Word Specificity: A Measure to Investigate Semantic Abstraction and Linguistic Creativity

Abstract

ABSTRACTION is a fundamental process in human cognition, which facilitates the formation of general concepts by extracting similarities and patterns from direct experiences, linguistic expressions, or existing concepts. It is commonly understood as Conceptual Concreteness, i.e., the capacity to think and communicate in terms of conceptual categories detached from perceptual experiences. However, abstraction includes the ability to think and talk in terms of conceptual categories that vary in inclusiveness, or better Categorical Specificity. For instance, "cat" represents a less inclusive concept with specific semantic and perceptual features, whereas "mammal" is not directly characterized by perceptual features as it encompasses various subcategories. This talk presents the role that Specificity could play in linguistic research, by looking at three aspects: i) how Specificity affects conceptual representation (through the lens of linguistic distributions), ii) how linguistic productivity relies on analogical mechanisms (by varying Specificity of words in constructions), and iii) what role word Specificity plays in generating creative language. The talk will present two ongoing studies. The first study investigates how the interplay between Concreteness and Specificity influences the distribution of concept words, specifically examining "contextual variability." By relying on distributional methodologies, this study explores how the degree of Specificity affects the contextual usage of nouns in both English and Italian languages. The hypothesis posits that more Specific words, irrespective of their concreteness, tend to be associated with narrower domains, thereby exhibiting reduced contextual variability compared to generic words. The second study investigates how Specificity influences linguistic productivity, focusing on the interpretation of novel compounds. The idea is that people can interpret novel compounds by abstracting from their knowledge of past experiences with similar conceptual combinations, which provide an analogical basis for the interpretation of novel constructs. By manipulating lexicalized (“birthday cake”) compounds and observing the challenges faced by Language Models (LLMs) in interpreting generated compounds (“birthday dessert”), this study sheds light on how altering Specificity impacts the interpretative processes of LLMs. Concluding remarks will contemplate potential cognitive and computational avenues for exploring the relation between abstraction processes and linguistic creativity.

Short Bio

Dr. Giulia Rambelli is currently a Post-doc for the ABSTRACTION project (a 5-years project founded by the European Research Council, supervised by prof. Marianna Bolognesi) at the University of Bologna. She obtained a joint PhD degree in Linguistics from the University of Pisa (Italy) and Aix-Marseille University (France) under the joint supervision of prof. Alessandro Lenci and prof. Philippe Blache. Her research is at the intersection between theoretical linguistics, cognitive science, and machine learning, with a specific interest in the underlying mechanisms of sentence interpretation and productivity from a usage-based perspective. Her papers have won international awards, including the Baidu AACL-IJCNLP Best Paper Award 2020, the Best Paper Award at *SEM 2021, and the Outstanding Paper Award at MWE 2023. Starting this year, she is also one of the organizers of the yearly workshop on Cognitive Modeling and Computational Linguistics (CMCL).

Venue

Large Conference Room, O’Reilly Institute