Diacritics are orthographic marks that clarify pronunciation, distinguish similar words, or alter meaning. They play a central role in many writing systems, yet their impact on language technology has not been systematically quantified across scripts. While prior work has examined diacritics in individual languages, there’s no cross-linguistic, data-driven framework for measuring the degree to which writing systems rely on them and how this affects downstream tasks. We propose a data-driven framework for quantifying diacritic complexity using corpus-level, information-theoretic metrics that capture the frequency, ambiguity, and structural diversity of character-diacritic combinations. We compute these metrics over 24 corpora in 15 languages, spanning both single- and multi-diacritic scripts. We then examine how diacritic complexity correlates with performance on the task of diacritics restoration, evaluating BERT- and RNN-based models. We find that across languages, higher diacritic complexity is strongly associated with lower restoration accuracy. In single-diacritic scripts, where character-diacritic combinations are more predictable, frequency-based and structural measures largely align. In multi-diacritic scripts, however, structural complexity exhibits the strongest association with performance, surpassing frequency-based measures. These findings show that measurable properties of diacritic usage influence the performance of diacritic restoration models, demonstrating that orthographic complexity is not only descriptive but functionally relevant for modeling.

The Private-Use Area (PUA) designation plays an important role in the Unicode standard. It covers several ranges of Unicode code points with no official character assignments. A PUA range is primarily used as a temporary representation mechanism for characters falling outside the official standard, to facilitate text entry and display of orthographies that are not otherwise adequately represented. The primary downside of PUA use is that characters lose their semantics if the pairing with the corresponding display font is broken, in which case they cannot be faithfully displayed in a general setting. Large-scale multilingual web corpora invariably contain PUA code points of unclear provenance, which may commonly be treated as noise and discarded. We investigate the distribution of PUA characters within large-scale web corpora, and analyze the resulting distributions across both scripts and writing systems. We show that, while the proportion of PUA-bearing paragraphs in the original corpora are small, PUA-bearing tokens can signal texts from under-represented languages. We additionally explore whether an off-the-shelf large language model (LLM) can classify PUA characters as constituting relevant orthographic signals versus punctuation or other noise. Our methods identify millions of paragraphs making use of such characters, and we argue that such data is important for the long tail of data-scarce orthographies. Moreover, as a primary Unicode mechanism for poorly represented writing systems, PUA characters are here to stay.

In this paper, we present HAnnoI – short for Handwriting Annotation Interface –, an open-source GUI application developed in Python that allows its users to identify and annotate so-called Regions of Interest (ROIs) within digital images. Several meta data such as their coordinates are retained for each ROI and they can be annotated on user-defined annotation layers. HAnnoI comes with a function to export all annotations to a CSV file, enabling further processing as well as quantitative analyses. HAnnoI also has a function to extract single PNG image files of all ROIs. It was developed to mark and annotate single letters in scans of handwritten (alphabetic) texts for linguistic analyses, yet it is not limited to this particular use case. In this paper, we first provide information on HAnnoI’s conception and technical details as well as an overview of alternative applications. We then showcase HAnnoI’s capabilities in a letter annotation task, where five annotators marked instances of lower case <s> in handwritten texts. Finally, we report on an exploratory analysis of this data, showing what kinds of investigations are enabled by using HAnnoI. The tool is available for free use at https://github.com/pywielR/HAnnoI.

Phoneticians and phonologists have developed featural systems that enable systematic description of human speech sounds. However, no such systems exist for describing the visual features of writing systems. It is critical to understand the features of writing systems given their central role in many language users’ everyday experience. Just as phonetic and phonological features provide insight into speech perception, visual features can play a similar role for studying reading. In this paper, we introduce SoriGraph, a database of visual feature descriptions and IPA transcriptions for the full lexicon of Korean, drawing on a recent large-scale study of the visual features of writing systems. This database enables analysis of the visual and phonological properties of Korean and will be a critical resource for researchers. We describe the construction of the database and provide an overview of several potential uses of the database, and demonstrate one potential usage (information-theoretic analysis of lexicon structure).

The Abkhaz-Adyghe and Nakh-Daghestanian language families encompass 35 living languages that possess arguably the most complex modern Cyrillic orthographies due to their very sophisticated phonology. The relevant online data displays idiosyncratic patterns among which the use of confusable characters in input methods is the most prevalent. This work studies one such character—letter palochka—that is shared by most of the writing systems in question. We investigate whether patterns including variants of this character can act as markers of these languages in large-scale web-crawled data. We use GlotLID, a wide-coverage off-the-shelf language identification (LID) model, to label paragraph-level web text that contains a palochka confusable, and estimate the effect of confusable character normalization on the quality of GlotLID’s predictions in 14 supported North Caucasian languages. According to GlotLID, the normalization significantly increases the recall (discovery of new language data) for some languages, while degrading it for others. However, manual evaluation reveals that overall, only 41% of ensuing wins and 46% of losses are accurate due to GlotLID prediction errors. We argue that, despite finding useful signals, higher precision LID approaches tailored to these long-tail languages are needed to improve the quality of mined data.

Abbreviations are an entrenched feature of the majority of the world’s writing systems, and the ability to expand abbreviations in context is important for speech technologies and language understanding tasks. This study presents several experiments applying large language models, prompt engineering, and fine-tuning to the expansion of ad-hoc abbreviations in English, showing substantial improvements against noisy channel models previously used for this task.

This paper evaluates three data augmentation strategies for training misspelling detection models in Spanish. Using the Spanish CORRSIC corpus of naturally occurring misspellings, we compare three misspelling generation methods: random perturbations, keyboard-based errors, and a statistical model derived from empirical edit patterns encoded as weighted finite-state transducers. We also analyze two word selection strategies (random and length-based) and two augmentation configurations designed to balance data diversity and reduce spurious correlations. This study shows that the statistical model produces misspellings most similar to real data, showing the lowest Jensen–Shannon divergence (0.148 nats) with the empirical distribution. In downstream detection experiments, performance improves with training size, and differences between word selection strategies remain minimal. Overall, the results highlight the value of statistically grounded misspelling generation for realistic and effective data augmentation in spell-checking tasks in Spanish.

Kazakh is written in the Arabic, Cyrillic, and Latin script which present unique challenges for OCR and post-OCR correction research. Despite this complexity, NLP research on Kazakh and its low-resource scripts remains extremely scarce. We analyze common OCR error patterns in all three Kazakh scripts using Tesseract and evaluate four large language models (LLMs) for post-OCR correction using minimal, confusion-aware, and few-shot prompting strategies. Our results reveal three systematic, writing-system-driven failure modes in LLM-based post-OCR correction: script switching, hallucination, and instruction-following breakdown. Arabic script post-OCR correction remains unsuccessful across all setups. In the Cyrillic script, post-OCR correction improvements are minimal due to the high baseline OCR performance on Cyrillic. For the Latin script, few-shot prompting with Gemini 2.5 Flash yields substantial improvements, reducing CER by 8.58 points and WER by 32.49 points to levels better than high-resource Kazakh Cyrillic script OCR. These findings demonstrate that LLM post-OCR correction failure modes are predictable from writing system properties such as script resource asymmetry and co-existing script dominance and demonstrate the need for typology-aware evaluation frameworks for multi-script and under-resourced languages.

Grapheme-to-phoneme (G2P) conversion plays a central role in speech technologies. This paper introduces G&P2P, a multi-source framework that integrates multiple pronunciation dictionaries to enhance G2P modeling. We evaluate both expert-curated and crowd-sourced resources using attentive LSTM, pointer-generator LSTM, and transformer architectures. Results indicate that combining high-quality expert dictionaries yields substantial improvements, achieving an 11.26-point absolute (22% relative) reduction in word error rate. In contrast, incorporating noisy crowd-sourced resources may degrade performance. Statistical analyses further suggest that dataset quality exerts a greater influence on outcomes than the choice of fusion strategy, offering practical guidance for the design of multi-source G2P systems.

We present a lightweight, corpus-based approach to abbreviation expansion that relies solely on contextual N-gram statistics. The method models local context using two-sided and one-sided bigram and trigram counts extracted from a large domain-specific corpus. Candidate expansions are selected through linear interpolation of context-specific evidence, enhanced with reliability-based scaling to mitigate sparse data effects. The approach does not require external linguistic resources, pretrained language models, or explicit morphosyntactic analysis, making it suitable for domain-specific and resource-constrained settings. Experiments conducted on a large Slovene medical corpus demonstrate that interpolation generally outperforms strict backoff strategies, with notable improvements for medium- and low-frequency abbreviations. Despite its simplicity, the proposed framework achieves robust performance while remaining computationally efficient and scalable.

Streaming multilingual speech recognition benefits from unified systems that produce numbers in their written form ‘34’ rather than their spoken form ‘thirty-four’. By generating digits directly, these systems eliminate the post-processing latency inherent in cascaded architectures that require a separate inverse text normalization (ITN) step. Arabic presents a formidable challenge for ITN; the system must not only determine the correct numerical value but also navigate complex rules for gender, number, and case marking that are determined by the counted noun. For instance, the digit ‘7’ (as in ‘47’) exhibits gender polarity: it must take a masculine form if modifying a feminine noun (e.g., Halala) and a feminine form if modifying a masculine noun (e.g., Riyal). While Arabic dialects typically exhibit simplified numeral systems by omitting case and gender markers, they vary significantly in verbalization patterns. This study explores the efficacy of a unified streaming Automatic Speech Recognition (ASR) system with integrated ITN features, comparing it against a traditional cascaded approach utilizing a post-processing rule-based ITN module. We utilize a FastConformer cache-aware streaming model trained on English and a diverse Arabic corpus spanning Modern Standard (MSA), dialectal, and Classical Arabic, while maintaining diacritics where contextually appropriate. We evaluate the system using Word Error Rate (WER) for ASR accuracy and exact match for ITN capability. Our results demonstrate that integrating ITN does not degrade core ASR performance and that the unified model achieves accuracy competitive with cascaded systems across Arabic variants. However, error analysis reveals that the primary failures in ITN are rooted in diacritization, gender polarity, and orthographic variation, highlighting the challenges of Arabic’s unique linguistic features in end-to-end modeling.