Tokenization Falling Short: On Subword Robustness in Large Language Models

TL;DR

The findings reveal that scaling model parameters can mitigate the issue of tokenization; however, LLMs still suffer from biases induced by typos and other text format variations and experiments show that subword regularization such as BPE-dropout can mitigate this issue.

Abstract

Language models typically tokenize raw text into sequences of subword identifiers from a predefined vocabulary, a process inherently sensitive to typographical errors, length variations, and largely oblivious to the internal structure of tokens--issues we term the curse of tokenization. In this study, we delve into these drawbacks and demonstrate that large language models (LLMs) remain susceptible to these problems. This study systematically investigates these challenges and their impact on LLMs through three critical research questions: (1) complex problem solving, (2) token structure probing, and (3) resilience to typographical variation. Our findings reveal that scaling model parameters can mitigate the issue of tokenization; however, LLMs still suffer from biases induced by typos and other text format variations. Our experiments show that subword regularization such as BPE-dropout can mitigate this issue. We release our evaluation code and data at https://github.com/FloatAI/TKEval.

Figures (11)

• Figure 1: Compositional challenges in token embeddings. (a) "assignment" decomposed into "assign" and "ment" shows a cosine similarity of 0.21 and an angle of 78.16°. (b) "import" decomposed into "im" and "port" shows a cosine similarity of 0.13 and an angle of 82.47°. These results indicate that existing LLMs do not accurately capture surface form composition.
• Figure 2: $K$-shot performance on Word Unscrambling (WU) and Cycled Letters (CL) tasks. The plots illustrate that increasing the number of demonstration examples ($K$-shot) does not consistently enhance performance. However, models with larger parameter sizes generally exhibit better performance across both tasks.
• Figure 3: The relationship between the length of scrambled words and the Exact Match (EM) score of Llama3-8B and Llama3-70B on the word unscrambling task under one-shot evaluation. The models tend to correctly reorder anagrams of shorter lengths, while struggling with longer words.
• Figure 4: $K$-shot performance on intra-token probing tasks (CCV, CC, NC, NCR). The plots demonstrate that increasing the number of demonstration examples ($K$-shot) generally results in an improvement from zero-shot to one-shot, with performance stabilizing thereafter.
• Figure 5: $K$-shot performance on various inter-token probing tasks. For edit distances, lower is better.