This paper explores the growing importance of Environmental, Social, and Governance (ESG) criteria in ifnancial assessments and conducts an AI-driven analysis of ESG concepts' evolution from 1980 to 2022. Focusing on media sources from the United States and the United Kingdom, the study utilizes the Dow Jones News Article dataset for a comprehensive analysis focused on the environmental domain. The research introduces an innovative information extraction technique, transforming extracted data into a knowledge graph. Key findings highlight recent trends in ESG aspects, with a notable emphasis on climate change, renewable energy sources, and biodiversity conservation in the environmental dimension.
CEUR ceur-ws.org
Over the past few years, there has been a growing significance in using Environmental, Social, and Governance (ESG) criteria for assessing financial investments 1 . The European Parliament has recognized the importance of ESG ratings in its legislative endeavors to foster an economy that truly serves the interests of the people. This recognition has resulted in specific initiatives, including the implementation of the EU taxonomy for sustainable activities2, a resource aiming to define a set of ESG standards for organizational conduct. It serves as a valuable tool for socially conscious investors assessing potential investments.
Monitoring and analyzing the portrayal and evolution of ESG-related concepts is crucial for
assessing changing perceptions in media and public opinion on sustainability and diversity [
To address this constraint, researchers have proposed diferent methods to create structured,
interconnected, and machine-readable data frameworks for analyzing news [
In this paper, we present an AI-powered examination of ESG concepts and their development spanning 1980 to 2022. The focus is on media outlets in the United States and the United Kingdom, encompassing notable publications like The Guardian, The New York Times, and The Times. The primary dataset employed for this investigation is the Dow Jones News Article dataset9, recognized for its extensive and high-quality compilation of news articles.
Our approach utilizes advanced information extraction techniques to condense relevant information from articles into structured statements, represented as triples (<subject, predicate, object>). The operational pipeline developed for this process is versatile. It can be implemented on a standard server, eliminating the need for extensive computational resources typically required by current large-scale language models for processing vast data sets. The primary advantage of this innovative approach lies in its capacity to analyze various entity types (e.g., 2EU taxonomy for sustainable activities - https://finance.ec.europa.eu/sustainable-finance/tools-and-standards/ eu-taxonomy-sustainable-activities_en 3Brandwatch - https://www.brandwatch.com/ 4Brand24 - https://brand24.com/ 5Repustate - https://www.repustate.com/ 6Cision Communication Cloud - https://www.cision.com/ 7SentiOne - https://sentione.com/ 8Meltwater - https://www.meltwater.com/ 9Dow Jones News Article dataset - https://developer.dowjones.com/datasets/details/news organizations, persons, topics) while establishing meaningful relationships between entities based on predicates extracted from the articles. Consequently, it serves as an efective tool for analyzing substantial volumes of news content, gaining insights into key concepts, and comprehending the evolution of discourse over time.
In detail, our paper contributes in the following ways: • We ofer an AI-driven analysis of the news discourse on ESG concepts spanning from 1980 to 2022. • We introduce a comprehensive and automated pipeline designed for creating a Knowledge
Graph (KG) from a collection of news documents. • We provide various analytics on ESG concepts, delving into entities and statements derived from a KG extracted from the news.
Section 2 explores prior works related to KGs in news. Section 3 outlines the general pipeline utilized for KG generation. In Section 4, the data source is outlined, and an overview of the resultant KG centered on ESG aspects is provided. Section 5 delves into the analysis results. Finally, Section 6 concludes the paper.
The central aim of incorporating Knowledge Graphs (KGs) into news analysis is to depict and
establish connections among diverse entities within the news domain, encompassing individuals,
locations, events, topics, and factual information. This systematic representation enables a
more insightful examination of shifts in discourse over time. For example, Al-Obeidat et al. [ 19]
constructed a KG focused on COVID-19-related news, providing a platform for researchers
and data analysts to address the challenges posed by the pandemic. Gangopadhyay et al. [20]
analyzed a knowledge graph of online claims showing how misinformation can spread on the
web and the extensive work for verifying online discussed facts. Tan et al. [
In contrast, our paper introduces a KG as a pivotal element in our AI-driven analysis of ESG sectors. Uniquely, our KG is crafted to specifically facilitate the exploration of the evolution of ESG discourse over time, providing a detailed representation of various entity types.
The constructed pipeline comprises two main phases. In its initial stage, a Text Parsing Module is employed to extract entities and their relationships from a set of news articles. The subsequent stage involves a three-step process to generate the knowledge graph. Firstly, the Entity Extraction Module identifies crucial entities and classifies them by type. Subsequently, the Relationship Extraction Module discerns relationships among these entities from the news articles. Lastly, the Triple Refinement Module concludes the process by refining the resulting triples, yielding the finalized knowledge graph.
The Text Parsing Module relies on the Stanford CoreNLP10 suite, an extensive collection of natural language processing tools developed in Java. Utilizing the Part-of-Speech (PoS) Tagger, this module assigns tags to each word in the provided text, identifying and classifying tokens based on their grammatical categories (e.g., preposition (PRP), verb (VB), noun (NN), adjective (JJ), etc.). Additionally, the module constructs a dependency tree for each sentence.
Within the Entity Extraction Submodule, nominal phrases are identified as entities for the knowledge graph. Nominal phrases constitute word groups with a noun or pronoun as the primary word, accompanied by modifiers, determiners, and complements ofering additional information about the noun (e.g., ‘long news article’). On the other hand, the Relationship Extraction Submodule identifies connections between entities. For every sentence , all the shortest paths of the dependency tree between each pair of entities ( , )| , ∈ containing a verb are selected. This process yields various types of paths between entities, and the analysis of these paths determines the most suitable ones for identifying relationships in the given context. The paths used in [25] were employed for this purpose.
The Triple Extraction Submodule performs three primary tasks: 1) relation refinement, 2) entity refinement, and 3) triple refinement. The set of triples , generated in the preceding step, may include triples with similar meanings but expressed through diferent verbs, for example, <company, build, 200-unit motel>, <company, construct, buildings>, <craftsmen, create, accommodation>. Relation refinement aims to identify the most suitable predicate label for each relation verb in a triple < , , > and map to in the resulting triple. This phase reduces the space of possible relationships by analyzing the resultant verbs and clustering them into a more concise set of well-defined relationships [ 25]. The method was applied to the 393 verbs found in all the triples extracted from the ESG news dataset, resulting in a final set of 57 predicates.
The Entity Refinement module establishes an index based on the tokens contained within the entities. This index links each token to all entities that contain it. For instance, the token Obama is linked to entities such as Barack Obama, President Obama, former President Barack Obama, Barack Obama’s Administration, Michelle Obama, and so on. Entities and ∈ are compared if they share at least one token. This comparison is executed using the state-of-the-art framework SentenceTransformers11, encoding the entities with the all-mpnet-base-v212 transformer model. If the cosine similarity between entity and exceeds 0.9, they are grouped into the same cluster. The reader notices that this value has been selected based on an empirical analysis that considered the values 0.7, 0.8, 0.85, 0.9, and 0.95.
For the Triple Refinement module, akin to the entity refinement step, a sentence transformer model is used to detect and merge triples with the same meaning. As a final step, the resulting 10Stanford CoreNLP - https://stanfordnlp.github.io/CoreNLP/ 11SentenceTransformers - https://huggingface.co/sentence-transformers 12all-mpnet-base-v2 - https://huggingface.co/sentence-transformers/all-mpnet-base-v2 triples are linked to the original papers and employed to construct the knowledge graph. Each triple is associated with its support, indicating the number of news articles from which it was extracted. To evaluate the pipeline’s accuracy, a sample of triples underwent assessment by three reviewers, considering both the triple and the original sentences from which it originated. The reviewers marked the triple as 1 if it accurately reflected the news articles’ content and 0 if it did not. The average agreement between annotators was 0.89, indicating substantial consensus. The pipeline’s accuracy, evaluated against the majority vote of the three annotators for 200 statements, was 0.85, with individual rater estimates ranging from 0.85 to 0.93, demonstrating the pipeline’s ability to extract triples with high accuracy. More information about this pipeline can be found in [26].
The Dow Jones News Datasets encompass an extensive compilation of 15, 105, 283 news articles spanning various languages. The dataset incorporates 13 English sources, including renowned ones like The Wall Street Journal, New York Times, and The Guardian, contributing to a total of 7.3 million distinct news items. In assembling a repository of news articles concerning ESG, we considered all news from 1980 to 2022 containing keywords related to Environmental, Social, and Governance, either within the text body or metadata fields. The ultimate collection comprises approximately 850, 000 news articles, distributed as 500, 000 on environmental topics, 290, 000 on social issues, and 60, 000 on governance. The pipeline detailed in Section 3 was applied to this set of 850, 000 ESG news articles, resulting in a KG that includes over 7.2M statements and 4M entities.
For structuring the statements, we utilized a lightweight ontology tailored to the primary purpose of aiding news analysis. The ontology defines four main classes: i) aggregated statement, ii) fine-grained statement, iii) News, and iv) Entity. It also specifies 57 object properties derived from the predicates outlined in Section 3. Furthermore, the ontology maps the statements using the original verb alongside their version utilizing the 57 predicates obtained by clustering them.
Each statement in ESG-KG incorporates: - rdf:subject, rdf:predicate, and rdf:object, providing the reification of triples within an rdf:Statement; - provo:wasDerivedFrom, supplying provenance information and listing the DNA-IDs of the news from which the statement is derived; - esgkg:statement_negated, a boolean indicating whether the statement was derived from a negative sentence (True) or not (False); - esg-kg:original_triple, listing the fine-grained versions of the statement.
Additionally, each news ID is linked to xsd:date, ofering the news publishing date, and esg-kg:source, providing the original journal source name.
The analysis presented in this section relies on various analytics derived from the knowledge graph.
Figure 1 illustrates the distribution of the three primary topics (Environmental, Social, and Governance) over time. Historically, environmental issues have consistently dominated, representing 55% to 75% of coverage. However, a noticeable trend shows an increasing focus on social issues, growing from approximately 20% in 1980 to nearly 40% by 2022. This shift seems to be propelled by heightened interest in subjects such as ethics, racism, gender identity, and global human rights.
The environmental component encompasses 2 million entities, the social aspect covers 361,000 entities, and the governance section comprises 209,000 entities. These entities are categorized based on the Named Entity Recognition (NER) tool provided by Spacy, and their types and frequencies are detailed in Table 1.
The KG includes 3.8 million statements: 3 million statements about environmental topics, 600,000 related to social issues, and 236,000 concerning governance.
We present the foremost ten political groups (Table 2), geopolitical entities (Table 3), prominent individuals (Table 4), and organizations (Table 5) for each category.
An impactful observation underscores the USA’s significant role in ESG discourse, evident in the top three groups comprising Democrats, Republicans, and Americans. Moreover, frequently mentioned individuals center around US Presidents, encompassing figures like Bush, Obama, Clinton, and Trump. The United States emerges as the most cited country in articles related to Environmental and Social aspects, underlining its perceived leadership status in these domains. Conversely, China takes the lead in discussions on Governance, with Beijing and Japan often mentioned in conversations about employee rights, securing the fourth and fith positions among the most referenced countries for Governance.
Our exploration then delves into the Environmental domain, scrutinizing the evolving trends of key entities. This involves computing the annual frequency of each entity and applying linear regression to discern the trajectory of these entities’ yearly distributions. The slope of the regression line serves as an indicator of the trend’s momentum, with a steeper slope signifying a more rapid surge in media coverage for the specified entity. This analytical technique, commonly employed to detect key trends, finds application in areas such as research topics [ 27]. Results are presented in Table 6, where slope_10 denotes the trend over the last 10 years, and slope_5 indicates the trend over the last 5 years. To accentuate common themes, related entities have been manually clustered and highlighted in the same color.
Entity
The environmental facet of ESG centers on assessing a company’s influence on the natural world and its approach to managing environmental risks. Table 6 presents entities that have exhibited notable increases in mentions over the past decade and the last five years. Freq.
Slope 10 Years
Slope 5 Years • Climate and Carbon Emissions 6: Central to discussions involving the measurement of carbon footprints, implementation of initiatives to reduce greenhouse gas emissions, and the formulation of strategies to mitigate the efects of climate change [ 28]. The increased visibility of these entities in news narratives underscores the growing importance of taking concrete measures to combat climate change. • Renewable Energy: Positive trends suggest a rising emphasis on the use of renewable energy sources and the adoption of energy-saving practices in public discourse. This shift towards energy eficiency reflects broader societal and economic recognition of the beneifts associated with sustainable energy practices. With the increasing urgency to address climate change, the push for more eficient energy usage and the transition to renewables becomes a central theme in policy, industry, and community conversations [29]. • Biodiversity and Land Use: The upward trend over the past five years underscores the significance of these issues. Specifically, the focus on deforestation and biodiversity highlights the media’s growing concern and interest in the conservation of natural habitats, ecosystems, and biodiversity. This focus aligns with global eforts to achieve biodiversity conservation targets and sustainable development goals [30], emphasizing the need for a holistic approach to environmental stewardship that includes protecting diverse ecosystems and ensuring responsible land use.
In this manuscript, we provide a preliminary examination of ESG concepts and their evolutionary trajectory spanning from 1980 to 2022, concentrating on news content sourced from the United States and the United Kingdom. Employing the Dow Jones Article dataset, our investigation encompasses news articles from well-known newspapers like The Guardian, The New York Times, and The Times. To execute this analysis, we initially applied an extraction pipeline to the news articles, involving the organization of extracted data into a Knowledge Graph (KG). The methodology employed advanced information extraction techniques to distill pertinent information from articles into structured statements represented as triples. These triples underwent aggregation, and verification, and were used in constructing a comprehensive knowledge graph. The implemented pipeline is versatile, applicable across domains, and facilitates the analysis of various entity types while establishing semantic relationships between them based on information extracted from news articles. The information extraction pipeline underwent rigorous evaluation by three annotators, achieving an accuracy of 0.85. Subsequently, the resulting knowledge graph was utilized to scrutinize the three core components of Environmental, Social, and Governance (ESG), with a specific focus on the environmental domain. In future work, we plan to overcome some limitations that we have encountered to further improve the generation pipeline. First, we would like to experiment and develop a novel model to merge entity mentions that refer to the same entity (for example, in Table 4 Mr. Reagan and President Reagan are not merged). Second, we would like to assign news into categories to make it simple the explore both the news and the KG content. Finally, we intend to explore additional datasets beyond Dow Jones Article dataset to enrich our analysis and validate the ifndings across diferent sources, thereby enhancing the robustness and generalizability of our research.
We acknowledge financial support under the National Recovery and Resilience Plan (NRRP), Mission 4 Component 2 Investment 1.5 - Call for tender No.3277 published on December 30, 2021 by the Italian Ministry of University and Research (MUR) funded by the European Union – NextGenerationEU. Project Code ECS0000038 – Project Title eINS Ecosystem of Innovation for Next Generation Sardinia – CUP F53C22000430001- Grant Assignment Decree No. 1056 adopted on June 23, 2022 by the Italian Ministry of University and Research (MUR). methodology for knowledge graph generation within the tourism domain, IEEE Access (2023). [11] M. Rizun, et al., Knowledge graph application in education: A literature review, Acta
Universitatis Lodziensis. Folia Oeconomica 3 (2019) 7–19. [12] A. Tchechmedjiev, P. Fafalios, K. Boland, M. Gasquet, M. Zloch, B. Zapilko, S. Dietze, K. Todorov, Claimskg: A knowledge graph of fact-checked claims, in: The Semantic Web–ISWC 2019: 18th International Semantic Web Conference, Auckland, New Zealand, October 26–30, 2019, Proceedings, Part II 18, Springer, 2019, pp. 309–324. [13] A. Meloni, S. Angioni, A. Salatino, F. Osborne, D. R. Recupero, E. Motta, Integrating conversational agents and knowledge graphs within the scholarly domain, Ieee Access 11 (2023) 22468–22489. [14] T. M. Nguyen, H.-W. Chun, M. Hwang, L.-N. Kwon, J.-M. Lee, K. Park, J. J. Jung, Sociopedia+: a visual analytics system for social knowledge graph-based event exploration, PeerJ Computer Science 9 (2023) e1277. [15] S. Angioni, A. Salatino, F. Osborne, A. Birukou, D. R. Recupero, E. Motta, Leveraging knowledge graph technologies to assess journals and conferences at springer nature, in: International Semantic Web Conference, Springer, 2022, pp. 735–752. [16] A. ZEKIYE, A. ALPKOCAK, Extracting and analyzing covid-19-related entities and relationships: A knowledge graph approach, in: Proceedings of 14th Turkish Congress of Medical Informatics, volume 16, 2023. [17] D. Dessí, R. Dessí, Diving into knowledge graphs for patents: Open challenges and benefits (2022). [18] P. Manghi, A. Mannocci, F. Osborne, D. Sacharidis, A. Salatino, T. Vergoulis, New trends in scientific knowledge graphs and research impact assessment, 2021. [19] F. Al-Obeidat, O. Adedugbe, A. B. Hani, E. Benkhelifa, M. Majdalawieh, Cone-kg: A semantic knowledge graph with news content and social context for studying covid19 news articles on social media, in: 2020 Seventh International Conference on Social Networks Analysis, Management and Security (SNAMS), 2020, pp. 1–7. doi:10.1109/ SNAMS52053.2020.9336541. [20] S. Gangopadhyay, K. Boland, D. Dessí, S. Dietze, P. Fafalios, A. Tchechmedjiev, K. Todorov,
H. Jabeen, Truth or dare: Investigating claims truthfulness with claimskg (2023). [21] D. Liu, T. Bai, J. Lian, G. Sun, W. X. Zhao, J. rong Wen, X. Xie, News graph: An enhanced knowledge graph for news recommendation, in: KaRS@CIKM, 2019. URL: https://api. semanticscholar.org/CorpusID:204777685. [22] M. Rospocher, M. Van Erp, P. Vossen, A. Fokkens, I. Aldabe, G. Rigau, A. Soroa, T. Ploeger, T. Bogaard, Building event-centric knowledge graphs from news, Journal of Web Semantics 37 (2016) 132–151. [23] L. Fu, H. Peng, S. Liu, Kg-mfend: an eficient knowledge graph-based model for multidomain fake news detection, The Journal of Supercomputing (2023) 1–28. [24] A. L. Opdahl, T. Al-Moslmi, D.-T. Dang-Nguyen, M. Gallofré Ocaña, B. Tessem, C. Veres, Semantic knowledge graphs for the news: A review, ACM Comput. Surv. 55 (2022). URL: https://doi.org/10.1145/3543508. doi:10.1145/3543508. [25] D. Dessí, F. Osborne, D. R. Recupero, D. Buscaldi, E. Motta, Scicero: A deep learning and nlp approach for generating scientific knowledge graphs in the computer science domain,