We propose a corpus-dependent alternative to byte encoding that learns fixed-length atomic codes for characters directly from text, which we refer to as Latom (Learned Atom-based Encoding).We instantiate this framework by training an HMM on N -repeated character sequences to estimate "atom" posteriors, followed by a Hungarian assignment yielding a globally optimal one-to-one character-code mapping.Across 14 languages, the encodings improve intrinsic metrics, including token counts after subword tokenization and bigram perplexity, with appropriate code lengths.On Amazon Reviews in six languages, Latom improves text classification accuracy and reduces decoding errors in language model generation.Overall, these results demonstrate that character encodings can be learned from corpus statistics while remaining reversible and compatible with standard tokenization pipelines.

Munachiso S Nwadike, Zangir Iklassov, Toluwani Aremu, Tatsuya Hiraoka, Benjamin Heinzerling, Velibor Bojkovic, Hilal AlQuabeh, Martin Takáč, Kentaro Inui. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

NLP datasets may still contain annotation errors, even when they are manually annotated. Researchers have attempted to develop methods to automatically reduce the adverse effect of errors in datasets. However, existing methods are time-consuming because they require many trained models to detect errors. This paper proposes a time-saving method that utilizes a tokenization technique called subword regularization to simulate multiple error detection models for detecting errors. Our proposed method, SubRegWeigh, can perform annotation weighting four to five times faster than the existing method. Additionally, SubRegWeigh improved performance in document classification and named entity recognition tasks. In experiments with pseudo-incorrect labels, SubRegWeigh clearly identifies pseudo-incorrect labels as annotation errors. Our code is available at https://github.com/4ldk/SubRegWeigh.

This paper investigates how LLMs encode inputs with typos. We hypothesize that specific neurons and attention heads recognize typos and fix them internally using local and global contexts. We introduce a method to identify typo neurons and typo heads that work actively when inputs contain typos. Our experimental results suggest the following: 1) LLMs can fix typos with local contexts when the typo neurons in either the early or late layers are activated, even if those in the other are not. 2) Typo neurons in the middle layers are the core of typo-fixing with global contexts. 3) Typo heads fix typos by widely considering the context not focusing on specific tokens. 4) Typo neurons and typo heads work not only for typo-fixing but also for understanding general contexts.

We present a subword regularization method for WordPiece, which uses a maximum matching algorithm for tokenization. The proposed method, MaxMatch-Dropout, randomly drops words in a search using the maximum matching algorithm. It realizes finetuning with subword regularization for popular pretrained language models such as BERT-base. The experimental results demonstrate that MaxMatch-Dropout improves the performance of text classification and machine translation tasks as well as other subword regularization methods. Moreover, we provide a comparative analysis of subword regularization methods: subword regularization with SentencePiece (Unigram), BPE-Dropout, and MaxMatch-Dropout.

We propose a method to pay attention to high-order relations among latent states to improve the conventional HMMs that focus only on the latest latent state, since they assume Markov property. To address the high-order relations, we apply an RNN to each sequence of latent states, because the RNN can represent the information of an arbitrary-length sequence with their cell: a fixed-size vector. However, the simplest way, which provides all latent sequences explicitly for the RNN, is intractable due to the combinatorial explosion of the search space of latent states. Thus, we modify the RNN to represent the history of latent states from the beginning of the sequence to the current state with a fixed number of RNN cells whose number is equal to the number of possible states. We conduct experiments on unsupervised POS tagging and synthetic datasets. Experimental results show that the proposed method achieves better performance than previous methods. In addition, the results on the synthetic dataset indicate that the proposed method can capture the high-order relations.

For unsegmented languages such as Japanese and Chinese, tokenization of a sentence has a significant impact on the performance of text classification. Sentences are usually segmented with words or subwords by a morphological analyzer or byte pair encoding and then encoded with word (or subword) representations for neural networks. However, segmentation is potentially ambiguous, and it is unclear whether the segmented tokens achieve the best performance for the target task. In this paper, we propose a method to simultaneously learn tokenization and text classification to address these problems. Our model incorporates a language model for unsupervised tokenization into a text classifier and then trains both models simultaneously. To make the model robust against infrequent tokens, we sampled segmentation for each sentence stochastically during training, which resulted in improved performance of text classification. We conducted experiments on sentiment analysis as a text classification task and show that our method achieves better performance than previous methods.