Publications

Abstract

Tokenization significantly influences language models (LMs)’ performance. This paper traces the evolution of tokenizers from word-level to subword-level, analyzing how they balance tokens and types to enhance model adaptability while controlling complexity. Despite subword tokenizers like Byte Pair Encoding (BPE) overcoming many word tokenizer limitations, they encounter difficulties in handling non-Latin languages and depend heavily on extensive training data and computational resources to grasp the nuances of multiword expressions (MWEs). This article argues that tokenizers, more than mere technical tools, should drawing inspiration from the cognitive science about human language processing. This study then introduces the “Principle of Least Effort” from cognitive science, that humans naturally seek to reduce cognitive effort, and discusses the benefits of this principle for tokenizer development. Based on this principle, the paper proposes that the Less-is-Better (LiB) model could be a new approach for LLM tokenizer. The LiB model can autonomously learn an integrated vocabulary consisting of subwords, words, and MWEs, which effectively reduces both the numbers of tokens and types. Comparative evaluations show that the LiB tokenizer outperforms existing word and BPE tokenizers, presenting an innovative method for tokenizer development, and hinting at the possibility of future cognitive science-based tokenizers being more efficient.

Abstract

Words typically form the basis of psycholinguistic and computational linguistic studies about sentence processing. However, recent evidence shows the basic units during reading, i.e., the items in the mental lexicon, are not always words, but could also be sub-word and supra-word units. To recognize these units, human readers require a cognitive mechanism to learn and detect them. In this paper, we assume eye fixations during reading reveal the locations of the cognitive units, and that the cognitive units are analogous with the text units discovered by unsupervised segmentation models. We predict eye fixations by model-segmented units on both English and Dutch text. The results show the model-segmented units predict eye fixations better than word units. This finding suggests that the predictive performance of model-segmented units indicates their plausibility as cognitive units. The Less-is-Better (LiB) model, which finds the units that minimize both long-term and working memory load, offers advantages both in terms of prediction score and efficiency among alternative models. Our results also suggest that modeling the least-effort principle for the management of long-term and working memory can lead to inferring cognitive units. Overall, the study supports the theory that the mental lexicon stores not only words but also smaller and larger units, suggests that fixation locations during reading depend on these units, and shows that unsupervised segmentation models can discover these units.