Tokenization Falling Short: On Subword Robustness in Large Language Models

• URL: http://arxiv.org/abs/2406.11687v3
• Date: Fri, 04 Oct 2024 13:06:24 GMT
• Title: Tokenization Falling Short: On Subword Robustness in Large Language Models
• Authors: Yekun Chai, Yewei Fang, Qiwei Peng, Xuhong Li,
• Abstract summary: This study systematically investigates these challenges and their impact on language models. Our findings reveal that scaling model parameters can mitigate the issue of tokenization. Our experiments show that subword regularization such as BPE-dropout can mitigate this issue.
• Score: 12.193639356480851
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• Broken Tokens? Your Language Model can Secretly Handle Non-Canonical Tokenizations [83.93566096400723]
  We find that instruction-tuned models retain up to 93.4% of their original performance when given a randomly sampled tokenization.<n>Character-level segmentation improves string manipulation and code understanding tasks by up to +14%.<n>Right-aligned digit grouping enhances large-number arithmetic by +33%.
  arXiv  Detail & Related papers  (2025-06-23T18:02:26Z)
• Spelling-out is not Straightforward: LLMs' Capability of Tokenization from Token to Characters [25.430820735194768]
  Large language models (LLMs) can spell out tokens character by character with high accuracy, yet they struggle with more complex character-level tasks.<n>We investigate how LLMs internally represent and utilize character-level information during the spelling-out process.
  arXiv  Detail & Related papers  (2025-06-12T12:27:41Z)
• Comparative analysis of subword tokenization approaches for Indian languages [5.012314384895538]
  Tokenization is the act of breaking down text into smaller parts, or tokens, that are easier for machines to process.<n>Subword tokenization enhances this process by breaking down words into smaller subword units.<n>It is useful in capturing the intricate structure of words in Indian languages (ILs), such as prefixes, suffixes, and other morphological variations.<n>This paper examines how different subword tokenization techniques, such as SentencePiece, Byte Pair, and WordPiece Tokenization, affect ILs.
  arXiv  Detail & Related papers  (2025-05-22T16:24:37Z)
• ICL CIPHERS: Quantifying "Learning'' in In-Context Learning via Substitution Ciphers [20.65223270978325]
  We introduce ICL CIPHERS, a class of task reformulations based on substitution ciphers borrowed from classic cryptography. In this approach, a subset of tokens in the in-context inputs are substituted with other (irrelevant) tokens, rendering English sentences less comprehensible to human eye. We show that LLMs are better at solving ICL CIPHERS with BIJECTIVE mappings than the NON-BIJECTIVE (irreversible) baseline.
  arXiv  Detail & Related papers  (2025-04-28T00:05:29Z)
• Fast Controlled Generation from Language Models with Adaptive Weighted Rejection Sampling [90.86991492288487]
  evaluating constraint on every token can be prohibitively expensive. LCD can distort the global distribution over strings, sampling tokens based only on local information. We show that our approach is superior to state-of-the-art baselines.
  arXiv  Detail & Related papers  (2025-04-07T18:30:18Z)
• Enhancing LLM Character-Level Manipulation via Divide and Conquer [74.55804812450164]
  Large Language Models (LLMs) have demonstrated strong generalization capabilities across a wide range of natural language processing (NLP) tasks. They exhibit notable weaknesses in character-level string manipulation, struggling with fundamental operations such as character deletion, insertion, and substitution. We propose Character-Level Manipulation via Divide and Conquer, a novel approach designed to bridge the gap between token-level processing and character-level manipulation.
  arXiv  Detail & Related papers  (2025-02-12T07:37:39Z)
• Vulnerability of LLMs to Vertically Aligned Text Manipulations [108.6908427615402]
  Large language models (LLMs) have become highly effective at performing text classification tasks. modifying input formats, such as vertically aligning words for encoder-based models, can substantially lower accuracy in text classification tasks. Do decoder-based LLMs exhibit similar vulnerabilities to vertically formatted text input?
  arXiv  Detail & Related papers  (2024-10-26T00:16:08Z)
• Impact of Non-Standard Unicode Characters on Security and Comprehension in Large Language Models [0.0]
  We present a comparative analysis of the performance of fifteen distinct models. The models are assessed based on the total occurrences of jailbreaks, hallucinations, and comprehension errors. By incorporating alphanumeric symbols from Unicode outside the standard Latin block and variants of characters in other languages, we observed a reduction in the efficacy of guardrails.
  arXiv  Detail & Related papers  (2024-05-23T12:24:38Z)
• Evaluating Subword Tokenization: Alien Subword Composition and OOV Generalization Challenge [10.721272718226848]
  We propose a combined intrinsic-extrinsic evaluation framework for subword tokenization. Intrepid evaluation is based on our new UniMorph Labeller tool that classifies subword tokenization as either morphological or alien. Our empirical findings show that the accuracy of UniMorph Labeller is 98%, and that alien tokenization leads to poorer generalizations.
  arXiv  Detail & Related papers  (2024-04-20T06:49:15Z)
• Revisiting subword tokenization: A case study on affixal negation in large language models [57.75279238091522]
  We measure the impact of affixal negation on modern English large language models (LLMs) We conduct experiments using LLMs with different subword tokenization methods. We show that models can, on the whole, reliably recognize the meaning of affixal negation.
  arXiv  Detail & Related papers  (2024-04-03T03:14:27Z)
• Identifying and Analyzing Task-Encoding Tokens in Large Language Models [55.03191279766383]
  In this paper, we identify and analyze task-encoding tokens on whose representations the task performance depends. We show that template and stopword tokens are the most prone to be task-encoding. Our work sheds light on how large language models (LLMs) learn to perform a task from demonstrations, deepens our understanding of the varied roles different types of tokens play in LLMs, and provides insights for avoiding instability from improperly utilizing task-encoding tokens.
  arXiv  Detail & Related papers  (2024-01-20T20:55:21Z)
• The first step is the hardest: Pitfalls of Representing and Tokenizing Temporal Data for Large Language Models [10.414206635385632]
  Large Language Models (LLMs) have demonstrated remarkable generalization across diverse tasks. A notable obstacle emerges when feeding numerical/temporal data into these models, such as data sourced from wearables or electronic health records. We discuss recent works that employ LLMs for human-centric tasks such as in mobile health sensing and present a case study showing that popular LLMs tokenize temporal data incorrectly.
  arXiv  Detail & Related papers  (2023-09-12T13:51:29Z)
• Exposing Attention Glitches with Flip-Flop Language Modeling [55.0688535574859]
  This work identifies and analyzes the phenomenon of attention glitches in large language models. We introduce flip-flop language modeling (FFLM), a family of synthetic benchmarks designed to probe the extrapolative behavior of neural language models. We find that Transformer FFLMs suffer from a long tail of sporadic reasoning errors, some of which we can eliminate using various regularization techniques.
  arXiv  Detail & Related papers  (2023-06-01T17:44:35Z)
• Idioms, Probing and Dangerous Things: Towards Structural Probing for Idiomaticity in Vector Space [2.5288257442251107]
  The goal of this paper is to learn more about how idiomatic information is structurally encoded in embeddings. We perform a comparative probing study of static (GloVe) and contextual (BERT) embeddings. Our experiments indicate that both encode some idiomatic information to varying degrees, but yield conflicting evidence as to whether idiomaticity is encoded in the vector norm.
  arXiv  Detail & Related papers  (2023-04-27T17:06:20Z)
• Towards preserving word order importance through Forced Invalidation [80.33036864442182]
  We show that pre-trained language models are insensitive to word order. We propose Forced Invalidation to help preserve the importance of word order. Our experiments demonstrate that Forced Invalidation significantly improves the sensitivity of the models to word order.
  arXiv  Detail & Related papers  (2023-04-11T13:42:10Z)
• Experiments with adversarial attacks on text genres [0.0]
  Neural models based on pre-trained transformers, such as BERT or XLM-RoBERTa, demonstrate SOTA results in many NLP tasks. We show that embedding-based algorithms which can replace some of the most significant'' words with words similar to them, have the ability to influence model predictions in a significant proportion of cases.
  arXiv  Detail & Related papers  (2021-07-05T19:37:59Z)
• On the Robustness of Language Encoders against Grammatical Errors [66.05648604987479]
  We collect real grammatical errors from non-native speakers and conduct adversarial attacks to simulate these errors on clean text data. Results confirm that the performance of all tested models is affected but the degree of impact varies.
  arXiv  Detail & Related papers  (2020-05-12T11:01:44Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.