Usage-based constructionist approaches and large language models
The constructionist framework is more relevant than ever, due to efforts by a broad range of researchers across
the globe, a steady increase in the use of corpus and experimental methods among linguists, consistent findings from laboratory
phonology, neuroscience, sociolinguistics, and striking progress in transformer-based large language models. These advances
promise exciting developments and a great deal more clarity over the next decade. The constructionist approach rests on two
interrelated but distinguishable tenets: a recognition that constructions pair form with function at varying levels of specificity
and abstraction, and the recognition that our knowledge and use of language are dynamic and based on language use.
Article outline
- 1.Introduction
- 1.1Constructions all the way down
- 1.2Definition
- 2.The usage-based nature of language
- 2.1Hazard a guess
- 2.2Skewed input
- 3.The usage-based nature of language is a challenge for symbolic formalisms
- 3.1Symbolic, feature-based formalisms
- 3.2Combining constructions: An example
- 4.A game changer: Large Language Models
- 4.1Lossy compression and interpolation
- 4.2Conform to conventions
- 4.3Complex dynamic network of constructions at varying levels of abstraction and complexity
- 4.4Context-dependent interpretation
- 4.5Semantic relationships among discontinuous elements
- 4.6NEW: Goal is to be helpful
- 4.7With great power comes great responsibility
- 5.GPTs at work
- 5.1Intention-reading and social inferences
- 5.2GPT correctly interprets unusual examples
- 5.3GPT4 on a simple math problem
- 5.4GPT4 appropriately characterizes conceptual metaphors
- 5.5Over-reliance on associations can lead GPT models (and humans) astray
- 6.Looking ahead
- 7.Conclusion
- Acknowledgements
- Notes
-
References