ABOUT Abstraction

Words, the building blocks of language, are signs that represent different kinds of categories. Some words define categories of concrete entities (such as cat or table) while others define abstract entities (such as belief, empathy). Some words define general categories that include within them different entities (such as the word vehicle, or art) while others define more specific categories (such as tandem, or Impressionism).  

Humans construct meaning by transforming experience into mental categories. This process of abstraction is closely connected to language, which provides the labels we use to organize and communicate knowledge. Two fundamental properties of words contributing to these processes are concreteness, the degree of connection to sensory experience, and specificity, the degree of categorical inclusiveness of a concept. Understanding their roles helps us understand how language mediates human abstraction.

Researchers from different disciplinary backgrounds tend to focus on different aspects of abstraction when investigating its mechanisms and effects. In particular, psychologists and cognitive scientists tend to focus on concreteness, while linguists and computer scientists often consider specificity. These partial perspectives can lead to misunderstandings in interdisciplinary debates and hinder theoretical development.

A multidimensional perspective on abstraction

ABSTRACTION has helped distinguish concreteness from specificity by investigating their relationship and their different contributions to the representation and processing of meaning. This distinction allows us to recognize that an abstract concept can be highly specific and that a concrete concept can be general.

The project’s findings have informed theoretical discussions about the role of specificity in models of language comprehension grounded in bodily and sensory experience. The project has also contributed to interdisciplinary work aimed at clarifying the terminology used in the study of meaning, facilitating dialogue across disciplines.

Read more:

• Beyond concreteness: why word specificity is the missing piece in theories on embodied language comprehension — Bolognesi (2025)

• What we mean when we say semantic: Toward a multidisciplinary semantic glossary — Reilly and colleagues

• Specificity: Metrics and Norms — Villani and Bolognesi (2026)

• Conceptual Metaphor — Bolognesi (2026)

New data and tools for studying meaning

At the start of the project, extensive collections of word concreteness judgments were already available, whereas resources for lexical specificity based on human judgments were much more limited.

To help address this gap, the team collected word specificity data in Italian and English through rating tasks and innovative gamification techniques. In particular, the Word Ladders application allows participants to construct word ladders by connecting concepts through relations of categorical inclusion.

The project has documented how this tool works and investigated the advantages of gamification for collecting linguistic data. It has also contributed to resources covering other lexical properties, such as iconicity: the perceived resemblance between a word’s form and its meaning.

Read more:

• Word Ladders: A Mobile Application for Semantic Data Collection — Bolognesi and colleagues (preprint)

• The advantages of gamification for collecting linguistic data: A case study using Word Ladders — Genovese and colleagues (2024)

• IconicITA: Iconicity ratings of the Italian affective lexicon — de Varda and colleagues (2025)

Three research areas: thought, language, and creativity

Drawing on these data and other lexical resources, ABSTRACTION has investigated the roles of specificity and concreteness through a combination of quantitative and qualitative methods across three areas:

• Thought: how humans organize experience into categories, represent concepts, and process their meanings.

• Language: how word properties contribute to textual clarity and informativeness, and to communication dynamics.

• Creativity: how familiar concepts are combined and reinterpreted through metaphors, novel expressions, and other forms of linguistic variation.

The pages dedicated to these three areas present the main findings and links to the related publications.

Human generalization and artificial intelligence

Understanding how people move from particular experiences to more general categories and concepts is also central to comparisons with artificial intelligence.

ABSTRACTION has contributed to interdisciplinary discussions of the similarities and differences between human and artificial generalization. The paper published in Nature Machine Intelligence compares how cognitive science and AI research define, implement, and evaluate generalization, identifying key challenges for collaboration between humans and artificial systems.

Other project studies have examined how language models organize categories, interpret generalizations, and represent concreteness and inclusiveness in context. Comparisons with human data have shown that producing plausible responses does not guarantee full correspondence with speakers’ conceptual organization. For example, the study of Italian subordinate categories found limited alignment between humans and models, with variation across semantic domains.

Read more:

• Aligning generalization between humans and machines — Ilievski and colleagues (2025), Nature Machine Intelligence

• How Humans and LLMs Organize Conceptual Knowledge: Exploring Subordinate Categories in Italian — Pedrotti and colleagues (2025)

• Wordnet and Word Ladders: Climbing the abstraction taxonomy with LLMs — Puccetti, Esuli, and Bolognesi (2025)

• Quantifying Generalizations: Exploring the Divide Between Human and LLMs’ Sensitivity to Quantification — Collacciani, Rambelli, and Bolognesi (2024)

• Interpretation of Generalization in Masked Language Models: An Investigation Straddling Quantifiers and Generics — Collacciani and Rambelli (2023)

• ABRICOT – ABstRactness and Inclusiveness in COntexT: A CALAMITA Challenge — Puccetti and colleagues (2025)

• Yet another approximation of human semantic judgments using LLMs… but with quantized local models on novel data — Ravelli and Bolognesi (2024)

Abstraction, social perspectives, and stereotypes

Comparisons between human and artificial language also concern the social perspectives expressed in texts. Research associated with the project examined how levels of abstraction change when language models are prompted to adopt different sociodemographic identities.

The findings highlight the limitations of this strategy in representing social groups’ perspectives and indicate a risk of reproducing stereotypes even when a model appears to speak from the perspective of a marginalized group.

Read more:

• Simulating Identity, Propagating Bias: Abstraction and Stereotypes in LLM-Generated Text — Sommerauer, Rambelli, and Caselli (2025)

Overall, ABSTRACTION contributes to understanding how language enables us to construct meaning from experience and formulate generalizations. Distinguishing concreteness from specificity provides tools for studying cognition, designing communication, and evaluating the capabilities and limitations of artificial intelligence systems.