Large Language Models: Language abilities and limitations

Identifier:  TDX:4499

Handle: https://hdl.handle.net/20.500.11797/TDX4499

Authors:  Dentella, Vittoria

Abstract:
Large Language Models (LLMs) are recruited in applications that generate strings of text that closely resemble natural language. On this basis, LLMs have been invested with human-like capabilities such as understanding, reasoning, and intelligence. Additionally, their adoption as cognitive theories capable of explaining language processing in humans is increasingly called for. However, a closer look reveals that the language outputs of LLMs are replete with inconsistencies and content no neurotypical human would produce, casting doubt on the human-likeness of the linguistic knowledge they encode. In this thesis, (1) an outline is drawn of the knowledge over morphosyntax, semantics and pragmatics that LLMs are attributed with; (2) experimental evidence is provided that similarities between LLMs’ outputs and natural language hold only superficially and thus, the two are not qualitatively comparable; and (3) a discussion is offered on the choice of methods used to evaluate the linguistic abilities of LLMs.