arXiv:2609. 21362v1 Announce Type: new Abstract: Conventional tokenizers represent text as characters or statistically derived subwords, overlooking the internal phonological structure of syllables and often requiring large vocabularies.
By Nghia Hieu Nguyen, Thai Bao Huynh, Binh-An Dinh-Le, Phu Gia Hoang, Dat Tien Nguyen, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen
The paper introduces a token‑level extension of Omni‑Temporal Classification (OTC) for automatic speech recognition, allowing unsupported tokens to be bypassed while preserving supervision for the rest of the word. Across 19 languages and three corpora, this token‑level OTC consistently outperforms standard CTC, achieving the lowest mean word error rate on every dataset and a 9.45% average relative WER reduction. A predictive‑entropy‑indexed schedule replaces epoch‑based relaxation, reducing training‑length dependence while maintaining performance.
By Saurabh Kumar, Diptiman Mohanta, Prasanta Kumar Ghosh
arXiv:2607. 23319v1 Announce Type: cross Abstract: Standard subword tokenization algorithms such as Byte-Pair Encoding (BPE) and SentencePiece are trained predominantly on modern language corpora and produce inefficient segmentations when applied to classical Indian languages.
By Poornima Kumaresan, Pavithra Muruganantham, Lakshmi Rajendran, Santhosh Sivasubramani
TokEval is a tokenizer evaluation suite that introduces metrics beyond traditional fertility and compression rate, capturing linguistically and structurally meaningful properties such as UTF‑8 character boundary integrity and digit place‑value alignment for mathematics. The authors validate these metrics by pretraining language models with varied tokenizers and measuring downstream performance on bits‑per‑byte and benchmarks covering linguistic understanding, mathematical reasoning, and code generation. Their results show that information‑theoretic metrics predict language modeling performance, while structure‑sensitive metrics correlate with task accuracy, suggesting TokEval can guide tokenizer selection more principledly.
By Clara Meister
The study investigates how the choice of tokenizer influences multilingual language models across 54 tokenizers and 123 models. It finds that tokenizer impact is greater for languages with less training data, and that excluding a language from tokenizer training consistently worsens its performance. While reallocating tokenizer training data to lower-resource languages can help, it does not guarantee improvement, and the best tokenizer properties vary by language, enabling predictive screening of tokenizer candidates.
By Clara Meister, G\"ul Sena Alt{\i}nta\c{s}, Antoine Bosselut
The paper investigates a failure mode called phonological interference in phoneme‑level multilingual speech models, where models incorrectly impose a single language’s phonology on mixed or unseen language input. Experiments show significant loss of language‑specific phonemes in code‑switched and unseen‑language scenarios, and the authors trace this effect to a low‑dimensional subspace of internal activations. They propose a windowed language estimation (WLE) repair that reduces interference by 34–69% while preserving monolingual performance.
By Moran Yanuka, Raja Giryes, Moris Alper