Ambiguity and underspecification change how language models learn words
Compact Language, Complex Model Shifts: How and Where Ambiguity and Underspecification Affect LLMs
Computation and Language
Summary
Language models can get confused by words that have multiple meanings or are unclear. The authors made fake words with multiple meanings and vague meanings to see how models learn them when trained more and more. They found that models get better overall with more ambiguous words but make more mistakes generating sentences with those tricky words. Inside the models, meanings of words with multiple interpretations get separated, but vague word meanings stay unclear as the model writes sentences.
What this means in practice
- •For natural language processing engineers: Improve training of language models by understanding the effects of ambiguous and underspecified words on model performance and internal representations.
- •For ai system maintainers: Better detect when AI-generated text may contain errors related to ambiguous or underspecified words and improve reliability.
Authors
Michaela Regneri, Nina Scheller, Sören Laue
Abstract
We analyze how lexical ambiguity and underspecification affect language model training. We create artificial homonyms and artificial hypernyms as pseudowords and analyze the generative performance of language models as they are trained with increasing amounts of these ambiguous or underspecified pseudoword types. We further analyze whether the models disambiguate ambiguous or underspecified statements and provide a first mechanistic account of how ambiguity and disambiguation are represented internally. Our main results show that both ambiguity and underspecification increase model performance in ways that scale with their influence on the language's type-token ratio. However, the accuracy of generating sequences containing ambiguous words or their synonyms decreases compared to other texts. We also show that internal representations of pseudowords reflect disambiguation of pseudo-homonyms, but underspecification of pseudo-hypernyms is maintained during the generative process.