Beyond buzzwords: NLP reveals common threads in sustainable and circular construction discourse

Circular and sustainable construction

In academia, precise use of scientific terms is essential for clear communication among scientists from various fields (Rheude et al., 2021). Sustainable development and CE in construction frequently co-occur in academic literature (Norouzi et al., 2021; Antwi-Afari, Ng & Hossain, 2021; Mirzyńska et al., 2021), showing their conceptual interrelation and closeness. The overlap of these concepts is confirmed in Sauvé, Bernard & Sloan (2016). Generally, literature shows that CE aligns with sustainable development goals, while, apart from making product lifecycle paths closed-loop, CE’s scope is limited and needs to encompass broader dimensions, such as environmental, economic, and social pillars (Muñoz, Hosseini & Crawford, 2023; Mahanty et al., 2021). Huovila & Westerholm (2022) emphasize that circularity is a strategic approach to achieving sustainable development. Geissdoerfer et al. (2017) considered CE as an essential prerequisite for sustainable development.

However, CE’s social benefits were somewhat limited, predominantly focusing on job creation. Some works have elucidated definitions of environmentally friendly practices in the construction industry (Charef, Lu & Hall, 2022; Rheude et al., 2021; Hasheminasab et al., 2022; Ossio, Salinas & Hernández, 2023). Other scholars suggest that specific SC methods may not align with the CE framework. For instance, Charef, Lu & Hall (2022) stated that “design for” approaches might not be considered truly circular, as reuse potential of some materials is not infinite, thereby not conforming to the closed-loop principle. However, some approaches, such as “design for disassembly,” can still be useful for implementing CE principles in construction. Such strategies help extract and reuse materials efficiently at the end of a building’s life cycle, reducing waste and promoting more circular resource flow. Furthermore, SC methods often include durability and adaptability principles, which support CE’s aim of reducing resource consumption and waste generation. While Charef, Lu & Hall’s (2022) critique is valid for a perspective focusing on “material lifecycle,” it’s crucial to recognize the broader benefits and potential of SC practices at the “project lifecycle” level in advancing CE principles.

In contrast, Kirchherr, Reike & Hekkert (2017) discovered that definitions demonstrate limited direct connections between CE and sustainable development. In another study, it is argued that sustainable development and CE diverge in their fundamental objectives: sustainable development focuses primarily on societal goals, whereas CE centers on altering traditional production and consumption models (Sauvé, Bernard & Sloan, 2016). CE, through its focus on production and consumption activities, prevents negative environmental effects on other sectors, thus helping achieve sustainable development objectives. Moreover, Costa & Rodrigues (2023) claimed that CE largely depends on participant willingness and market-driven solutions, whereas sustainability demands stronger involvement from governments and policymakers.

Luo, Zhuo & Xu (2024) tie optimal NGO staffing to stronger circular-economy governance; Qiao et al. (2024) show sustainability drives knowledge evolution in inclusive tourism; Shi, Hayat & Cai (2024) unveil a unified open-vocabulary network for dense visual prediction; Bibi et al. (2023) optimise low-cost seismic isolation for masonry; Hu et al. (2024) map China’s rare-earth innovation network and its governance-market nexus. Collectively, they reinforce our interdisciplinary, resource-efficient approach to sustainable construction.

Natural language processing

NLP is a field of artificial intelligence (AI) that focuses on interaction between computers and humans through natural language. NLP has roots in linguistics, computer science and AI, and involves various tasks such as text analysis, machine translation, sentiment analysis, and information retrieval (Sawicki, Ganzha & Paprzycki, 2023). A primary application of NLP is content analysis, where it can uncover patterns, themes, and insights from unstructured text data. Techniques such as tokenisation and stop-word removal are crucial in this stage (Torbarina et al., 2023). After preprocessing, more advanced methods, such as part-of-speech tagging, named entity recognition, and syntactic parsing, are employed to extract meaningful features from text (Chung et al., 2023). NLP techniques can analyse academic articles, industry reports, and news articles to identify common themes and terminologies (Kesgin & Amasyali, 2024). Recent advances in NLP, such as transformer models like BERT and GPT, have significantly enhanced the accuracy and efficiency of text analysis tasks (Torbarina et al., 2023). These models leverage deep learning to understand context and semantics thoroughly, allowing for more nuanced and sophisticated analyses.

Bibliometric analyses have also been employed to study CE integration in the built environment, highlighting key thematic developments and challenges (Mhlanga, Haupt & Loggia, 2024). Another bibliometric research has provided a conceptual framework for understanding the role of CE in the built environment, mapping key research trends and influential publications (Silva Martins et al., 2024). Similarly, scientometric reviews have analyzed sustainability and sustainable development trends, providing insight into the evolution of SC research across multiple disciplines (Olawumi & Chan, 2018). Despite this focus, significant gaps remain in the practical understanding of circular economy and sustainability differences and commonalities in the construction industry.

Criteria for selecting literature sources for analysis

Data was gathered through a systematic literature review, employing two distinct searches within the Scopus database. The first search utilised keywords “circular AND economy AND construction OR buildings”, whereas the second focused on “sustainability AND construction OR buildings”. Both searches were confined to review articles published in the last three years (2021–2024). Figure 2 illustrates substantial growth in publications over recent years for both keyword searches. This trend demonstrates increasing attention these topics have received in academic research, particularly after 2020. The rapid surge in review publications highlights emerging interest in synthesising existing knowledge to address practical challenges and knowledge gaps. By focusing on the 2021–2024 period, our study captures this peak of scholarly activity and ensures relevance to the most current developments in circular economy and sustainability within the construction sector. We focused on the 2021–2024 timeframe to capture recent advancements and trends in circular economy and sustainability, ensuring our analysis reflects contemporary practices and terminologies. This approach also allowed effective dataset management while maintaining analytical depth using selected NLP methods. This strategy yielded 400 results for the first query and 1,626 for the second. To ensure relevance to the research topic, each article was carefully evaluated through thorough reading of its abstract. This process led to selecting 107 documents from the first query and 373 documents from the second query, deemed suitable for further NLP data analysis.

Data collection

After collecting data in PDF and HTML formats, conversion to a usable form was required. PDFReader (Maupin, 2018), Apache Tika (The Apache Software Foundation, 2024), and PDFMiner (Shinyama & Guglielmetti, 2023) were selected to handle PDF articles due to their robust text extraction capabilities and handling of complex document structures. Table 1 presents a summarised description of each tool. Multiple tools were utilised because certain PDF files had specific characteristics preventing a single method from extracting information effectively. Beautiful Soup (Richardson, 2023) was chosen for HTML files due to its effective HTML parsing and content extraction. We parsed 381 PDF and HTML files and successfully extracted the necessary texts for analysis. Each sample represents a single research article, providing a foundation for subsequent analysis. This dataset, titled “Sustainable and Circular Construction Terms”, was published Karaca et al. (2024).

Upon constructing our dataset, we organised it into a structured format. The dataset includes the following components: author information, paper title, abstract text, article text (in two representations: (1) a contiguous text block containing the entire article and (2) a segmented list of individual sentences), and quantitative metrics (length and complexity of each article, measuring the number of symbols, words, and sentences).

Data cleaning

Data cleaning procedures were essential to minimise noise and enhance the accuracy of subsequent text analyses, ensuring the dataset was structured and clean. This section describes the steps to refine the collected dataset. To focus analysis on core article content, we removed abstracts from the article_text_string’ since they were already stored separately in abstract’, making their presence redundant and potentially skewing text analysis metrics such as word and sentence counts. Similarly, all embedded URLs were removed as they are non-informative for textual analysis, ensuring subsequent studies, like frequency distributions or topic modelling, are based solely on substantive content.

Another cleaning step involved extracting and isolating references from each document. References indicate research depth and breadth but are peripheral to primary text analysis. By storing references in a dedicated ‘references’ column, we maintained access to this data without cluttering the primary text analysis fields.

Data preprocessing

Following the initial data cleaning phase, we conducted data preprocessing to refine our dataset to transform the raw text data into a structured and analytically processable format.

We removed all numerical digits from article texts during preprocessing to focus analysis on textual content and avoid the potential bias that numbers might introduce to text analysis metrics. We also removed mathematical symbols, which could similarly skew text analytics results by being mistaken for textual content.

These elements—numerical values, units, and equations—often reflect dataset-specific noise rather than consistent semantic meaning across documents. Since our study aims to extract thematic and conceptual patterns rather than quantitative analysis, retaining such elements would distort frequency-based metrics (TF-IDF) and introduce irrelevant nodes into graph-based models (TextRank). Their exclusion ensured that term extraction focused solely on meaningful linguistic constructs.

Another significant preprocessing step was removing stop-words from each article text. Stop-words—commonly used words with minimal individual meaning—contribute little to understanding text content and can disproportionately affect text statistics (Manning, Raghavan & Schütze, 2008). Their elimination enhances focus on relevant vocabulary within text mining.

We used the default English stop-word list from the Natural Language Toolkit (NLTK) library, which offers a balanced set of commonly non-informative terms suitable for academic discourse. We intentionally did not apply stemming or lemmatisation to preserve original lexical forms of technical and domain-specific terms (e.g., “recycling,” “recycled materials,” “resource recovery”), which might otherwise be distorted. This decision ensured higher fidelity in concept recognition and maintained semantic granularity important for later clustering and semantic annotation.

After these preprocessing steps, the statistics (the number of words, symbols, and sentences) for each article were calculated (see Table 2). Table 2 represents the descriptive statistics of quantitative metrics for number of symbols, words, and sentences in the collected data corpus.

The histograms in Fig. 3 demonstrate the distribution of sentences and words in the articles after preprocessing. Understanding these distributions helps identify the general structure and length of the articles (especially by looking at the number of words since we are going to find the most valuable terms from the articles), which can influence the effectiveness of the subsequent NLP analysis. The histogram for the number of sentences shows a right-skewed distribution, with a high frequency of articles having fewer sentences and a tail extending towards articles with more sentences. This indicates that while most articles are concise, a few are significantly longer. Similarly, the histogram for the number of words is also right-skewed, with most articles having a lower word count and fewer articles being lengthy. This skewness suggests a prevalence of shorter articles in the dataset, with a small proportion being more detailed.

To further illustrate the impact of preprocessing, Fig. 4 presents a comparative distribution of word and sentence counts before and after preprocessing. The histogram highlights how the text cleaning process, including stop-word removal, numerical filtering, and symbol elimination, affected the dataset. The shift in distributions demonstrates that preprocessing significantly reduced text length while maintaining the overall structure. This comparison validates the necessity of preprocessing, as it ensures that the data used for further NLP analysis is more refined and free from non-informative elements.

Description of the used NLP techniques and tools

To analyze textual data related to SC and CC in the construction sector, we employed three complementary NLP techniques: TF-IDF, a statistical method that quantifies term importance based on document frequency (Rajaraman & Ullman, 2011); TextRank, a graph-based ranking algorithm inspired by PageRank that extracts keywords through term co-occurrence relationships (Page et al., 1999); and semantic annotation, a knowledge-based approach linking text tokens to external knowledge sources like Wikipedia (Brank, Leban & Grobelnik, 2018; Klopper & Lubbe, 2012). Studies such as Obi et al. (2022) have demonstrated that automated text-mining techniques effectively identify key sustainability and CE trends in construction research. These methods complement each other: TF-IDF identifies statistically discriminative terms, TextRank captures term relationships through graph structures, and semantic annotation introduces external contextual knowledge for term disambiguation and enrichment.

Evaluation methods

Evaluation used three extraction methods: TF-IDF, TextRank, and Wikification, providing comprehensive concept coverage across scientific articles about sustainable and circular construction. Concept clustering grouped similar terms using vector-based similarity measures, identifying dominant themes and facilitating interpretation. Semantic analysis examined contextual meaning through term co-occurrences and external knowledge base linkage via Wikification. Expert evaluation by sustainability and construction specialists validated extracted concepts and associations, ensuring accuracy and relevance of identified themes. This framework provides a systematic analysis of evolving terminologies with reproducibility and structured comparison capabilities.

The data analysis flowchart in Fig. 5 outlines the analysis procedure, and the data collection and construction flowchart in Fig. 6.

To ensure comparability across extraction methods, we maintained a similar number of clusters (typically 5–7 per method) for TF-IDF, TextRank, and Wikifier outputs. This alignment facilitated thematic comparison across techniques and avoided disproportionate clustering that could hinder cross-method analysis. The number of clusters was guided by the elbow method and refined through interpretability. Cluster labels were manually assigned based on semantic consistency within each group of terms. Furthermore, we assessed the overlap and divergence between methods using Venn diagrams and word cloud visualisations, identifying shared and method-specific terms. This comparative evaluation allowed us to highlight consistent domain concepts and better understand the unique strengths of each extraction technique.

TF-IDF results and analysis

The results of the term clustering and their comparative analysis of terms related to CC and SC using the TF-IDF method are provided in Table 3. It is seen how CC and SC are similar and differ in terms of their content. To start with similarities, both CC and SC have such themes as “environmental impact assessment”. This indicates a shared concern for understanding and mitigating the negative effects of construction activities on the environment. Environmental impact assessment is crucial for both concepts as it provides a systematic process for evaluating the potential environmental consequences of proposed projects or policies, helping to ensure that such initiatives are sustainable and environmentally responsible.

Despite these similarities, there are notable differences between the terms. CC dataset is more inclined towards “waste and resource utilisation”, being more focused on “resource efficiency”. This emphasis on optimising resource use, minimising waste, and promoting recycling and reuse highlights CE’s goal of creating closed-loop systems where materials are continuously cycled back into production processes (Korhonen, Honkasalo & Seppälä, 2018).

In contrast, SC emerges as a more foundational and comprehensive concept, comprising a broader scope that extends beyond the building sector to include cities and energy usage. This reflects the three-pillar nature of sustainability, which balances environmental, social, and economic goals. Sustainability in construction encompasses a more holistic approach (Geissdoerfer et al., 2017; Muñoz, Hosseini & Crawford, 2023; Mahanty et al., 2021), contrasting with CC’s more targeted focus on resource cycles within specific industries or sectors.

The comparative analysis using the TF-IDF method further reveals that while CC focuses on practical applications and processes related to waste management and resource efficiency, SC covers a broader vision that integrates aspects of urban planning, community development, and energy management. For instance, terms related to “cities” and “land use” frequently appear in the SC dataset, emphasising its relevance to urban sustainability and regional planning.

TextRank results and analysis

TextRank method results for CC and SC key concepts are summarised in Table 4. Based on the obtained results, it is seen that both concepts have energy-related themes; however, their underlying key concepts differ. That is, CC is more focused on resource recovery, while SC emphasises energy-efficient building design and related materials, e.g., timber. It is interesting to note that the CC dataset contains the term “sustainability”, while the SC dataset does not contain the term “circularity”, which can be justified by sustainable development being a more fundamental concept than CE. The presence of the term “sustainability” within the CE dataset suggests that CE frameworks often incorporate sustainability principles, reflecting a nested relationship where CE practices contribute to broader sustainability goals (Korhonen, Honkasalo & Seppälä, 2018).

Both concepts share a theme of assessment and lifecycle analysis. However, CE is more inclined towards the resources lifecycle, focusing on aspects such as recycling, reusing, and minimising waste throughout the production and consumption cycles. On the other hand, sustainability pays more attention to the building lifecycle, including energy consumption, material durability, and overall environmental impact over time. This distinction demonstrates the different scales and scopes at which each concept operates, with CC being more process-oriented and SC being more holistic in its approach to built environments.

The SC concept operates more at a regional level, using keywords like “land use” and “cities.” This regional focus highlights sustainable development’s importance in adopting policies and practices impacting communities, infrastructure, and ecosystems at local and regional scales (Newman & Jennings, 2008; Orenstein & Shach-Pinsley, 2017; Thacker et al., 2019).

In contrast, CC focuses more on industry level, resources, and building use, reflecting practical implementation of circular practices within manufacturing, construction, and other sectors. The differences between SC and CC also imply varying approaches to policy and innovation. SC might drive policy development toward urban sustainability strategies, influencing zoning laws, building codes, standardisation, and transportation planning. Meanwhile, CC could stimulate innovation in product design and industrial processes, encouraging businesses to adopt circular practices.

Semantic annotations results and analysis

Table 5 presents the results of the semantic annotation method for the analysis of CC and SC key concepts. CC and SC share a focus on economic and technological systems, governance frameworks, and environmental protection. Both emphasise business, methodology, financial instruments, technology, and policy to guide construction practices. They also incorporate building materials, planning, and manufacturing to improve efficiency. However, CC prioritises trade, medium of exchange, alternative currency, and blockchain, highlighting its emphasis on economic value and resource circulation. It also focuses on data storage, software development methodology, and energy storage, underlining its reliance on technological innovation. SC, in contrast, integrates urban planning, infrastructure, logistics, and climate classification, demonstrating a broader concern for long-term environmental impact and societal well-being. SC also highlights environmental change, greenhouse gases, pollutants, and soil conditioners, indicating a stronger focus on ecological health. While CC is centred on optimising material and energy exchange, SC takes a more systemic approach, addressing governance, administrative entities, and political systems.

Table 6 shows the co-occurrence of concepts of Sustainability and CE. In general, this table is a sign of similarity between both concepts, as there is significant overlap in co-occurrence across all terms. The terms themselves also co-occur frequently, which aligns with existing literature (Norouzi et al., 2021; Mirzyńska et al., 2021; Antwi-Afari, Ng & Hossain, 2021). Talking about differences, it is evident that “Sustainability” more frequently co-occurs with all of the concepts compared to “Circular.” Particularly, examining the field of “Built Environment,” we see a significant difference in co-occurrence, with 0.13 for Sustainability and 0.05 for CE. This discrepancy suggests that sustainability principles are more deeply integrated into building and urban planning practices than the principles of the CE. In almost all of the concepts, the difference in co-occurrence frequency is large; however, in the case of “Waste Management,” it is almost equal. This could indicate that sustainable development is a broader topic than CE but also a fundamental concept that is related to all others.

This observation is further supported by the TF-IDF analysis (Table 7), which reveals that terms associated with Sustainable Construction appear more prominently in discussions concerning environmental impact, urban planning, and resource management. The high TF-IDF values for these concepts indicate their greater importance in sustainability discourse when compared to circular economy-related terms. This suggests that while CE is an integral part of sustainability, it remains a subset of a larger and more widely applied framework.

A more detailed perspective on the structural differences between CC and SC is provided by the TextRank analysis (Table 8). This method highlights that key nodes in the SC network are centred around energy efficiency, waste recycling, and life cycle management, reflecting its strong environmental focus. In contrast, CC places greater emphasis on digital technologies, resource exchange, and financial instruments, reinforcing its economic and technological orientation. These differences illustrate that while both frameworks aim for sustainability, they approach it from distinct perspectives: SC takes a more systemic and long-term environmental approach, while CC is driven by innovation and optimisation of economic resources.

Comparative analysis

This analysis aims to highlight the strengths and limitations of each approach and their agreement in identifying key terms related to SC and CE. By leveraging multiple extraction techniques—TF-IDF, TextRank, and Wikifier—we can observe how each method interprets textual importance through statistical frequency, graph-based relationships, or conceptual connections with external knowledge bases. Examining their individual outputs and intersections provides valuable insights into the consistency and uniqueness of extracted terms.

To visualise and quantify overlaps and divergences among these methods, we analyse their intersections using a Venn diagram¹ , followed by exploring word clouds² , which offer an intuitive representation of term prominence within each approach.

Venn diagram

The results presented in Fig. 7 allow us to assess the degree of agreement among the three methods to extract key terms. The highest number of unique terms was identified by the TF-IDF method (254 words), which can be explained by its statistical approach, where the importance of a word is determined by its frequency in the text relative to the entire corpus. Wikifier identified 200 unique terms, indicating its focus on conceptual connections with external knowledge bases rather than simple frequency analysis. TextRank, on the other hand, was the most conservative, extracting only 11 unique terms, as its algorithm is based on a graph model and identifies only the most significant words in terms of their interconnections.

The intersection of two methods also provides important insights. Specifically, 99 terms were simultaneously identified by both TF-IDF and Wikifier, indicating the presence of terms that are significant both in terms of statistical frequency and conceptual relevance. The methods TF-IDF and TextRank identified 171 common terms, highlighting their similarity in extracting key words despite their different approaches (statistics vs. graph-based model). The overlap between Wikifier and TextRank was minimal (14 terms), which can be explained by the fact that Wikifier relies on knowledge from external sources, while TextRank identifies key terms based on internal text relationships.

Of particular importance is the intersection of all three methods, where 45 terms overlapped. This number can be interpreted as an indicator of method agreement. Despite differences in algorithms, these words were recognized as important by all three approaches, confirming their special significance in the text. The higher this number, the more methods agree on the most important terms, indicating the high stability of these terms regardless of the extraction methodology. In our case, 45 is a relatively small proportion of the total number of words extracted, suggesting significant differences in the approaches of the methods. Thus, using multiple methods simultaneously provides a more comprehensive representation of key terms than applying a single method alone, as each contributes uniquely to text analysis.

Word clouds

To further compare the performance of TF-IDF, TextRank, and Wikifier in extracting key terms, word clouds were generated for each method (Fig. 8). A key observation is that TF-IDF tends to extract a broad set of terms, particularly those appearing frequently in the text. This method highlights both general and specific terms related to SC and CE. In particular, TF-IDF identifies a significant number of technical and engineering-related words, such as “yield stress”, “composite strength”, and “mechanical properties”. However, this statistical approach does not account for the contextual meaning of words, leading to the inclusion of some less conceptually relevant terms.

TextRank, in contrast, is more selective, extracting fewer but more contextually significant terms. The words identified by TextRank are often central to the thematic structure of the text, focusing on fundamental concepts such as “building energy efficiency”, “construction material”, and “sustainable building”. This reflects the method’s ability to capture interrelations between words through its graph-based ranking algorithm. However, TextRank appears to be more aligned with SC than CE, as fewer circular economy-specific terms emerge compared to TF-IDF and Wikifier.

Wikifier differs from both TF-IDF and TextRank by integrating external knowledge bases, resulting in a more conceptually rich set of extracted terms. This method emphasises broader sustainability concepts such as “sustainability”, “carbon footprint”, and “industrial revolution”, making it particularly effective in linking terms to well-established knowledge domains. Additionally, Wikifier identifies interdisciplinary terms such as “intellectual property” and “geographical information system”, which are not as prominent in the other methods. This suggests that Wikifier is better suited for capturing domain-specific knowledge beyond mere word frequency or internal text structure.

When comparing the agreement between methods, we observe that:

1. TF-IDF and TextRank share a notable number of key terms, particularly in the domain of construction materials and energy efficiency. This suggests that despite their methodological differences, both approaches recognise frequently occurring words that are also structurally important within the text.

2. TF-IDF and Wikifier exhibit strong overlap in statistical and conceptually relevant words, highlighting terms that are both frequent and knowledge-driven.

3. TextRank and Wikifier, however, have minimal intersection, as TextRank prioritises structural importance within the text, while Wikifier derives importance from external domain knowledge.

Overall, the word clouds demonstrate that each method brings own perspective to keyword extraction. While TF-IDF provides a broad statistical perspective, TextRank refines the selection by focusing on structurally significant words, and Wikifier extends the analysis by integrating external knowledge. The combination of these approaches allows for a more comprehensive understanding of key terms within the domains of SC and CC.

Reproducibility

Dataset (DOI/URL): https://data.mendeley.com/datasets/6w74d7x8s4/2 (DOI: 10.17632/6w74d7x8s4.2); DOI: 10.5281/zenodo.15510952, v1.0 AlexPak/paper-2024-nu-sustainable-constr.

Computing infrastructure (operating system, hardware etc) requirements:

All experiments were conducted on the following computing environment:

• –

  CPU: 2 $\times$ Intel® Xeon® Gold 6354

• –

  RAM: 128 GB DDR4

• –

  GPU: 2 $\times$ NVIDIA A100 (40 GB each)

• –

  Storage: High-speed NVMe SSD

• –

  OS: Ubuntu 22.04 LTS

• –

  Python: Version 3.10

Software and dependencies:

The analysis was implemented in Python using the following major packages:

$\bullet$ pandas $\bullet$ numpy $\bullet$ scikit-learn $\bullet$ nltk $\bullet$ matplotlib $\bullet$ networkx $\bullet$ gensim $\bullet$ transformers $\bullet$ sentence-transformers.

A full list of Python packages with pinned versions is provided in the accompanying requirements.txt file, which is submitted as supplementary material alongside this manuscript. This ensures full reproducibility of the results.

To install all dependencies, the following command can be used:

pip install -r requirements.txt

Description of models used

Justification for model type used:

The selection of models and methods in this study is driven by the need to explore, analyze, and differentiate the nuances between circular construction and sustainable construction in the literature. Each model was carefully chosen to highlight different dimensions of this relationship:

• Concept Matrix approach: The Concept Matrix model was chosen to capture contextual associations between key terms, as it provides a detailed mapping of how circular construction is situated within the broader discourse on sustainability. This model allows for a comparison between the sector-specific focus of circular construction and the more general nature of sustainable construction.

• Clustering methods: Clustering was employed to group terms found in the literature and visually represent the conceptual landscape. This method was selected for its ability to reveal patterns and groupings in a way that is both interpretable and meaningful for understanding the relationship between different concepts in the dataset.

• TextRank algorithm: TextRank was selected due to its strength in ranking terms based on importance within a graph-based structure. This method highlights the hierarchical and nested relationships between circular economy and sustainability concepts, aligning with the study’s goal of showing how circular construction fits within broader sustainability frameworks.

• TF-IDF approach: The TF-IDF method was selected for its effectiveness in distinguishing between terms that are more specific to circular construction compared to those associated with sustainability. This aligns with the objective of pinpointing the more targeted nature of circular construction as compared to the broader scope of sustainable construction.

  Assessment metrics (Justification):

• NLP similarity: A measure of semantic similarity between terms, based on contextual embeddings. This helped identify closely related concepts within the corpus.

• Rank: For terms extracted via TextRank, a rank value was assigned based on their centrality in the term co-occurrence graph. Higher-ranked terms are considered more crucial for understanding the key topics in the text.

• Importance (TF-IDF): The TF-IDF scores were used to assess the relative importance of terms in individual documents. Terms with higher TF-IDF scores are indicative of their significance within a particular document or set of documents.

• Entropy: The entropy of the term distribution was calculated to assess the diversity of terms within the corpus. A higher entropy value indicates a wider variety of terms used across the dataset.

• Graph metrics: The TextRank graph was analyzed using various network-based metrics, including:

  • Node centrality: Identifies the most influential terms based on their connectedness in the co-occurrence network.

  • Degree: The number of direct connections a term has with other terms.

  • Clustering coefficient: Measures how tightly connected the neighbors of a term are, providing insights into local term clusters.

Repository link: https://github.com/AlexPak/paper-2024-nu-sustainable-constr.

This repository is organized into several Jupyter Notebooks that represent the different methods used in the study:

ConceptMatrix.ipynb: Implements the semantic annotation approach to discover contextual associations between key terms in the dataset, revealing how circular construction is more operational and sector-specific compared to sustainable construction.

Clustering.ipynb: Contains code for clustering the dataset based on the terms found using NLP methods, providing a clear visual representation of how different concepts are grouped within the literature.

Evaluation_Methods.ipynb: Includes the evaluation metrics used to assess the performance and relevance of the NLP methods in identifying key themes in the literature.

TextRank_approach.ipynb: Applies the TextRank algorithm to reveal the relationships between circular economy and sustainability concepts, showing the nested nature of circular economy principles within broader sustainability goals.

TFIDF_approach.ipynb: Demonstrates the use of the TF-IDF method to differentiate between the targeted approach of circular construction and the more comprehensive nature of sustainable construction.