net acquisition of assets in the context of the financial account, instead of the standard term “net acquisition of financial assets”, because only financial assets are relevant in the financial account). As we had focussed our search on RMs, this made us aware of the fact that reference documents and manuals may provide term candidates that we should not select as terms.
Secondly, we cleaned class names, i.e. expressions that do not designate concepts, but classes of things that may be of different natures. As mentioned in Section 3, statisticians group things in aggregates. Designations matching patterns like “_NN and _NN”11, “_NN not included elsewhere” / 10 For details on queries and results obtained, see https://github.com/SCarsenty/Ontology-based-terminology-of-theBOP/tree/main/Corpus/Queries%20on%20English%20corpus. 11 This pattern does not correspond to the ones mentioned in Table 1. It was added at an intermediary stage, based on our knowledge of designations of account elements like “equity and investment fund shares” and “currency and deposits”, and on frequency analysis of collocates in the corpus. “_NN n.i.e.”, 12 “_NN except * _NN” and “other _NN” may signal aggregates grouping heterogeneous things. For example, the patterns “_NN not included elsewhere” and “_NN n.i.e.” indicate elements, which statisticians have not been able to record in any other category. Another example of a pattern signalling class names is “other _NN”. Most term candidates matching that pattern were rejected. However, we kept those, which do not designate classes but concepts denoting entities. These are names of objects, which play a central role in the structuring of the BOP, like “other changes in financial assets and liabilities account” (designation of an account, in which all changes occurring during a period and not pertaining to transactions are recorded) and “other investment” (term denoting a functional category, which gathers specific investment relationships between residents and non-residents).
After this first selection and validation process, we obtained a shortlist of 148 term candidates in English.
The selected term candidates in English were all entered in the ontoterminology editor Tedi 3.713 [22]. Tedi is a software environment that allows creating multilingual ontoterminologies and exporting them into different formats (RDF, HTML, TBX, and CSV).14 Based both on our understanding of the domain and on the list of term candidates, we defined in Tedi seven upper categories: 1. <Entity>: this category allows defining entities, whose activities are observed and analysed by statisticians for establishing the BOP. 2. <Event>: this category is the genus of all concepts representing activities and processes that lead to entries in the BOP. 3. <Instrument>: this category groups concepts representing instruments used by statisticians to collect, record, and present data on the BOP. 4. <Location>: this category models the residence of entities involved in a transaction, as relevant transactions mostly concern a resident and a non-resident entity. Interactions between entities that are resident in the same economy or that are both non-residents are, with very few exceptions, outside the scope of the BOP. 5. <Principle>: this category gathers concepts pertaining to principles, which statisticians have to adhere to when establishing the BOP. These principles are for example accounting rules or rules for data classification. 6. <Product>: this category encompasses concepts representing outcomes of production activities that are supplied or received by entities observed in the BOP, namely goods and services that are exported or imported. 7. <Resource>: this category is the genus of concepts representing objects used by entities to perform an economic activity. These objects are so-called economic assets. Those that are based on a financial contract, namely financial assets, can also be used by entities to perform an economic activity.
Each concept in the ontology was created as a combination of essential characteristics after one of these categories, and associated with a term in English. 12 Abbreviation of “not included elsewhere”. 13 https://ontoterminology.com/tedi. 14 https://ontoterminology.com/export.
This research is still in progress. It has led so far to the creation of a trilingual ontology-based terminology resource on the BOP encompassing 140 concepts. Not all concepts of the domain have been analysed, nor all terms been extracted. The ontoterminology encompasses 150 terms in English, 151 in French, and 169 in German. In other words, the number of terms in each language is bigger than the number of concepts. Nevertheless, not all concepts were denoted by BOP terms: 10 concepts have no designation in any language. Among them are the seven upper categories (see Section 4.2.3): although they can be named in a natural language (e.g. “entity”, “event”…), they are not denoted by BOP terms. Still, they are necessary for the structuration of the concept system. Interestingly, the number of concepts without denotation is bigger in French (12) and in German (20). We interpreted this as a confirmation that English is the language, in which the BOP has been conceptualised and standardised.
The ontoterminology was submitted to experts in order to assess whether the modelling work was reusable by others.15 The validation of a terminology resource is a very important step because it confirms that the knowledge represented corresponds to a consensus among experts. Moreover, it should allow assessing the quality of definitions in natural language. To that end, we exported the ontoterminology in Tedi into two human readable formats: as an HTML electronic dictionary (see Figure 3) and as a concept map that can be edited in the software CmapTools16 [24].
Experts gave a valuable feedback that allowed among others underlining modelling errors and discarding irrelevant concepts and terms. Modelling errors were the consequence of our limited knowledge of the BOP (as explained in Section 4.1) and of misinterpretation of corpus data. As for irrelevant concepts and terms, they resulted from the fact that the BOP is at the intersection of different disciplines (as mentioned in Section 3), and that we included in our corpus documents pertaining to neighbouring statistics, e.g. national accounts, because the BOP shares with these statistics some common concepts, and definitions are more precise in the reference manual for national accounts. We have not been able to reject straightforwardly those terms of national accounts that are not shared with international accounts and that are consequently irrelevant.
After the validation by experts, a formal validation of the ontology with competency questions (CQs) [25] was performed. CQs are “natural language sentences that express patterns for types of question people want to be able to answer with the ontology. The ability to answer questions of the type indicated by a CQ meaningfully can be regarded as a functional requirement that must be satisfied by the ontology”[26]. In Tedi, the ontoterminology was converted into a knowledge graph in RDF format, which allows editing in Protégé.17 15 The validation stage was limited due to time constraints. 16 https://cmap.ihmc.us/. 17 https://protege.stanford.edu/.
Figure 4 is an excerpt of Tedi’s help. It shows the different vocabularies used for the conversion into RDF. These include OWL, RDF, RDFS, SKOS, and OTV, a vocabulary conceived for the expression of essential characteristics as instances of classes.
The RDF knowledge graph was uploaded to the server http://www.ontologia.fr/OTB/BOP.rdf. The following CQs were defined: 1. Designation of Concepts in Different Languages: Which are the names of services relevant for the balance of payment in English, in French, and in German? 2. Structuration of the Concept System by Generic Relations: Which are, in English, the names of all economic assets that are based on a financial contract? 3. Partitive Relations: Which are the names of the different parts of the current account in English? Which are the names of the accounts making up the balance of payments in English? 4. Associative Relations: Can you designate in English all entities, which can own financial assets, non-produced non-financial assets or goods? 5. Definitions: What is the definition of a currency union? What is the difference between a customs union and a currency union?
All CQs were expressed with SPARQL syntax. BOP.rdf was queried through the SPARQL endpoint http://sparql.org/sparql.html. We present hereafter the expression of the second of the two CQs for partitive relations, namely “Which are the names of the accounts making up the balance of payments in English?” with SPARQL syntax.
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX skos: <http://www.w3.org/2004/02/skos/core#> PREFIX owl: <http://www.w3.org/2002/07/owl#> PREFIX otv: <http://www.ontologia.fr/OTB/otv#> PREFIX bop: <http://www.ontologia.fr/OTB/BOP#>
?cpt skos:prefLabel "balance of payments"@en.
?partCpt bop:partOf ?cpt.
?partCpt skos:prefLabel ?termEn.
} ORDER BY ?termEn
The query searches for particular concepts in the BOP ontology (?partCpt bop), which are linked by the relation partOf with the concept denoted by “balance of payments” as its preferred term in English (skos:prefLabel). It returns the following answer, which is correct: termEn
We mention in this section challenges that were faced in this research. Some are linked with the structuration of knowledge in the BOP (Section 5.4.1), others concern knowledge representation with the software environment chosen (Section 5.4.2), and a last category pertain to the size and composition of subcorpora (Section 5.4.3).
Not all knowledge units in the BOP can be defined straightforwardly by specific differentiation.
This can be illustrated with the concept of residence. Entities (i.e. institutional units, see definition
in Figure 3) that have their “centre of predominant economic interest” [
Secondly, BOP compilers define broad categories that group different elements (entities, resources, products, etc.). These categories are relevant for the structuration of knowledge in the domain, because they reflect the way experts classify things. Nevertheless, they do not correspond to clear-cut concepts. They are easy to identify because they are clearly lexicalised, with linguistic means like “_NN n.i.e.” or “other _NN” (see Section 4.2.2). We either split them into their component (e.g. “maintenance and repair services n.i.e.” was split into “maintenance service” and “repair service”), rejected them and searched for the members of the class they designate (e.g. “other financial corporations”) or kept designations that represent classes of objects that can be defined intensionally (e.g. “other investment”).
Moreover, for certain concepts, the corpus supplied only extensional definitions, and it was not always unproblematic to determine the corresponding intension.
The ontoterminology created in Tedi can be exported into RDF, HTML, CSV, and TBX. RDF guarantees the interoperability of the resource on the Semantic Web [27] (see Section 5.3). Challenges were faced, among others because of missing data categories: at the time of writing, Tedi allows recording, for each term, a definition and one or more notes and contexts. However, it is not possible to record the source of notes and contexts in dedicated fields. We thus had to enter the information in the fields themselves.
As for the HTML dictionary, it provides very valuable conceptual information like the position of the concept in the concept system, its inherited and its specific differences, its genus, and the relations that may link it with other concepts (see lower part of Figure 3). That information is expressed in the formal language used by Tedi, with concept IDs that can be very long. The information is thus not immediately accessible for a human user. One possible improvement could be replacing each concept ID mentioned in that section with the term that denotes it – provided it is denoted in language, which is the case of most concepts in an ontoterminology.
To ensure interoperability with computer-assisted translation tools (CAT tools), the ontoterminology can be exported into TBX. However, the source of definitions is not included in the TBX export. Including a source for each piece of information in a terminology resource belongs to best practice in terminology management. This is why we regard this aspect in the TBX export as presenting room for improvement.
The small size of the subcorpus in German has limited the productivity of corpus analysis for that language. Whereas the size of the English subcorpus is 2 505 648 tokens (out of which 1 519 279 in reference documents and manuals [RMs]), and the subcorpus in French has 1 922 459 tokens (out of which 1 280 028 in RMs), the subcorpus in German encompasses only 633 046 tokens (out of which 217 591 in RMs). As a consequence, a certain number of terms in German could not be found in the corpus, and it was not possible to extract KRCs in German for most terms.
The BOP is a standardised domain. In that domain, concepts and terms are stable and based on a
consensus among experts over a period of time (2009-2024 in the case of the research presented in
this paper). We have acquired domain knowledge through the translation of texts, but no expertise.
Moreover, we have direct access to experts (statisticians) establishing the BOP in our institution.
Last, it was possible to compile a corpus of texts on the BOP in three languages, enabling us to
complement our knowledge with authentic texts, i.e. “knowledge in action” [
The RDF format chosen for knowledge graphs guarantees the interoperability of the resource created. The definition of a concept in a formal language is the basis for generating an expression in natural language of that definition. However, improvements can certainly be brought regarding interoperability with CAT tools and reusability by humans, be they translators or not.
This research is as mentioned still in progress. Challenges that were faced provide directions for future work. We believe that interviewing German-speaking experts will allow filling remaining gaps in the BOP terminology in that language. Additionally, the validation by experts shall be strengthened. Concerning knowledge modelling, while potentially challenging, modelling both concepts and classes of objects that structure knowledge in the BOP is necessary. The main reason for this is that experts do structure at least part of the domain with umbrella categories.
Additionally, it would be highly interesting to link the BOP ontoterminology with existing ontologies on neighbouring domains. Last but not least, a study of the diachronic dimension would be valuable, especially in view of the publication of the new BOP reference manual in March 2025.
The author has not employed any Generative AI tools. [13] European Commission. COMMISSION REGULATION (EU) No 555/2012 of 22 June 2012 Amending Regulation (EC) No 184/2005 of the European Parliament and of the Council on Community Statistics Concerning Balance of Payments, International Trade in Services and Foreign Direct Investment, as Regards the Update of Data Requirements and Definitions, 2012. [14] European Communities, International Monetary Fund, Organisation Economic Co-operation, United Nations Development, and World Bank, System of National Accounts 2008 (SNA 2008).
New York: United Nations, 2009. [15] J. Pearson, Terms in Context. Amsterdam/Philadelphia: John Benjamins Publishing Company, 1998. [16] International Organization for Standardization, ISO 5078:2025(en) Management of
Terminology Resources — Terminology Extraction, 2025. [17] D. Maingueneau, Analyser les textes de communication. 3e ed. Paris, France: Armand Colin, 2016. [18] P. Charaudeau, Dis-moi quel est ton corpus, je te dirai quelle est ta problématique, Corpus 8:37–66. doi: 10.4000/corpus.1674), 2009. [19] A. Koester, Building Small Specialised Corpora, in: The Routledge Handbook of Corpus
Linguistics. Abingdon: Routledge, 2010, Pp. 66–79. [20] G. Doualan, De la représentativité à la spécialisation : exemple d’un petit corpus sur la synonymie, Corpus 18 | 2018. doi: 10.4000/corpus.3331. [21] L. Anthony, AntConc [Computer Software], 2023. [22] C. Roche, Tedi [Computer Software], 2024. [23] I. Meyer, Extracting Knowledge-Rich Contexts for Terminography: A Conceptual and Methodological Framework, in: Recent Advances in Computational Terminology. Amsterdam/ Philadelphia: John Benjamins Publishing Company, 2001, Pp. 279–302, doi: 10.1075/nlp.2.15mey. [24] Institute for Human & Machine Cognition (IHMC), ‘CmapTools [Computer Software]’ 2019. [25] M. Uschold, and M. Grüninger, Ontologies: Principles, Methods and Applications, The
Knowledge Engineering Review 11(2):93–136, 1996. [26] Y. Ren, A. Parvizi, C. Mellish, J. Z. Pan, K. van Deemter, and R. Stevens, Towards Competency Question-Driven Ontology Authoring, in V. Presutti, C. d’Amato, F. Gandon, M. d’Aquin, S. Staab, and A. Tordai (Ed.) The Semantic Web: Trends and Challenges, Cham: Springer International Publishing, 2014, pp. 752–67. doi: 10.1007/978-3-319-07443-6_50. [27] T. Berners-Lee, J. Hendler, and O. Lassila, The Semantic Web. A New Form of Web Content That Is Meaningful to Computers Will Unleash a Revolution of New Possibilities, Scientific American, May, 2001, 34–43.