The Singleton Fallacy: Why Current Critiques of Language Models Miss the Point.

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Cognitive Science

Background:

• Critiques of neural network-based Natural Language Understanding (NLU) models, often termed language models, are prevalent.
• The debate frequently centers on the 'singleton fallacy,' an assumption that language, meaning, and understanding represent a singular, uniform phenomenon.
• This assumed phenomenon is often considered unattainable by current language models.

Purpose of the Study:

• To address the 'singleton fallacy' in the critique of language models.
• To argue that language models offer a viable computational approach to language understanding.
• To reframe language understanding as a computational means to an end, rather than an absolute state.

Main Methods:

• Conceptual analysis of the 'singleton fallacy' in NLU debates.
• Theoretical argumentation regarding dualistic positions versus computational approaches.
• Historical context of distributional methods in computational linguistics.

Main Results:

• The 'singleton fallacy' is identified as a key error in current critiques of language models.
• Positing a mental 'unobtanium' for understanding leads to dualism, a position distributional methods were developed to avoid.
• Language models represent a theoretically and practically sound method for achieving computational language understanding.

Conclusions:

• Language models provide the most promising current approach for computers to achieve language understanding.
• The critique against language models is often based on a flawed assumption about the nature of understanding.
• Understanding in AI should be viewed as a computational capability, a means to an end.