This paper investigates the identification and ontological classification of domain-specific entities to enable large-scale analysis of environmental discourse. While general-purpose Named Entity Recognition (NER) systems reliably detect standard categories such as persons, organizations, and locations, specialized domains like environmental communication require the recognition of additional, domain-relevant entities. These entities, often realized as common nouns, represent abstract, evolving concepts that are highly dependent on context and vary across languages. To address this challenge, we compare two pipelines for identifying domain-specific environmental entities in a bilingual corpus of WWF Living Planet Reports: (i) a traditional NLP pipeline that extracts noun phrases using dependency syntax parsing and matches them to BabelNet and GEMET, and (ii) a Large Language Model (LLM)-based pipeline that uses prompt-based instructions to both extract noun phrases and generate corresponding ontology matches. We evaluate the coverage of each approach and analyze the most frequent mapped entities to identify key environmental concepts emphasized in WWF discourse. To further assess the capabilities of LLMs in ontology-based annotation, we also prompted the LLM to generate GEMET-style definitions for phrases not found in the ontology. Our findings contribute practical insights for developing robust, ontology-enriched methods for environmental discourse analysis and knowledge extraction. Though tested on environmental texts, the framework can generalize to other domains via suitable ontologies and extraction rules.
CLiC-it 2025: Eleventh Italian Conference on Computational Linguistics, September 24 — 26, 2025, Cagliari, Italy * Corresponding author. † These authors contributed equally. $ elisa.chierchiello@unito.it (E. Chierchiello); pchiril@uchicago.edu (P. Chiril); apagano@ufmg.br (A. Pagano) © 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License adapted extensions show good performance to detect Attribution 4.0 International (CC BY 4.0).
domain mentions [3, 7, 8]. Hybrid architectures, such as the Ontology-Attention Layer, demonstrate that coupling LLMs with explicit ontology guidance further improves Generic NER has been extensively studied as a foundaaccuracy in specialized contexts [9]. tional NLP task, with systems reliably detecting persons,
Despite this progress, to the best of our knowledge, organizations, and locations [1, 2]. However, as Marrero no study has systematically compared rule-based NLP et al. [4] and Zhang et al. [3] observe, such systems perpipelines and LLM-based methods for domain entity form poorly when applied to specialized domains because recognition in multilingual environmental discourse. Our domain-specific concepts are often expressed through study addresses this gap by focusing exclusively on common noun phrases rather than proper names, and domain-specific entities, with conventional named enti- thus lack the distinctive lexical or orthographic cues that ties such as persons, organizations, and locations to be standard NER methods exploit. examined in future work, and implementing two distinct To address these limitations, domain-specific NER has pipelines for their identification and classification. by im- been pursued to handle technical and abstract terminolplementing and comparing two pipelines on a bilingual ogy in specialized texts. In the biomedical field, for incorpus of WWF Living Planet Report Executive Sum- stance, Zhang et al. [10] review biomedical entity recogmaries (2014–2024). The first pipeline uses dependency nition as an example of domain-focused extraction, highparsing to extract noun phrases and matches them to lighting the essential role of ontologies for semantic preciBabelNet and GEMET using string-based similarity. The sion. In geosciences, Villacorta Chambi et al. [11] pursue second pipeline uses a prompt-based LLM to both detect NER improvement through the use of specialized geolognoun phrases and suggest ontology matches directly. We ical schemas. then assess the coverage of each approach and qualita- Ontology-based approaches to domain-specific entity tively examine the most frequent shared mapped entities recognition have been widely explored. García-Silva et to highlight the core concepts that characterize WWF en- al. [5] proposed an ontology-based pipeline for environvironmental discourse. To address these aims, this study mental data that uses dependency parsing to identify investigates the following research questions: candidate terms and maps them to structured environmental ontologies. Zhou and El-Gohary [6] developed • How much coverage do a dependency syntax- a syntax-driven framework to extract provisions from based pipeline and a prompt-based LLM pipeline environmental regulations and link them to a complieach achieve when extracting and mapping noun ance ontology, demonstrating high precision for domainphrases to BabelNet and GEMET in a bilingual specific phrases. Wei et al. [ 9] integrated an ontologycorpus of WWF Living Planet Reports? attention mechanism within BERT to improve medical • Which environmental concepts emerge as the entity recognition, while Dai et al. [12] emphasized the most frequently mapped entities and what do combination of entity recognition and ontology linking these frequent concepts reveal about thematic to build domain-specific knowledge graphs. emphases in WWF environmental discourse? More recently, LLMs have emerged as powerful tools for general and domain-specific entity recognition. These
In order to answer these questions, we assess the two LLMs can complement ontology-based systems by propipelines in terms of coverage — that is, the number viding contextual understanding for domain terms that and proportion of extracted noun phrases that can be lack consistent surface forms. mapped to domain concepts in BabelNet and GEMET — Cross-lingual and multilingual methods support conand then examine the most frequently mapped entities sistent domain entity alignment across languages. Navto highlight key environmental concepts emphasized in igli and Ponzetto [13] presented BabelNet, a multilinWWF discourse. gual lexical network used for semantic linking. GEMET
Finally, to explore how LLMs can contribute to ex- [14] serves as a domain-focused environmental thesaurus, panding domain ontologies, we prompted the LLM to while Ryu et al. [15] and Zhao et al. [16] show how generate GEMET-style definitions for entities that could such resources help maintain terminological coherence not be matched in GEMET. in translation and cross-lingual NLP.
By comparing these two pipelines, we highlight practi- A disciplinary field that directly benefits from precise cal considerations for building ontology-based workflows domain entity recognition supported by environmenfor semantic search and discourse analysis in environ- tal thesauri and ontologies is environmental discourse mental texts. While applied here to environmental texts, analysis. Dryzek [17] and Doyle [18] examine how lanthe general approach can be tested in other domains us- guage shapes environmental policy debates and public ing suitable ontologies and tailored extraction strategies. narratives. Nerlich and Koteyko [19] explore competing frames in climate change discourse, while recent computational studies by Jørgensen et al. [20] and Chen et
matically realized as noun phrases, we applied two different methods to extract noun phrases from the corpus. As described in the following sections, the first method is rule-based and the second is LLM-based. al. [21] apply NLP and machine learning to large-scale climate communication data.
The aforementioned studies demonstrate the benefits of combining NLP methods with structured ontologies for domain-specific entity recognition across multiple domains. However, comparative studies on these methods in the context of multilingual environmental discourse remain limited. This work builds on these foundations to advance ontology-enriched environmental text analysis.
formed rule-based noun phrase extraction relying
on annotations following the Universal
Dependen3. Methodology cies (UD) guidelines.2 Sentences were annotated
morphologically and syntactically using a
neuThis section introduces the corpus and outlines the two ral state-of-the-art dependency parser [23] using
methodological pipelines, which combine noun phrase the language models english-gum-ud-2.15 and
extraction with ontology mapping. italian-isdt-ud-2.15. For each sentence in
CoNLL-U format, we identified head tokens tagged
3.1. Corpus as NOUN or PROPN and expanded them by recursively
including adjectival modifiers ( amod), compounds
The corpus used in this study is the English and Italian (compound), and nominal modifiers ( nmod). The
subcorpus of the TreEn corpus [22], which compiles en- extraction algorithm builds each noun phrase starting
vironmental discourse from the 2014 to 2024 editions of from the head and prepending modifiers according to
the WWF Living Planet Report.1 WWF typically pub- their dependency links. The lemma column in each
lishes a suite of documents tailored to diferent audiences, CoNLL-U representation was used in order to reduce
including a full report (a comprehensive publication con- lexical variation and support downstream concept
taining detailed data, methodology, case studies, visual- mapping. For instance, from the sample sentence:
izations, and policy analysis) and an executive summary,
which distills the key findings and recommendations for (
Both the English and the Italian texts were man- amodcompoundroot ually cleaned to retain only the plain text, with all non-textual content — such as images, captions, info- Adequate funding mechanisms graphics, footnotes, and bibliographic references — sys- ADJ NOUN NOUN tematically removed to support syntactic and semantic annotation. For each English and Italian edition amod root of the WWF Living Planet Report published between compound 2014 and 2024, we computed the number of sentences, protective area management words, and lemmas using a custom Python pipeline ADJ NOUN NOUN built with pandas and language-specific spaCy models ({en,it}_core_web_sm). Table 1 presents the result- Figure 1: Dependency syntax annotation for sample noun ing counts across reporting years. phrases.
method of noun phrase extraction employed GPT-o3.3
able to retrieve the Italian editions of the report, starting from the earliest available publication up to the most recent: https://www. 2https://universaldependencies.org/guidelines.html wwf.it/cosa-facciamo/pubblicazioni/living-planet-report/ 3https://platform.openai.com/docs/models/o3 sentences tokens unique words lemmas avg. sentence length 2014 2016 2018 2020 2022 2024
Specific prompts were iteratively developed for each lan- Translate,4 and then applied basic normalization (e.g.,
guage under analysis (i.e., Italian and English), with in- lowercasing, removal of diacritics) before comparing it to
structions highlighting syntactic constraints, lemmati- the set of Italian noun phrases extracted from the aligned
zation, and complete modifier preservation in order to sentence using the same syntactic rules. If a match was
ensure consistency with the rule-based noun phrase ex- found, the corresponding GEMET concept was propagated
traction method. Figure 4 (see Appendix A) presents the to the Italian sentence. For example, in the English
senEnglish prompt used for this task. tence:
(
the noun phrases science and world were mapped to Following the extraction of candidate noun phrases, we GEMET concepts. Their Italian equivalents, scienza and performed concept-level mapping using two distinct on- mondo, appeared among the extracted noun phrases in tologies: GEMET and BabelNet. The objective was to the aligned Italian sentence “In tutto il mondo si usano link each phrase to an unique identifier representing an molte lingue per comunicare la scienza”. In this way, we environmentally relevant concept within a structured could propagate the annotations to the Italian side, even semantic resource. though the GEMET concept is originally linked to the En
Our multilingual setting required diferent strategies glish noun phrase. for the two resources. For GEMET, which is primarily In contrast, BabelNet provides multilingual support designed around English entries and ofers more lim- by design. Therefore, we queried noun phrases directly in ited multilingual coverage, we relied on aligned sentence both English and Italian, allowing us to retrieve languagepairs in English and Italian to propagate annotations. specific senses without relying on sentence alignment. Specifically, we used an alignment file where each English This approach enabled broader coverage and avoided the sentence was paired with its Italian equivalent. Once need for cross-lingual projection.
GEMET concepts were identified in the English sentence, we transferred them to the Italian version whenever the GEMET. We queried GEMET via its public REST API.5 same noun phrase (or a direct translation) was present. For each noun phrase, we attempted an exact string This allowed us to enrich the Italian portion of the corpus match using the getConceptsMatchingKeyword endeven when direct GEMET matches were not available in point. To maximize recall, we also applied fallback strateItalian. To support this transfer, we first checked whether gies by decomposing multiword expressions and querythe same noun phrase annotated in English occurred ver- ing each component token separately (e.g., climate vulbatim in the aligned Italian sentence. If no exact match nerability → climate, vulnerability). Concept URIs (Uniwas found, we used automatic translation to bridge the form Resource Identifier) returned from GEMET were gap between the two languages. Specifically, we translated the English noun phrase into Italian using Google 4https://cloud.google.com/translate/docs/reference/rest 5https://www.eionet.europa.eu/gemet/en/webservices/ stored along with the original phrase to support later semantic grouping and analysis. In addition, we retrieved the semantic group associated with each concept using the getAllConceptRelatives endpoint (with relation group), allowing us to categorize entities into high-level thematic domains (e.g., BIOSPHERE, SOCIETY, WASTES).
Figure 2 shows the GEMET entry for climate change6and all the fields we extract: concept URI, label, definition, related terms, and group. These fields were stored to facilitate both downstream semantic analysis and explainability of the mappings.
geological epoch, Anthropocene. Figure 3 shows the BabelNet entry for Anthropocene, highlighting the fields we extract, namely, definition, categories, relations, synonyms, and semantically related terms.
BabelNet. We accessed BabelNet via its getSenses and getSynsetIds endpoints,7 querying each noun phrase both in Italian and in English. This bilingual querying strategy was adopted to maximize coverage and mitigate cases where a concept might be present only in one of the two languages. Unlike GEMET, BabelNet returns disambiguated senses associated with synset identifiers. We retained only senses with part-of-speech NOUN 3.3.2. LLM-based Ontology Mapping and applied a filtering step to discard irrelevant or ambiguous senses based on glosses and semantic domains. Our second method for performing the concept-level In multiword expressions that failed to return a direct mapping of the extracted noun phrases relies on GPT-o3 match, we again decomposed the phrase into component through prompts with expected output. Upon extraction, tokens and aggregated partial matches when available. a manual analysis of the concept-level mapping process As an example, consider the following sentence: (cf. Table 2) revealed that several (multi-word) noun phrases were not found in either GEMET or BabelNet.
For example, in the following sentence:
noun phrases, highlighting the wider lexical and multilingual coverage of BabelNet. This example also demonstrates the complementary nature of the two resources: GEMET provides high precision within the environmental domain, while BabelNet ensures broader recall across a wider conceptual space.
This dual mapping strategy enabled both
domainspecific grounding (via GEMET) and broader lexical
disambiguation (via BabelNet), intensifying the robustness of
concept alignment across heterogeneous texts.
(
(
We need nature positive by 2030 – which, in simple terms, means more nature by the end of this decade than at its start.
To conduct a diachronic analysis of concepts mapped
with GEMET and BabelNet across the 2014–2024 corpus,
we identified concepts that appear in all WWF report
editions and analysed their frequency per report to gather
insights into the evolution of environmental discourse.
(
Across the 2014–2024 WWF reports, the LLM-based ing in the health, abundance, diversity and resilience of method extracts a number of unique noun phrases com- species, ecosystems and processes”. GPT-o3, on the other parable to the rule-based method. However, for the hand, generates the following definition: “a future state 2022 Italian edition, the LLM extracts substantially more in which nature—biodiversity, ecosystem services and natuphrases. This diference appears to result from the way ral capital—is restored and enhanced relative to its current nested structures were handled: the LLM returned entire condition”. While both definitions are valid, the LLMnoun phrases with several nested noun phrases. This pat- generated one captures more accurately the forwardtern is especially evident in Italian, where nested noun looking, goal-oriented nature of the nature positive conand prepositional phrases are common. For instance, cept. Unlike BabelNet’s definition, which frames the extracted spans such as “negoziato internazionale della concept mainly as a set of measurable biodiversity outconvenzione quadro delle Nazioni Unite sul cambiamento comes, the LLM definition presents it as a “future state” climatico e della convenzione sulla diversità biologica” 8 or in which nature is restored and enhanced. This distinction is significant, as BabelNet treats the concept as a 8Original in English: “international negotiations under the United Nations Framework Convention on Climate Change and the Convention on Biological Diversity”. 10https://wwf.panda.org/nature_positive/ 11Acronym for Sustainable Development Goals. 5. Discussion result, while the LLM version treats it as a trajectory/vision, which aligns more closely with how the term is currently used in WWF discourse (e.g., WWF defines nature positive as a goal to “halt and reverse nature loss by 2030” ).10 This suggests a promising direction for scaling these semantic resources, with domain-relevant entities extracted from domain-specific literature. However, as highlighted above, given the nuanced conceptual distinctions, expert validation remains crucial in order to ensure accuracy and to account for the subtle semantic distinctions that such models may overlook.
tions of rule-based and LLM-based pipelines for ontologyoriented entity annotation in environmental discourse, aligning with insights from previous work on domainspecific NLP [3, 4, 5, 6].
First, in terms of extraction, the LLM-based approach demonstrated coverage comparable to the rule-based method in line with recent research highlighting LLMs’ strong performance for entity detection [7, 8]. However, the LLM’s tendency to generate longer, contextually rich Temporal analysis of concept dynamics. Our di- noun phrases — particularly in Italian, where nesting is achronic analysis of the concepts mapped via GEMET and frequent — resulted in both higher phrase counts and BabelNet across the six-year corpus (2014–2024) yielded a greater proportion of partial matches. This confirms the following results. observations by Marrero et al. [4] that domain-relevant
For GEMET, we identified 59 English concepts that concepts often appear as complex, nested noun phrases appeared consistently in all years, including domain- that challenge standard NER boundaries. specific terms such as climate change, biodiversity, ecosys- Second, our results show that while GEMET provides tem, and habitat loss, reflecting the controlled and envi- reliable coverage for core environmental concepts, conronmentally focused nature of the thesaurus. A parallel sistent with its controlled and domain-focused design, analysis of the Italian portion revealed a partially over- BabelNet ofers a wider conceptual coverage. This lapping core set, with terms such as ambiente (environ- aligns with prior findings that general-purpose lexical ment), specie (species), and risorsa (resource) persistently networks like BabelNet can capture more entities, but appearing. at the same time can include discourse or general entities
In contrast, BabelNet yielded a smaller set of consis- not so relevant to characterize a domain [13, 15]. tently recurring concepts, such as biodiversity, consump- Third, the quality of LLM-generated definitions for untion, and development, but also revealed a much broader mapped phrases suggests potential for semi-automated and more dynamic tail of emerging concepts (i.e., less fre- ontology enrichment. For instance, for the concept naquent terms that vary widely across documents and cap- ture positive, the LLM produced a forward-looking definiture context-dependent discourse). Notably, BabelNet tion more aligned with current environmental discourse annotations surfaced many general-purpose or discourse- framing than the existing BabelNet entry. This supports driven terms (e.g., ambition, alarm, goal, confidence limit), recent arguments for integrating LLMs into domain onoften reflecting the rhetorical framing of environmental tology extension workflows [ 9], but also highlights the narratives in the source texts. importance of expert validation, given possible subtleties
We also tracked emerging and declining concepts in sense distinctions. across both resources. For GEMET, emergent concepts Finally, our diachronic analysis, though conducted on since 2018 include soil biodiversity, plastic, ocean acidifi- a very low scale, showed interesting aspects about how cation, and urbanisation, many of which correspond to sustainability narratives evolve rhetorically, in line with increasingly salient ecological issues. Conversely, con- work by Dryzek [17] and Nerlich and Koteyko [19] on cepts such as ammonia, energy consumption, and ozone shifting environmental frames. peaked before 2018 and gradually disappeared, suggest- Taken together, our results demonstrate that combining shifting topical focus in environmental discourse. ing rule-based and LLM-based pipelines may provide Similar trends were found in BabelNet, where contem- complementary strengths for environmental concept anporary discourse introduced terms like sdg,11 carbon se- notation: the rule-based method ensures syntactic preciquestration, and digital storytelling, while older narrative sion and consistent granularity, while the LLM broadens anchors like anthropocene, habitat loss, and even ocean semantic reach and can supply draft definitions for novel saw relative decline. A detailed overview of the five most or evolving terms. However, consistent ontology covfrequent concepts per year, derived from both GEMET and erage remains an issue, as a substantial proportion of BabelNet annotations, is provided in Tables 4 and 5 (see relevant phrases were not found in either resource, unAppendix A). derscoring the need for ongoing ontology expansion and domain adaptation, as stressed in recent surveys [10, 24].
Future work should explore refining LLM prompts to better constrain phrase boundaries, integrating syntactic cues during generation, and developing semiautomatic curation workflows to incorporate validated LLM-generated definitions into existing ontologies. This is a promising path for scaling high-quality, domainadapted semantic annotation in support of environmental discourse analysis.
ing and semantically annotating noun phrases in multilingual environmental texts using both GEMET and BabelNet ontological frameworks. The two resources were used in complementary ways: GEMET provided structured domain-specific knowledge, while BabelNet contributed broader lexical coverage and multilingual lfexibility. Through a combination of ontology matching, fallback decomposition strategies, and cross-lingual projection, we achieved wide and meaningful semantic enrichment across languages. Looking ahead, the approach we propose could also support the ongoing evolution of domain ontologies themselves. For instance, GEMET is periodically updated with new concepts and definitions. 12 Automatically extracting candidate terms and associating them with existing or missing concepts, especially through LLM-based suggestion and contextual generalization, might provide curators looking to add to the thesaurus with insightful information.
Several directions can be pursued for the future development of this work. For instance, alternative approaches to named entity propagation — such as alignment-based techniques [25, 26] — can be tested, and additional inventories for entities and concepts can be explored, such as [27].
Finally, it is important to note that our study focused on the task as performed by LLMs. In future work, we will compare these results with human annotations provided by domain experts in order to examine whether more or diferent entities are extracted from the texts. This comparison will help determine whether more finegrained analyses are necessary (e.g., to resolve partial matches involving nested entities or syntactically complex modifier structures). Moreover, incorporating expert judgment will allow us to account for diverse disciplinary perspectives (e.g., biology, ecology, chemistry, physics, geography) on environmental issues.
the TreEn project team. Special thanks to the collaborators who contributed to the treebank annotation. 12https://www.eionet.europa.eu/gemet/en/about/
tional project CN-HPC-Spoke1-Future HPC & Big Data, PNRR MUR-M4C2. Adriana Pagano has a grant from Brazil’s National Council for Scientific and Technological Development (CNPq 404722/2024-5; 313103/2021-6) and Minas Gerais State Agency for Research and Development (FAPEMIG) to conduct research in collaboration with the University of Turin. ical entity recognition: A systematic review of ap- [24] S. Li, Z. Zhou, L. Huang, F. Wu, A survey on proaches and challenges, Briefings in Bioinformat- ontology-based named entity recognition, IEEE ics 22 (2021) bbaa180. Access 10 (2022) 113192–113210. [11] S. P. Villacorta Chambi, M. Lindsay, J. Klump, [25] S. Tedeschi, R. Navigli, MultiNERD: A multiK. Gessner, E. Gray, H. McFarlane, Assessing lingual, multi-genre and fine-grained dataset for named entity recognition by using geoscience do- named entity recognition (and disambiguation), in: main schemas: the case of mineral systems, Fron- M. Carpuat, M.-C. de Marnefe, I. V. Meza Ruiz tiers in Earth Science 13 (2025) 1530004. (Eds.), Findings of the Association for Computa[12] X. Dai, S. Karimi, C. Paris, Building domain-specific tional Linguistics: NAACL 2022, Association for knowledge graphs for named entity linking: A case Computational Linguistics, Seattle, United States, study of cancer research literature, in: Proceedings 2022, pp. 801–812. URL: https://aclanthology.org/ of the 58th Annual Meeting of the Association for 2022.findings-naacl.60/. doi: 10.18653/v1/2022.
Computational Linguistics, 2020, pp. 7460–7471. findings-naacl.60. [13] R. Navigli, S. P. Ponzetto, Babelnet: The automatic [26] F. Martelli, A. S. Bejgu, C. Campagnano, J. Čibej, construction, evaluation and application of a wide- R. Costa, A. Gantar, J. Kallas, S. P. Koeva, K. Koppel, coverage multilingual semantic network, Artificial S. Krek, M. Langemets, V. Lipp, S. Nimb, S. Olsen, Intelligence 193 (2012) 217–250. doi:10.1016/j. B. Sanford Pedersen, V. Quochi, A. Salgado, L. Siartint.2012.07.001. mon, C. Tiberius, R.-J. Ureña-Ruiz, R. Navigli, XL[14] European Environment Agency, Gemet — gen- WA: a gold evaluation benchmark for word aligneral multilingual environmental thesaurus, https: ment in 14 language pairs, in: F. Boschetti, G. E. //www.eionet.europa.eu/gemet/en/themes/, 2025. Lebani, B. Magnini, N. Novielli (Eds.), Proceedings Accessed: 2025-06-15. of the 9th Italian Conference on Computational [15] J. Ryu, J. Lee, J. Kang, Cross-lingual entity link- Linguistics (CLiC-it 2023), CEUR Workshop Proing with multilingual bert and knowledge graph ceedings, Venice, Italy, 2023, pp. 272–280. URL: embedding, volume 546, 2021, pp. 663–674. https://aclanthology.org/2023.clicit-1.34/. [16] Y. Zhao, W. Chen, X. Xie, Z. Liu, J. Li, Entity-aware [27] G. Martinelli, F. Molfese, S. Tedeschi, A. Fernándezneural machine translation, in: Proceedings of the Castro, R. Navigli, CNER: Concept and named en58th Annual Meeting of the Association for Com- tity recognition, in: K. Duh, H. Gomez, S. Bethard putational Linguistics, 2020, pp. 2858–2866. (Eds.), Proceedings of the 2024 Conference of the [17] J. S. Dryzek, The Politics of the Earth: Environmen- North American Chapter of the Association for tal Discourses, Oxford University Press, 2013. Computational Linguistics: Human Language Tech[18] J. Doyle, Media and the Environment, Polity Press, nologies (Volume 1: Long Papers), Association for 2020. Computational Linguistics, Mexico City, Mexico, [19] B. Nerlich, N. Koteyko, Competing representations 2024, pp. 8336–8351.
of the ’climate change’ frame in uk news media,
Nature Climate Change 3 (2009) 423–427. [20] P. S. Jørgensen, colleagues, Machine learning and natural language processing in environmental research, Environmental Research Letters 17 (2022) 023003. [21] Z. Chen, T. Zhang, H. Su, Analyzing climate change discourse with nlp: A review, Current Opinion in
Environmental Sustainability 61 (2023) 101237. [22] A. Pagano, P. Chiril, E. Chierchiello, C. Bosco,
Treen: A multilingual treebank project on environmental discourse, in: Eighth Workshop on Universal Dependencies (UDW, SyntaxFest 2025), 2025, p. 80. [23] M. Straka, UDPipe 2.0 prototype at CoNLL 2018
UD shared task, in: Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, Association for Computational Linguistics, Brussels, Belgium, 2018, pp. 197–207. URL: https://www.aclweb.org/anthology/
K18-2020. doi:10.18653/v1/K18-2020.
WWF report
ecosystem (20), world (20), species (18), energy (18), resource (16)
specie (18), ecosistema (14), energia (13), mondo (13), biodiversità (12)
ecosystem (31), species (27), resource (22), food (22), energy (18)
ecosistema (26), specie (25), risorsa (23), habitat (15), consumo (15)
biodiversity (56), species (35), loss (26), indicator (15), land (14)
biodiversità (39), specie (26), perdita (20), indicatore (
Top five BabelNet Concepts (with frequency)
earth (11), country (
Declaration on Generative AI During the preparation of this work, the author(s) did not use any generative AI tools or services.