Information Architecture Model for Ontology Learning in the Electronic Consumer Invoice (NFCe) environment

ABSTRACT

Introduction:  Ontology Learning works with Machine Learning, Data Mining, and Text Mining algorithms and provides a method to represent and share domain knowledge.

Objective:  This article aims to present and discuss the evaluation of the results of the Information Architecture used to organize data that will be retrieved and used by Machine Learning in the construction of an ontology about the product beer.

Methodology:  The literature review will address topics such as: Machine Learning, Information Architecture, Ontology and Ontology Learning, a summary of the method and the evaluation of the Organization of Information with Term Mining in the context of Electronic Consumer Tax Invoices, a receipt used in the commercialization of goods in the States, applied to the product beer due to its large number of qualifiers and the difficulties in its correct identification. The research begins with an exploratory bibliographic survey and analysis of examples that deal with adopted resolution models and develops a literature review within the terms in the research context, whose sources of information used were scientific and academic texts.

Results:  The methodology adopted to obtain the data was based on the first layer of “Term Mining” for Ontology Learning as a response to the elements of Information Architecture in guiding the construction of the ontology, without requiring the presence of the domain expert.

Conclusion:  The Information Architecture Model for Ontology Learning focuses on the form of requirements acquisition, which helps to facilitate understanding of the process to identify products in the e-invoicing environment.

KEYWORDS:
Information architecture; Text mining; Machine learning; Ontology learning; Ontology

RESUMO

Introdução:  A Ontology Learning trabalha com algoritmos de Aprendizado de Máquina, Mineração de Dados e Mineração de Texto e fornece um método para representar e compartilhar conhecimento de domínio.

Objetivo:  Tem como objetivo apresentar e discutir os resultados da aplicação dos Requisitos da Arquitetura da Informação para organizar dados que serão utilizados por algoritmos não supervisionados de Aprendizado de Máquina para construção de uma ontologia do produto cerveja.

Metodologia:  A revisão da literatura aborda temas como: Aprendizado de Máquina, Arquitetura da Informação, Ontologia e Ontology Learning, um resumo do método e da avaliação da organização da informação com Mineração de Texto no contexto de Notas Fiscais de Consumidor Eletrônicas (NFC-e), documento fiscal utilizado na comercialização de mercadoria nos Estados aplicado ao produto cerveja pelo seu grande número de qualificadores e dificuldades para sua correta identificação. A pesquisa inicia com levantamento bibliográfico exploratório e análise de exemplos que tratem de modelos de resolução adotados e elabora uma revisão de literatura dentro dos termos no contexto da pesquisa, as fontes de informação são textos científicos e acadêmicos.

Resultados:  A metodologia adotada para obter dados se baseia na ‘Mineração dos Termos Relevantes’ para Ontology Learning como resposta aos Requisitos da Arquitetura da Informação para orientar a construção da ontologia, sem a necessária presença do especialista do domínio.

Conclusão:  O Modelo de Arquitetura da Informação para Ontology Learning concentra-se na forma de aquisição de dados, que contribui e facilita o entendimento do processo para identificar produtos no ambiente de Notas Fiscais de Consumidor Eletrônicas.

PALAVRAS-CHAVE:
Arquitetura da Informação; Mineração de texto; Aprendizado de máquina; Sistema de organização do conhecimento; Ontologia

1 INTRODUCTION

Artificial intelligence (AI) is already an integral part of everyday life, solving realworld problems. It is an evolution of computer programs that, years ago, used codes to process logical rules extracted from expert knowledge. This knowledge was extracted from interviews, but this process was hindered by human subjectivity and a lack of cooperation from the expert.

Other difficulties, such as computational complexity, data volume, different data formats, and knowledge scattered in free texts in various storage media, also motivated the development of sophisticated computational tools that are independent of human intervention. These tools have been developed in the field of AI and its subarea of Machine Learning (ML), which is transforming the way we interact with computers through natural language processing, facial recognition, computer vision, and intelligent behavior based on user choices. This makes it possible to deliver advanced computer systems quickly.

Ontology provides an appropriate technical means for sharing and exchanging knowledge between humans and/or machines, improving the machine's reasoning and comprehension capabilities (Du et al., 2024, p. 1).

Ontology learning, or ontology learning environments, is a fundamental area within this domain. It is responsible for extracting, representing, and automatically refining conceptual knowledge for ontology engineering. AI methodologies such as text mining, machine learning (ML) and natural language processing (NLP) are used for this purpose.

Exploring the relationship between an object and its properties and behaviors using ML faces difficulties, mainly in configuring AI and selecting and organizing information to be processed for learning and adapting the environment. However, AI can provide guidance on the selection and treatment of data, facilitating the acquisition of conceptual knowledge for ontology engineering by helping to automatically discover data in free texts that have not yet been fully explored in AM studies.

Several methodologies for developing ontology learning environments have already been suggested, but there is a need for a preliminary study before developing the ontology to organize data in a way that improves retrieval efficiency, security, and speed while avoiding duplication of information and the absence of specialists and reducing the effort required to mine unnecessary information.

In this context, information architecture requirements are inserted iteratively and incrementally into the layers of ontology learning to form a model that assists in obtaining terms, classes, and relationships-an artifact that improves modeling and speeds up the search. A scenario of electronic consumer invoices (NFC-e) is evaluated based on the requirements obtained at the end of the model's application.

This article aims to present and discuss the evaluation of the Information Architecture used to organize data to be retrieved by text mining and used by AM in ontology learning for the construction of an ontology describing beer products marketed in NFC-e tax documents, also known as tax coupons.

The literature review will address topics relevant to the proposed discussion, namely: AM, information architecture, ontology, and ontology learning. It will provide a summary of the method and evaluation of information organization for ontology engineering using term mining in the context of NFC-e.

This study evaluates information collected from text mining using the structural elements of information architecture and an ontology to assist with NFC-e auditing. Exploring this hypothesis and proposing a possible solution using theoretical frameworks associated with information science creates a connection of approximation or familiarity with the subject. The research begins with a plan for a bibliographic survey and analysis of examples dealing with adopted resolution models. The exploratory bibliographic survey reviews literature within the context of the research. Sources of information included books, articles, journals, theses, and dissertations.

The methodology adopted to obtain the data was based on the first layer of 'relevant term mining' for ontology learning. This presented a response from each element of information architecture (IA) as a guideline for constructing the ontology without the presence of a domain expert being necessary.

To better organize the information, this work is divided into sections on the literature review, related works, the development of the information architecture model for ontology learning, the results and discussion, and finally, the final considerations.

2 LITERATURE REVIEW

Architecture, as an ancient activity, focuses on organizing human construction possibilities, combining, according to Vitruvius' vision, beauty (Venustas), utility (Utilitas), and structure (Firmitas) (Kuroki Junior, 2018), compatible with an organization capable of recovering what impacts people. The author adds that the term "architect" has been associated with the term "information," as in "information architecture," in the sense of subjecting information to an order for better human appreciation, of organizing by first deciding how one wishes to search for and find something in the “[...] construction of the information structure that allows others to subsequently understand it” (Wurman, 1997, p.17) to retrieve the necessary and sought-after information. For the author, “[...] the information structure must relate something that is already understandable to those being instructed so that it is possible to relate something understandable to something unknown” during the search process (Wurman, 1997, p.17).

This organization has some requirements for its construction, as explained by the author, which are: Location, Alphabet, Time, Category, and Hierarchy of data (Wurman, 1997), with the purpose of designing, constructing, and shaping an order recognized as appropriate by human cognition. Location indicates the place or warehouse where the documents with the necessary and desired data are sometimes scattered without an order that is understandable to humans or machines; the Alphabet has a structural equivalence with the data dictionary, the organized and systematic manifestation of words constructed from symbols of a finite alphabet; Time indicates the initial and final period where data research will be restricted to delimit a scope of search and evaluation of results; Category manifests the obtaining of terms and groups of terms from equivalent meanings; Hierarchy manifests taxonomic or nontaxonomic relationships that may become essential in the search for more elements that facilitate ontology.

Many authors, including Siqueira (2012) and Hessen (2003), recognize the phenomenon of knowledge construction based on the correlation between subject and object, whereby reality (object) is accessible through human experience and thought (subject). Therefore, before organising information in these systems, it is necessary to understand the subject, restrict the search environment objectively, and familiarize oneself with the set of words and expressions in the dictionary and/or alphabet that are part of the domain, as well as their meanings and relationships, to search for intelligent resources for the structural design of the information environment. Applying Information Architecture requirements to Text Mining and using IA tools enables a more complete, accessible and accurate observation of the Subject and Object, ensuring optimization in terms of time and resources, interoperability, and deductive reasoning for ontology (Guidalia et al., 2023).

When using computational tools, it is important to consider the term 'artificial intelligence', which, according to some authors (Nilsson, 2009; Wang, 2019), refers to machines' ability to solve complex problems and adapt to environments with limited knowledge and resources. Kaplan and Haenlein (2019) define AI as an intelligent system's ability to correctly interpret data, learn from it, and use that learning to achieve specific goals and tasks through flexible adaptation.

Machine learning (ML), a subarea of AI, enables learning from experiences using the principles of inference and induction to extract generic conclusions from a particular set of examples (Faceli et al., 2021). Another advantage of ML is that it can work with imperfect data, such as a dataset containing noise, inconsistencies, missing data, and redundancy.

In MA, the development of a model, i.e., the learning of rules, must be sufficiently robust and generic so that these rules apply to the training dataset and to data outside the training set but within the domain or context. MA tasks can be categorized as either predictive or descriptive.

Predictive tasks involve predicting the value of an object based on recognized predictive attributes from a given training set. Descriptive tasks involve extracting patterns from the predictive values of a dataset to search for similar objects or rules of association between objects within the same domain (Faceli et al., 2021). The authors also point out that '[...] a predictive model can generate descriptions of a dataset, and a descriptive model can provide predictions after being validated' (Faceli et al., 2021, p. 4).

Unsupervised MA algorithms, such as Apriori¹, are efficient descriptive association algorithms responsible for mining frequent items (ItemSet) to discover knowledge in the form of association rules between frequent items in a database containing multiple transactions. Each transaction 'supports' a specific subset of frequent items (Agrawal & Srikant, 1994). For Faceli et al. (2021), providing support means 'testifying in favor'of a given set of items in relation to the dataset.

Thus, the support of an item set is the fraction of transactions in a database that contain the set. In Apriori, when considering a set of frequent items, it is possible to derive probabilistic association rules in the form of "antecedent consequent." The degree of uncertainty or certainty of the rule is determined by verifying the confidence and lift (support) of the rule (Gorayeb & Duque, 2024). This process establishes the discovery of association rules based on their efficiency, ensuring that each rule supports the association between a set of frequent items (Alpaydin, 2014; Sumithra & Paul, 2010). In the context of model induction, each AM algorithm represents possible biases in representation or search, ensuring the learning and generalization of knowledge acquired in the training process. This knowledge will be used in the "Ontology Learning Environment" or "Ontology Learning."

In ontology, classification arises from the need to define the search object and provide a definition for each term of interest while grouping similar definitions together. This creates categories of interest, adding semantic richness and forming a hierarchical, aggregated structure with defined relationships. Thus, something structured and prepared to produce an organized system of knowledge is presented that can be shared between people and systems. An ontology is constructed to define the relationships between items or key terms. According to Mori (2009), an ontology relates formally and consensually represented concepts within a given domain. It is a portion of reality reproduced logically, with which different computerized information systems can operate.

Conceptualization provides a vocabulary that allows informational resources to be used and reused, be interoperable, and be applied by keyword search engines for information retrieval. According to Gruber (1995), concepts in ontologies are formalized through classes of objects containing properties (or attributes). These ontologies contain functions and relationships in the form of a set of assertions that are used to model a given domain and define the vocabulary used by the application (Astrova, Koschel, & Lee, 2020). They also propose axioms about these elements as a means of sharing knowledge (Evaristo & Duque, 2011).

Formally, an ontology can be described as the following tuple:

$O=\langle C,H,R,A\rangle$

In ontology, it represents a set of classes (concepts), hierarchical links between concepts (taxonomic relations), conceptual links (non-taxonomic relations), and rules and axioms (Zouaq, Gasevic, & Hatala, 2011).

Some studies on ontology engineering advocate a construction process corresponding to a life cycle shared by several widespread Information Science methodologies. This process includes distinct stages for the extraction and classification of relevant elements, implementation, sharing, and the cyclical and incremental expansion of knowledge (Fox et al., 1993; Noy & McGuinness, 2001; Fernández, Gómez-Pérez, & Juristo, 1997; Ushioda & Gruninger, 1996).

Chart 1
Stages of ontology construction

Definition of scope, requirements, and specific knowledge	The project plan and competency issues for the ontology model are defined as means of monitoring the project objectives.
Modeling	Chain of concepts and information flow, rules, quality standards, restrictions, and other mechanisms of the classification process that support the representation of reality;
Formalization and implementation	Use of logical language to represent conceptualization;
Validation	Establishment of criteria for the relevance of concepts and presentation of informational content, such as the generality of the model, efficiency, and clarity in reasoning, transformability, scalability, and integration of the model into knowledge practices;
Evolution	Assessment of the degree of maintainability and updating of the model seeking to expand knowledge.

• Source: Adapted from Fox et al. (1993); Noy; Mcguinness (2001); Fernández; Gómez-Pérez; Juristo (1997); Uschod; Gruninger (1996)
• Description: Chart 1 presents the stages of ontology construction with two divisions, the first column identifying the “Definition of scope, requirements, and specific knowledge” and the second column identifying: “The project plan and competency issues for the ontology model are defined as means of monitoring the project objectives.” The chart has four rows with a salmon pink background.

The incorporation of ML into the ontology construction process, such as Ontology Learning, seeks to extract new, useful, and relevant knowledge from a data set and processes the information from the beginning of the construction process. ML will allow the use of descriptive ML algorithms to define concepts, hierarchies, and rules, and ultimately, ML's predictive means will allow the prediction of useful descriptions and statements about the object of interest in the domain.

Thus, Ontology Learning relates rule discovery, enrichment, improved learning, automatic taxonomy proposal, and non-taxonomic relationships (Hassan; Rashid, 2021). According to the authors, AI-assisted techniques can offer pattern classification and knowledge mining by discovering various types of hidden relationships in knowledge, such as the “main term-attribute terms” relationship of the object of interest, the “terms-behavior” relationship of the object of interest, and the relationship between non-taxonomic “main termcomplementary terms,” but still within the domain where the object is found.

The extraction and organization of concepts and meaningful knowledge are fundamental to machine understanding and the reasoning ability of algorithms, since Ontology Learning is responsible for the extraction, representation, and refinement of structured ontologies that encapsulate the complexities of various domains (Du et al., 2024). According to the authors, learning techniques make it possible to describe terms from a given domain based on semantic understanding and to infer relationships between entities, such as the descriptive and predictive proposal of AM.

Ontology learning refers to semi-automatic or automatic support for the construction, instantiation, and evolution of an ontology (Hassan; Rashid, 2021) through knowledge discovery in different types of data sources and its representation through an intelligent approach to automate or semi-automate the process from raw data (text documents, images, numerical data, or even other ontologies).

Knowledge Discovery in Databases (KDD) is the process of identifying new knowledge extracted from a database or from texts in Text Mining (Fayyad et al., 1996). The process of constructing ontologies, Ontology Learning, consists of a set of layers with concepts, relationships, and axioms extracted from unstructured text: term mining, definition of meaning, concept and hierarchy, association of relationships, and extraction of knowledge for axioms and rules, as shown in Figure 1:

Figure 1
Ontology Learning LayerCake model

According to the authors, the Terms layer identifies relevant terms in documents or interviews with experts or through the statistical process of text mining. The Synonyms layer addresses acquiring semantic variants of terms, i.e., terms similar in meaning to or reciprocal of the object of interest. The Concept layer begins with the designation, induction, or formation of an intention and describes the realization of all overlapping terms and synonyms (Lisi, 2007). This involves organizing related terms into hierarchies or categories based on similarities, functionalities, or semantic relationships (Du et al., 2024). The hierarchy layer of concepts deals with relationships between words with specific and general meanings (hyponymy), ensuring future lexical-syntactic relationships so descriptions of the object of interest in a given domain are well-structured and make sense (Buitelaar, Cimiano, & Magnini, 2005). The nonhierarchical Relationship stage, which depends on term mining, relates essential terms to complementary terms to improve the lexical-syntactic meaning of the object description. This creates layers of new concepts and hierarchies obtained through association rules (Maedche & Staab, 2000). The axioms or rules stages relate to restricting relationships through specialized knowledge or AM algorithms to extract rules and identify data properties during text mining (Gorayeb & Duque, 2024). These stages define dependencies or logical relationships between entities or concepts. They seek to formalize domain knowledge and establish logical constraints within the ontology.

The data for this study were made available by the Amazonas State Finance Department (SEFAZ/AM) in the form of a .csv text file containing files from the NFC-e database from February 1 to May 31, 2023. Once the database was accessible, the Mercosur Common Nomenclature (NCM) product with the initial four digits, 2203.xxxx, was chosen for the application of the model. "Malt beer" was chosen for the application of the model. The NCM 2203.xxxx sample is relevant because, of all the logically and computationally structured NFC-e fields, the product description field allows for free descriptions without controlling the terms used to identify products sold. Additionally, beer products have many brands, additives, packaging types, and quantities. Together with the possibility of a free description, this leads to immense difficulty in marketing and inspection control by the competent authorities.

3 RELATED WORKS

Ontology and AM (automated reasoning) learning environments present many categories of study, including mapping related ontologies, enriching ontology terms, automatically populating ontologies, automatically discovering ontology rules, automatically constructing taxonomic and non-taxonomic relations, and reducing the granularity of ontology concepts for improved reasoning.

Within these categories, there are two main types of ontology learning: semi-automatic and automatic. There are also many challenges to analyze, such as work intensity, formulation of axioms, automatic knowledge acquisition, scalability, heterogeneity, evaluation and validation, and ambiguities (Du et al., 2024), as well as diversity of applied machine learning (ML) methods, inductive and deductive reasoning, and automatic construction of taxonomy (Guidalia et al., 2023).

However, considering the use of unsupervised machine learning (ML) algorithms to discover relationships and statistics when obtaining construction data, population, and ontology enrichment, some articles evaluate the proposed architecture for organizing input data for ontology learning, as described in Chart 2.

Chart 2
Evaluation of Information Architecture in Studies for Ontology Learning Using Unsupervised AM Algorithms

Authors	Information Architecture for Ontology Learning	Application	AI results
Chung; Yoo; Choe (2020)	Architecture based on data mining process: Data compression; Data selection; Data preparation; Data modeling; Data evaluation.	Risk assessment in the area of personal health.	The data is processed solely from a computational standpoint during the Data Mining and Text Mining process.
Yang (2020)	Architecture based on rational organization: reasoning, semantic analysis of terms, and use of experts to extract concepts (ontology engineering stages)	Risk assessment for logistics in financial security.	There is no data preparation prior to ontology engineering.
Maedche; Staab (2000)	Architecture based on a text management and processing component “to select domain texts explored in the discovery process” and choice of text preprocessing methods.	Evaluation of ML techniques using shallow and deep learning.	There is no data preparation before text preprocessing for input into the ontology modeling environment - Ontology Modeling Environment (OntoEdit4).
Haridy et al., (2023)	Architecture based on ontology engineering with the first phase Data acquisition requirements: Identification; Analysis; Specification.	Acquire a list of requirements that the output ontology identifies as the main information of the domain and user needs in the area of tourism.	The data is formatted as information requirements in the first stage of ontology engineering (ON-ODM methodology).

• Source: Chung; Yoo; Choe (2020); Yang (2020); Maedche; Staab (2000) e Haridy et al., (2023)
• Description: Chart 2 presents a description of the Information Architecture Assessment in studies for Ontology Learning with four columns. The first column lists the authors of the research. The second describes the “Information Architecture for Ontology Learning.” The third column describes the “Application,” and the last “IA Results.” The chart has red borders and a salmon pink background.

The studies always highlight the importance of the quality of the Corpus data for the construction of semi-automatic or automatic domain ontologies and how poor data quality influences the lack of depth in NLP techniques for extracting and integrating data and syntactic relationships between words, hindering depth and expressiveness in the domain and evidence about the relevance of concepts. Despite this, the methodologies studied do not focus on phases of knowledge or information organization, characteristics of Information Architecture, generating numerous terms that are not suitable as elements of domain ontology.

4 DEVELOPMENT OF THE INFORMATION ARCHITECTURE MODEL FOR ONTOLOGY LEARNING

The Information Architecture model proposed in Figure 2 is independent of any specific tool or framework. The process can be applied to any set of data from a given domain, and from this generic perspective, the input goes through Information Architecture requirements to identify and organize a corpus dedicated to extracting concepts relevant to a given existing domain. The output is a list of concepts and relationships of interest to the ontology.

Figure 2
Information Architecture Model for Ontology Learning

The proposal for an Information Architecture model to organize data in the construction layers of Ontology Learning covers the phases of Text Mining and the application of ML techniques such as PLN and unsupervised algorithms for statistics and association rules. It serves as a “best practices” manual for identifying which “input resources” are relevant, what format they are in, how they should be handled, and where they are located, that is, the availability of prior knowledge at the beginning of ontology construction.

The prior knowledge used in the ‘Term Mining’ layer, for example, will be transformed into the first version of the first layer of Ontology Learning (Figure 3), presenting statistics on mined terms that will later be classified and ordered into main terms, attributes, complements, and behaviors.

Figure 3
‘Term Mining’ layer, cloud of terms mined from the Information Architecture in the sample of the product of interest, beer.

The output version of the first layer, ‘Term Mining’, already classified, will undergo a new cycle of Information Architecture (Location, Alphabet, Time, Category, Hierarchy) for entry into the next layer, ‘Equivalence between Terms’, and so on for all other layers of Ontology Learning construction (Figure 1): ‘Equivalence between Terms’ for ‘Classes’; ‘Classes’ for ‘Class Hierarchy’; ‘Class Hierarchy’ for ‘Association of Preferred and NonPreferred Terms’; and from there to the last layer, ‘Behavior of Axioms and Rules’, favoring learning in an iterative and incremental process. At the end of the process, the builder is free to restart and repeat it as many times as necessary to add data and information sources and extract new terms, concepts, hierarchies, axioms, and rules for the ontology.

Considering the NFC-e environment, the challenge is to mine terms in the field of the invoice for product description, a free field without computational logical structures and without metadata, which represent the product of interest: beer, in such a way that it is possible to relate the main term beer to the other terms that qualify it in sufficient quantity and quality to recognize useful forms of product description at the end of the ontology. In addition to the useful description of the product sold, it is necessary to concatenate other logical, computationally structured data in the ontology that represent sales information such as value, unit, invoice issuer, consumer, etc., building an ontology suitable for monitoring the sale of goods by the competent state agencies.

The input resources of the NFC-e domain are three types of data: structured, semistructured, and unstructured:

• a) Structured resources are located in NFC-e database schemas (from which the study data sample was taken) or other existing product and beer sales ontologies (research source). The main issue in Ontology Learning with structured information sources is determining which pieces of structured information can provide adequate and relevant knowledge. Using the proposed Information Architecture model, the knowledge selected for structured resources is found in lines (#) 1, 2, and 4 of Table 3 and lines (#) 1, 2, and 4 of Table 4;

• b) Semi-structured resources are typically represented by manually compiled electronic dictionaries, which are usually open, well-documented, and free sources. Using the proposed Information Architecture model, the knowledge selected for semi-structured resources can be found in line (#) 3 of Table 3 and line (#) 3 of Table 4;

• c) Unstructured resources rely on AM techniques, natural language processing, and information retrieval with Text Mining to search for patterns and trends relevant to the domain. The results for this research, according to Information Architecture, are found in lines 5 and 6 of Table 4.

Next, we present the Information Architecture requirements for Ontology Learning in Charts 3 and 4, with examples of results found from the selected data types.

Chart 3
Cycle in the first layer, “Term Mining,” of the Information Architecture Model of Ontology Learning for NFC-e beer products

#	Data types	Information Architecture Requirement	Selected result
1	Logical and computationally recognized fields in rows and columns of the NFC-e database tables for product identification:	Localization	• Mercosur Common Nomenclature (NCM): 2203.xxxx; • cEAN (GTIN code); • Item number; • Quantity; • Unit; • Value; • Original description (extracted from the “product description” field on the invoice).
2	Logical and computationally recognized fields in rows and columns of NFC-e database tables for sales identification:		• State (UF); Municipality; • Issuer (Group C of NF-e); • Recipient (Group E of NF-e); • Invoice Numberal.
3	Grammar with terms (lexicon) that describe the product beer and that are initially identified.	Alphabet	• Extracted from State Resolution No. 028/20232; • Extracted from reference ontologies.
4	Period of interest for auditing NFC-e.	Time	1. Sample period provided by SEFAZ/AM: February 1, 2023, to May 31, 2023
5	Unstructured fields of the “product description”: classification by equivalence or reciprocity of words related to terms describing beer (by statistics).	Category	• Term of interest: • “BEER.”
6	Unstructured fields in the “product description”: associations constructed between terms in the	Hierarchy	Null
	alphabet (by the force of the rules between terms).

• Source: Survey data (2025)
• Description: Chart 3 shows the Cycle in the first layer, “Term Mining,” of the Ontology Learning Information Architecture model for NFC-e beer products, with four columns. The first column identifies the ‘hashtag’ symbol. The second column describes the “data types.” The third column contains the “Information Architecture Record.” The fourth column describes the “Selected Result.” Six lines are shown in the chart. The chart has red borders and a salmon pink background.

At the end of the cycle for the first layer, ‘Term Mining’ (Figure 2), the Information Architecture requirement Hierarchy was not used, as there is insufficient knowledge about the data to define which hierarchies will be useful and should be mined for the construction of the ontology. In addition, the Information Architecture model indicates that the only Category, line (#)5 of Table 3, that is part of the domain and is known is the object of interest itself: BEER.

When applying the Information Architecture model to other layers of Ontology Learning, Table 4, the requirements begin to be completed and provide guidance for organizing the information and retrieving it in the form necessary for the construction of the NFC-e ontology.

Chart 4
Complete cycle of the Ontology Learning Information Architecture Model for NFC-e beer products

#	Data types	Information Architecture Requirement	Selected result
1	Logical and computationally recognized fields in rows and columns of the NFC database tables-and for product identification:	Localization	• NCM (Brazilian Classification of • Goods and Services); cEAN (GTIN code); • Item number; • Quantity; • Unit; • Value; • Original description (extracted from the “product description” field on the invoice).
2	Logical and computationally recognized fields in rows and columns of the NFC-e database tables for sales identification:		• State (UF); • Municipality; • Issuer (Group C of NF-e); • Recipient (Group E of NF-e);
			Invoice number.
3	Grammar with terms (lexicon) that describe the product beer and that are initially identified.	Alphabet	• Extracted from State Resolution • No. 028/2023; • Extracted from reference ontologies.
4	Period of interest for auditing NFC-e.	Time	• Sample period provided by SEFAZ/AM: February 1, 2023, to May 31, 2023.
5	Unstructured fields of the “product description”: classification by equivalence or reciprocity of words related to terms describing beer (by statistics).	Category	• Identified examples: “BEER, BEER, BEER”; “CAN, CAN, CAN”; “LONGNECK, LONG, LN”
6	Unstructured fields in the “product description”: associations constructed	Hierarchy	• Examples identified: “BEER + SKOL + CAN + 269ML.”
	between terms in the alphabet (by the force of the rules between terms).		• “ ”+ “SKOL + LONGNECK + 350ML.
			Metadata example::

• Description: Chart 4 describes the complete cycle of the Ontology Learning Information Architecture Model for NFC-beer products, represented by four columns. The first column identifies the “hashtag” symbol. The second column describes the “data types.” The third column shows the “Information Architecture Record.” The fourth column describes the “Selected Result.” Six lines are shown in the chart. The chart has red borders and a salmon pink background.

The application of Information Architecture allows for in-depth semantic analysis, thereby extracting various blocks of knowledge such as domain terms, “is a” type relationships, and conceptual relationships, which are then validated and exported to a domain ontology.

5 RESULTS AND DISCUSSIONS

The objective of this session is to present and discuss the evaluation of the information architecture used to organize the data to be retrieved by text mining and used by AM to construct an ontology about beer.

These results are based on the requirements proposed at the end of the first layer, "Term Mining" for Ontology Learning (Table 3), as well as at the end of the entire cycle (Table 4). They represent a response from each element of the information architecture and serve as guidelines for constructing the ontology without the presence of a domain expert. Substituting the role of the expert with the organization of information by the information architecture requirements can ensure efficiency and time savings in data retrieval during text mining and analysis.

Furthermore, an analysis of the works listed in Table 2 reveals that they do not present an information architecture model prior to constructing the ontology (Table 1 steps) to assist in organizing, retrieving, and defining terms during mining. The techniques used are data mining preprocessing techniques intended to clean data collected randomly from the explored databases. However, data collection and selection are not performed according to organizational or architectural criteria. Some studies present requirements gathering as an initial phase. However, there is no guidance model or best practices for building functional requirements for searching terms and relationships.

Finally, applying information architecture iteratively and incrementally to the layers of ontology learning increases the likelihood of inserting new terms and concepts into the ontology while leaving a trace of the path used to define additional classes and relationships in the requirements documentation. This avoids unnecessary or duplicate information.

6 CONCLUSION

Ontologies are organized knowledge tools for various domains and interests. To overcome the time-consuming and difficult development process, AM is widely used to incorporate techniques into the data and text mining process for ontology learning. However, a major challenge is choosing and organizing significant data and information to accelerate learning and make the process efficient.

Domain experts in ontology learning generally provide data input for ontology development, and their input is derived from data mining and text mining performed on structured, semi-structured, and unstructured data volumes. Nevertheless, certain essential data and information organization activities for retrieving terms, concepts, or hierarchical relationships were not covered in the evaluated ontology construction models. With this in mind, activities could be improved in other cases.

The information architecture model for ontology learning focuses on requirements acquisition, which significantly contributes to the outcome. Architectural requirements are gathered from different perspectives and presented in five forms: Location, Alphabet, Time, Category, and Hierarchy. The model is iterative and incremental, serving as a reference that points out data and information for use in text mining and intelligent algorithms. This makes learning more efficient because the set of terms and their associations become more visible with each interaction.

The main step in the ontology construction cycle is improving conceptual modeling by incorporating information architecture theories to build a requirements engineering artifact. This step infers preferred paths for mining and natural language processing (NLP) techniques and extracts a list of new candidates for terms and relationships at each end of the cycle.

Finally, a presentation was given on an application for the electronic invoice environment for beer products, which generated results and facilitated understanding of the process.

Acknowledgements:

Not applicable.

References

• AGRAWAL, R.; SRIKANT, R. Fast Algorithms for Mining Association Rules. In: STONEBRAKER, M.; HELLERSTEIN, J. M. (ed.) Readings in database systems. 3. ed. San Francisco, CA: Morgan Kaufmann PublishersInc., 1994. p. 580-592. Disponível em: https://www.vldb.org/conf/1994/P487.PDF Acesso em: 13 jan. 2025.
  » https://www.vldb.org/conf/1994/P487.PDF
• ALPAYDIN, E. Introduction to Machine Learning. 3 ed. Massachusetts: MIT Press, 2014. Disponível em: https://dl.matlabyar.com/siavash/ML/Book/Ethem%20Alpaydin-Introduction %20to%20Machine%20Learning-The%20MIT%20Press%20(2014).pdf Acesso em: 13 jan. 2025.
  » https://dl.matlabyar.com/siavash/ML/Book/Ethem%20Alpaydin-Introduction %20to%20Machine%20Learning-The%20MIT%20Press%20(2014).pdf
• ASTROVA, I.; ARNE KOSCHEL, A.; LEE, S. How the Apriori Algorithm Can Help to Find Semantic Duplicates in Ontology. Joint Conference on Knowledge-Based Software Engineering. [S. l.]: JCKBSE, 2020. Disponível em: https://www.researchgate.net/publication/343195533_How_the_Apriori_Algorithm_Can_Hel p_to_Find_Semantic_Duplicates_in_Ontology Acesso em: 13 jan. 2025.
  » https://www.researchgate.net/publication/343195533_How_the_Apriori_Algorithm_Can_Hel p_to_Find_Semantic_Duplicates_in_Ontology
• BUITELAAR, P.; CIMIANO, P.; MAGNINI, B. Ontology learning from text: An overview. Ontol. Learn. fromtextMethods, Amisterdã,v. 123, p. 3-12, 2005. Disponível em: https://pub.uni-bielefeld.de/record/2497696 Acesso em: 7 jul. 2024.
  » https://pub.uni-bielefeld.de/record/2497696
• CHUNG, K.; YOO, H.; CHOE, D. Ambient context-based modeling for health risk assessment using deep neural network. Journal of Ambient Intelligence and Humanized Computing, Germany, v.11, p. 1387-1395, 2020. DOI: https://doi.org/10.1007/s12652-0181033-7
  » https://doi.org/10.1007/s12652-0181033-7
• DU, R.; AN. H.; WANG, K.; LIU, W. A short review for ontology learning: Stride to large language models trend. arXiv preprint arXiv:2404.14991, Nova York, v. 1, p. 1-14, 2024. Disponível em: https://arxiv.org/abs/2404.14991 Acesso em: 13 jan. 2025.
  » https://arxiv.org/abs/2404.14991
• EVARISTO, Emílio; DUQUE, Claudio Gottschalg. Recuperação da Informação em vídeo por meio de análise multimodal. In: DUQUE, C. G. (org.). Ciência da Informação Estudos e Práticas. Brasília: Thesaurus Editora de Brasília Ltda, 2011. p. 237-250.
• FACELI, Ana Carolina Lorena Katti et al. Inteligência artificial: uma abordagem de aprendizado de máquina. Rio de Janeiro: Editora LTC, 2021.
• FAYYAD, Usama.; PIATETSKY-SHAPIRO, Gregory; SMYTH, Padhraic. Da Mineração de Dados à Descoberta de Conhecimento em Bancos de Dados. AI Magazine, Arlington, v. 17, n. 3, p. 37-42, 1996. DOI: https://doi.org/10.1609/aimag.v17i3.1230
  » https://doi.org/10.1609/aimag.v17i3.1230
• FERNÁNDEZ, Mariano; GÓMEZ-PÉREZ, Asunción; JURISTO, Natalia. Methontology: from ontological art towards ontological engineering. In: AAAI SPRING SYMPOSIUM SERIES, 1997, Palo Alto. Proceedings… Palo Alto: AAAI Press, 1997. p. 25-43. Disponível em: https://aaai.org/papers/0005-ss97-06-005-methontology-from-ontological-art-towardsontological-engineering/. Acesso em: 27 set. 2023.
  » https://aaai.org/papers/0005-ss97-06-005-methontology-from-ontological-art-towardsontological-engineering
• FOX, Mark S. The TOVE project towards a common-sense model of the enterprise. In: BELLI, F.; RADERMACHER, F. J. (ed.) Industrial and engineering applications of artificial intelligence and expert systems. London: Springer-Verlag, 1993. p. 25-34.
• GHIDALIA, Sarah et al. Combining machine learning and ontology: A systematic literature review. arXiv preprint arXiv:2401.07744, Nova York, v. 2, p. 47-55, 2024. Disponível em: https://arxiv.org/abs/2401.07744 Acesso em: 2 set. 2024.
  » https://arxiv.org/abs/2401.07744
• GORAYEB, Diana Maria da Camara; DUQUE, Claudio Gottschalg. Proposta de metadados para descrição de produtos da Nota Fiscal de Consumidor Eletrônica (NFC-e) usando Apriori. P2P & Inovação, Rio de Janeiro, v. 11, n. 1, p. 1-24, e-7124, 2024. DOI: https://doi.org/10.21728/p2p.2024v11n1e-7124
  » https://doi.org/10.21728/p2p.2024v11n1e-7124
• GRUBER, Thomas R. Toward Principles for the Design of Ontologies Used for Knowledge Sharing. International Journal Human-Computer Studies, Oxford, v. 43, p. 907-928, 1995. DOI: https://doi.org/10.1006/ijhc.1995.1081
  » https://doi.org/10.1006/ijhc.1995.1081
• GRÜNINGER, Michael; FOX, Mark S. The role of competency questions in enterprise engineering. In: ROLSTADÅS, A. (ed.). Benchmarking: Theory and Practice. Boston: IFIP Advances in Information and Communication Technology, 1995. p. 22-31.
• HAENLEIN, Michael, KAPLAN, Andreas. A brief history of artificial intelligence: on the past, present, and future of artificial intelligence. California Management Review, Califronia, n. 61, p. 5-14, 2019. DOI: https://doi.org/10.1177/0008125619864925
  » https://doi.org/10.1177/0008125619864925
• HARIDY, Shaimaa et al. An ontology development methodology based on ontology-driven conceptual modeling and natural language processing: tourism case study. Big Data and Cognitive Computing, Basel, v. 101, n. 2, p. 1-23, 2023. DOI: https://doi.org/10.3390/bdcc7020101
  » https://doi.org/10.3390/bdcc7020101
• HASSAN, Bryar A.; RASHID, Tarik A. Artificial intelligence algorithms for natural language processing and the semantic web ontology learning. arXiv preprint arXiv:2108.13772, New York, p. 22-41, 2021. Disponível em: https://arxiv.org/abs/2108.13772 Acesso em: 2 ago 2024.
  » https://arxiv.org/abs/2108.13772
• HESSEN, Johannes. Teoria do conhecimento. Tradução João Vergílio Gallerani Cuter; revisão técnica Sérgio Sérvulo da Cunha. 2. ed. São Paulo: Martins Fontes, 2003.
• KUROKI JUNIOR, George Hideyuki; DUQUE, Claudio Gottschalg. Uma proposta de arquitetura da informação aplicada ao processamento de linguagem natural: uma proposta. Contribuições da Ciência da Informação no pré-processamento de dados para treinamento e aprendizagem de redes neurais artificiais. RDBCI, Campinas, e023002, 2023. DOI: https://doi.org/10.20396/rdbci.v21i00.8671396
  » https://doi.org/10.20396/rdbci.v21i00.8671396
• LISI, Francesca A. Building rules on top of ontologies for the semantic web with inductive logic programming. arXiv preprint arXiv:2108.13772, Ithaca, New York, v. 8, n. 3, p. 4760, 2007. Disponível em: https://arxiv.org/abs/0711.1814 Acesso em: 15 maio 2024.
  » https://arxiv.org/abs/0711.1814
• MAEDCHE, Alexander; STAAB, Steffen. Discovering conceptual relations from text. In: EUROPEAN CONFERENCE ON ARTIFICIAL INTELLIGENCE, 14., 2000, Berlin. Proceedings… Amsterdam: IOS Press, 2000. p. 321-325.
• MORI, Alexandre. Modelagem do conhecimento baseada em ontologias aplicada às Políticas Públicas e Habitação. 2009. Dissertação (Mestrado em Ciência da Informação) - Departamento de Ciência da Informação e Documentação, Universidade de Brasília, Brasília, 2009. Disponível em: https://repositorio2.unb.br/jspui/handle/10482/6880 Acesso em: 14 jul. 2023.
  » https://repositorio2.unb.br/jspui/handle/10482/6880
• NILSSON, N. J. Inteligência Artificial: uma abordagem moderna. 3. ed. Upper Saddle River: Prentice Hall, 2009.
• NOY, N. F.; MCGUINNESS, Debora L. Ontology development 101: a guide to creating your first Ontology. California:Stanford Knowledge Systems Laboratory Technical Report KSL-01-05; Stanford Medical Informatics Technical Report SMI-2001-0880, 2001. Disponível em: https://www.researchgate.net/publication/243772462_Ontology_Development _101_A_Guide_to_Creating_Your_First_Ontology Acesso em: 2 maio 2023.
  » https://www.researchgate.net/publication/243772462_Ontology_Development _101_A_Guide_to_Creating_Your_First_Ontology
• AMAZONAS. SEFAZ. Resolução Nº 0028 de 23 de novembro de 2023. Modifica a Resolução nº 011/2019-GSEFAZ, que estabelece o valor do preço médio ponderado a consumidor final - PMPF para cálculo do ICMS devido por substituição tributária nas operações com cervejas. Disponível em: https://sistemas.sefaz.am.gov.br/get/ Normas.do?metodo=viewDoc&uuidDoc=0262265b-8e4e-41ba-8eb0d73704adb5fa⪙=0b9bce66-dc46-47d2-ac78-dff7672f3dde Acesso em: 24 fev. 2024.
  » https://sistemas.sefaz.am.gov.br/get/ Normas.do?metodo=viewDoc&uuidDoc=0262265b-8e4e-41ba-8eb0d73704adb5fa⪙=0b9bce66-dc46-47d2-ac78-dff7672f3dde
• SIQUEIRA, André Henrique. Arquitetura da Informação: uma proposta para a fundamentação e caracterização de uma disciplina científica. 2012. Tese (Doutorado em Ciência da Informação) - Universidade de Brasília, Brasília, DF, 2012.
• SUMITHRA, R.; PAUL, Sujni. Using distributed apriori association rule and classical apriori mining algorithms for grid-based knowledge discovery. IEEExplorer, Piscataway, 2010. Disponível em: https://ieeexplore.ieee.org/document/5591577 Acesso em: 20 mar. 2023.
  » https://ieeexplore.ieee.org/document/5591577
• USCHOLD, Mike; GRUNINGER, Michael. Ontologies: principles, methods and applications. Knowledge Engineering Review, New York, v. 11, n. 2, p. 93-155, 1996. DOI: https://doi.org/10.1017/S0269888900007797
  » https://doi.org/10.1017/S0269888900007797
• WANG, Pei. On defining artificial intelligence. Journal of Artificial General Intelligence, Boston, v. 10, n. 2, p. 1-37, 2019. DOI: https://doi.org/10.2478/jagi-2019-0002
  » https://doi.org/10.2478/jagi-2019-0002
• WURMAN, R. S. Information Architects. New York: Graphis Inc, 1997.
• YANG, Bo. Construction of logistics financial security risk ontology model based on risk association and machine learning. Safety Science, Califórnia, v. 123, 2020. DOI: https://doi.org/10.1016/J.SSCI.2019.08.005
  » https://doi.org/10.1016/J.SSCI.2019.08.005
• ZOUAQ, Amal; GASEVIC, Dragan; HATALA, Marek. Towards open ontology learning and filtering. Information Systems, Illinois, n. 36, v. 7, p. 1064-1081, 2011. DOI: https://doi.org/10.1016/j.is.2011.03.005
  » https://doi.org/10.1016/j.is.2011.03.005

• 

  
  
• 

  
  
• Funding:
  Not applicable.
• 1
  It is an unsupervised AM algorithm frequently used in association rule learning.
• 2
  Current Resolution amending Resolution No. 011/2019-GSEFAZ, which establishes the weighted average price to the final consumer (PMPF) for calculating the ICMS due for tax substitution in beer transactions, within the scope of SEFAZ/AM.
• Ethical approval:
  Not applicable.
• Availability of data and material:
  Not applicable.
• Image:
  Extracted from the Lattes platform.
• JITA:
  BD. Information society
• SDG:
  16. Peace, justice, and effective institutions

Edited by

• Editor:
  Gildenir Carolino Santos https://orcid.org/0000-0002-4375-6815

Data availability

Not applicable.

Publication Dates

• Publication in this collection
  09 Mar 2026
• Date of issue
  2026

History

• Received
  15 Apr 2025
• Accepted
  26 Aug 2025
• Published
  09 Dec 2025