言語生成AIは文の意味を理解しているか：真理条件的意味論の立場から

Abstract

This study investigates whether language generative AIs understand the meaning of a sentence using Google Bard (currently Gemini) and ChatGPT. Based on the truth-conditional semantics, we define the meaning of a sentence as the condition under which the proposition expressed by the sentence is true. We use two sentences, “The morning star is the morning star.” and “If you don't give me candy, I'll play a trick on you.” to assess whether the AIs understood the meaning of the sentences. The former sentence is a well-known example of a noun phrase having a referent “Sinn,” distinguished from “Bedeutung” by Frege. The latter sentence is a conditional sentence that induces what is called an “invited inference,” which is different from logical inference. The results show that both AI models gave incorrect having a referent, “Sinn,” distinguished from “Bebentung” by Frege truth conditions for the former sentence. For the latter sentence, both AIs gave inconsistent answers mixing correct and incorrect answers depending on the wording of the input.

Ähnliche Arbeiten