- The paper demonstrates that current AI models exhibit strong observational adequacy but struggle with descriptive and explanatory linguistic clarity.
- It compares traditional neural language models with emerging neural grammar induction models that explicitly derive syntactic rules.
- The study suggests integrating neural grammar induction techniques to refine language learnability and advance the understanding of Universal Grammar.
Revisiting Chomskyan Theoretical Frameworks in Light of AI Linguistic Models
The paper "On the goals of linguistic theory: Revisiting Chomskyan theories in the era of AI" by Eva Portelance and Masoud Jasbi revisits foundational principles of generative linguistics within the context of advancements in AI LLMs. By examining these principles, the paper aims to evaluate whether AI models can effectively contribute to the goals of theoretical linguistics, a field long driven by the tenets laid out by Noam Chomsky. It specifically focuses on three critical areas: levels of theoretical adequacy, procedures for linguistic theory development, and language learnability and Universal Grammar.
Generative linguistics aims to describe human linguistic capability through formal grammars and explore how such grammars are acquired and processed. Chomsky's framework postulates several levels of adequacy for theoretical models, namely observational, descriptive, and explanatory adequacy. Observational adequacy pertains to generating grammatical sentences from a corpus, descriptive adequacy extends to explaining structural intuitions among native speakers, and explanatory adequacy addresses why a particular grammar is acquired over others. The paper draws parallels between these levels and Marr's computational, algorithmic, and implementational levels of analysis.
The authors argue that neural LLMs such as BERT and GPT-2 exhibit significant observational adequacy. However, these models often lack the capacity to provide a clear, human-readable grammatical structure, thus lagging behind in both descriptive and explanatory adequacy. Despite their robust performance in language production, they are frequently seen as "black boxes," where the underlying linguistic representations remain opaque.
In contrast, neural grammar induction models are presented as a promising alternative. These models do not merely replicate surface linguistic patterns but aim to infer explicit sets of syntactic rules akin to the grammars outlined in generative linguistics. Contemporary models, such as C-PCFG and L-PCFG, have demonstrated capabilities in deriving formal grammars with minimal prior linguistic assumptions. This positions them favorably as they can aspire to both descriptive and explanatory adequacy, making them valuable assets for revisiting the procedures for linguistic theory development.
Portelance and Jasbi also revisit the question of language learnability and Universal Grammar (UG) in light of AI models. Chomsky's UG hypothesis suggests that an innate linguistic structure constrains the space of possible grammars, a direct response to the learnability challenge. The paper posits that AI models, particularly those that implement probabilistic and neural network approaches, could provide insights into alternative frameworks, potentially refuting or refining long-standing theories like UG. These models present a learning paradigm that diverges from the traditional symbolic representation of grammatical knowledge, offering a more distributed and less biased approach to modeling language acquisition.
Despite the comprehensive evaluation of AI models in theoretical linguistics, the paper emphasizes the need for cautious integration. Neural network models, while robust, must be judged against criteria for model adequacy pertinent to linguistic theories before displacing or redefining existing frameworks. The authors argue for a balanced exploration of both neural LLMs and neural grammar induction models, particularly where they complement the theoretical goals of generative linguistics.
In conclusion, Portelance and Jasbi provide a nuanced examination of the synergies and divergences between AI models and traditional linguistic theories. They advocate for the potential of neural grammar induction models in evaluating and possibly reshaping existing linguistic tenets, particularly when extending research into multimodal and multichannel inputs. Future investigations might capitalize on these AI models to uncover deeper insights into language processing and acquisition, ultimately driving towards the overarching goal of comprehending the essence and structure of human language.
