Periodic Reporting for period 1 - ABSTRACTION (Abstraction. Unlocking meaning from experience, through language.)
Reporting period: 2022-06-01 to 2024-11-30
Words, the building blocks of language, serve as labels that define various categories. Some words define tangible entities like cats and tables, while others define abstract concepts such as legacy and empathy. Moreover, certain words define broad, inclusive categories like vehicles and art, while others pinpoint more specific ones, like sport cars or Impressionism. These two axes of variation are called concreteness and specificity, respectively.
In the pursuit of extracting meaning from our experiences, we engage in the construction of diverse categories through the mechanisms of abstraction. Concreteness and specificity emerge as pivotal variables that underpin these abstractions.
Yet, when delving into the study of abstraction, scholars across various disciplines tend to focus solely on specificity or solely on concreteness. This singular focus, grounded in different and partial definitions of abstraction, hinders the interdisciplinary dialogue, impairs the debate, and jeopardizes theoretical development. This challenge is exacerbated by the absence of human-generated resources to measure specificity.
The innovative ABSTRACTION team aims to address this gap by collecting specificity data for thousands of words in both English and Italian, employing a gamification technique. Leveraging the lexical resources obtained through gamification, alongside other lexical resources, our approach involves conducting extensive statistical analyses, designed to unravel the interplay between specificity and concreteness in three crucial domains:
Thought: Shedding light on conflicting findings previously attributed solely to concreteness, our research seeks to explain how specificity contributes to the formation of abstract thoughts.
Language: By investigating the optimal clarity and informativeness of texts for diverse readerships, we aim to construct language that transcends traditional boundaries and resonates with a broad audience.
Creativity: Exploring the role of specificity and concreteness in constructing effective metaphors across various contexts, our research aims to unravel the creative process and enhance metaphorical expression.
The ABSTRACTION project endeavors to elucidate how the interplay of word specificity and concreteness empowers us to extract meaning from experiences. This exploration extends to the realm of cognitive science, where the grounding of abstract concepts remains an open question, and in AI research, where the construction and utilization of concepts in a manner akin to human cognition is yet to be fully understood.
In the pursuit of extracting meaning from our experiences, we engage in the construction of diverse categories through the mechanisms of abstraction. Concreteness and specificity emerge as pivotal variables that underpin these abstractions.
Yet, when delving into the study of abstraction, scholars across various disciplines tend to focus solely on specificity or solely on concreteness. This singular focus, grounded in different and partial definitions of abstraction, hinders the interdisciplinary dialogue, impairs the debate, and jeopardizes theoretical development. This challenge is exacerbated by the absence of human-generated resources to measure specificity.
The innovative ABSTRACTION team aims to address this gap by collecting specificity data for thousands of words in both English and Italian, employing a gamification technique. Leveraging the lexical resources obtained through gamification, alongside other lexical resources, our approach involves conducting extensive statistical analyses, designed to unravel the interplay between specificity and concreteness in three crucial domains:
Thought: Shedding light on conflicting findings previously attributed solely to concreteness, our research seeks to explain how specificity contributes to the formation of abstract thoughts.
Language: By investigating the optimal clarity and informativeness of texts for diverse readerships, we aim to construct language that transcends traditional boundaries and resonates with a broad audience.
Creativity: Exploring the role of specificity and concreteness in constructing effective metaphors across various contexts, our research aims to unravel the creative process and enhance metaphorical expression.
The ABSTRACTION project endeavors to elucidate how the interplay of word specificity and concreteness empowers us to extract meaning from experiences. This exploration extends to the realm of cognitive science, where the grounding of abstract concepts remains an open question, and in AI research, where the construction and utilization of concepts in a manner akin to human cognition is yet to be fully understood.
Over the past 24 months, we have reached significant achievements aligned with our main objectives. Our efforts have resulted in important publications and presentations, presented on the project website: https://site.unibo.it/abstraction/en(opens in new window).
We developed and launched the Word Ladders mobile app, a gamified tool for collecting specificity ratings, in September 2023. The app, available on Android and iOS, has over 3500 active players and has generated data from 36,000 matches. This data is being analyzed to create a knowledge graph of word connections, with findings to be published in two upcoming journal articles. The app's dissemination is ongoing through various educational channels. Additionally, traditional psycholinguistic methods have gathered specificity ratings for over 1000 Italian and 1000 English words to benchmark against app data.
In relation with our challenge described as “Abstraction in Thought”, through behavioral experiments we demonstrated that word concreteness and word specificity, while correlated, are independent variables. Specificity impacts language processing beyond concreteness. We found that words differ in semantic content based on these variables and that abstract, generic words in conversations increase the need for clarification. Moreover, we show that human conceptual categorization differs significantly from that of Large Language Models (LLMs), highlighting gaps in AI’s ability to mirror human abstraction processes.
In relation with our challenge described as “Abstraction in Language”, we showed that word specificity significantly affects the richness and density of linguistic contexts. Specific words appear in fewer, more interconnected contexts than generic ones. This underscores the role of specificity in contextual variability, beyond concreteness. We also found that the same word’s specificity varies with context and that LLMs struggle with semantic distinctions that humans handle with ease. This has implications for AI's potential to foster stereotypes through overgeneralization.
The challenge described as "Abstraction in Creativity" is planned for 2025-2027.
We developed and launched the Word Ladders mobile app, a gamified tool for collecting specificity ratings, in September 2023. The app, available on Android and iOS, has over 3500 active players and has generated data from 36,000 matches. This data is being analyzed to create a knowledge graph of word connections, with findings to be published in two upcoming journal articles. The app's dissemination is ongoing through various educational channels. Additionally, traditional psycholinguistic methods have gathered specificity ratings for over 1000 Italian and 1000 English words to benchmark against app data.
In relation with our challenge described as “Abstraction in Thought”, through behavioral experiments we demonstrated that word concreteness and word specificity, while correlated, are independent variables. Specificity impacts language processing beyond concreteness. We found that words differ in semantic content based on these variables and that abstract, generic words in conversations increase the need for clarification. Moreover, we show that human conceptual categorization differs significantly from that of Large Language Models (LLMs), highlighting gaps in AI’s ability to mirror human abstraction processes.
In relation with our challenge described as “Abstraction in Language”, we showed that word specificity significantly affects the richness and density of linguistic contexts. Specific words appear in fewer, more interconnected contexts than generic ones. This underscores the role of specificity in contextual variability, beyond concreteness. We also found that the same word’s specificity varies with context and that LLMs struggle with semantic distinctions that humans handle with ease. This has implications for AI's potential to foster stereotypes through overgeneralization.
The challenge described as "Abstraction in Creativity" is planned for 2025-2027.
Our research has revealed significant insights into the roles of specificity and concreteness in word processing, challenging previous studies that confounded these variables. This has major implications for the cognitive sciences and psycholinguistics communities, as it suggests that some effects attributed solely to concreteness may actually be explained by specificity.
Additionally, we have uncovered how specificity and concreteness interact to shape linguistic contexts, revealing that specificity plays a substantial role in contextual variability, which was previously credited mainly to concreteness. These findings advance our understanding of conceptual abstraction and language modeling.
An unplanned breakthrough in our research showed that Large Language Models (LLMs) lack the world knowledge to accurately interpret generic statements. For instance, given statements like “mosquitoes fly” and “mosquitoes carry malaria,” humans understand that the first statement refers to most mosquitoes, while the latter refers to only a small fraction of mosquitoes. LLMs lack the necessary knowledge to differentiate between these statements and tend to overgeneralize, applying a generic statement to all instances of a category. This phenomenon can lead to the encoding of harmful stereotypes. This insight is crucial for improving the ethical application of AI systems, ensuring they better reflect human understanding and reduce bias.
Additionally, we have uncovered how specificity and concreteness interact to shape linguistic contexts, revealing that specificity plays a substantial role in contextual variability, which was previously credited mainly to concreteness. These findings advance our understanding of conceptual abstraction and language modeling.
An unplanned breakthrough in our research showed that Large Language Models (LLMs) lack the world knowledge to accurately interpret generic statements. For instance, given statements like “mosquitoes fly” and “mosquitoes carry malaria,” humans understand that the first statement refers to most mosquitoes, while the latter refers to only a small fraction of mosquitoes. LLMs lack the necessary knowledge to differentiate between these statements and tend to overgeneralize, applying a generic statement to all instances of a category. This phenomenon can lead to the encoding of harmful stereotypes. This insight is crucial for improving the ethical application of AI systems, ensuring they better reflect human understanding and reduce bias.