Formosan languages are a critically endangered group of Austronesian languages spoken in Taiwan, with severely limited representation in natural language processing (NLP) research and no support in existing language identification (LID) tools. We present the first systematic evaluation of machine learning models for the language identification of Kavalan, a Formosan language with fewer than 300 known speakers. We construct two benchmarks: a deployment-oriented benchmark with languages commonly confused with Kavalan by existing tools, and a linguistically motivated benchmark of typologically related Formosan languages. We evaluate random forest, support vector machine (SVM), and three pre-trained multilingual models using repeated stratified cross-validation. SVM models with character n-gram features achieve the strongest performance on both benchmarks, with a macro F1 of 0.993 on the deployment benchmark and a macro F1 of 0.906 on the Formosan benchmark, while remaining computationally inexpensive and effective with a low amount of data. Pre-trained multilingual models degrade significantly on the Formosan benchmark, with XLM-RoBERTa falling to a macro F1 of 0.505. These results demonstrate that traditional n-gram-based approaches are effective with low-resource Formosan LID and establish a foundation for downstream NLP tasks supporting the documentation and revitalization of low-resource Formosan languages.

Large-scale corpora of Holocaust testimonies preserve vast amounts of historical, emotional, and narrative information, but their size and complexity can make accurate, systematic analysis challenging. This paper presents a cross-modal computational analysis of emotional and thematic trajectories in the CORHOH corpus, containing 500 Holocaust survivor testimonies as a language resource for computational analysis. We segment each testimony into ten segments and apply sentiment analysis, emotion recognition, and topic modeling to each of these segments to reveal how theme and emotion evolve over time in Holocaust testimonies. Results reveal a sharp decline from pre-war life in wartime and camp experiences, with sentiment and emotion remaining negative in post-war segments. Emotion analysis reveals decreasing joy and increasing sadness and fear during segments related to deportation and concentration camps, with limited emotional recovery. Topic modeling identifies coherent themes that align closely with sentiment and emotional patterns. We systematically examine correlations between sentiment, emotion, and topic trajectories, which demonstrate many strong associations between topic and emotion. This work demonstrates that combining sentiment analysis, emotion recognition, and topic modeling can reveal systematic patterns in large oral history corpora, and shows the value of computational approaches for studying historical narratives like the Holocaust.

Language of study is an aspect of computational linguistics papers that is useful for analyses of trends and diversity in computational linguistics. This study introduces the first benchmark and evaluation of automated language of study extraction from computational linguistics publications. The benchmark containing 431 publications from the ACL Anthology, with 62 languages analyzed, was annotated. SciBERT and four large language models (LLMs), GPT-4o mini, Gemini 2.5 Flash, Claude 3.5 Haiku, and DeepSeek 3.2, were evaluated on the benchmark using different parts of the ACL Anthology papers. GPT-4o mini achieved the best exact match and Jaccard agreement scores of 0.646 and 0.687, respectively, which is slightly less than the agreement in human annotation. Gemini 2.5 Flash achieved the best micro F1 of 0.633. Models using the abstract for extraction were competitive with models using the full text, showing that accuracy can be achieved in language of study extraction without high computational costs. These findings demonstrate that LLMs are able to accurately identify the languages of study in computational linguistics papers, potentially reducing the time and cost of analyses in computational linguistics.

Formosan languages are a critically endangered branch of the Austronesian family spoken in Taiwan, and many of their dialects remain poorly understood and computationally understudied. Subgrouping relationships in these languages are often contested and unresolved. We provide the first evaluation of the ePark corpus as a dialectal NLP resource, identifying its strengths and gaps for future NLP work, and present the first large-scale corpus-based computational analysis of dialect similarity across all officially recognized Formosan languages. We use the ePark corpus to analyze 42 dialects in 16 Formosan languages, and through word-level TF-IDF cosine similarity, Jaccard similarity over shared vocabulary, and Levenshtein distance, we quantify pairwise dialectal relationships within the Amis, Atayal, Seediq, Bunun, Paiwan, Rukai, and Puyuma languages. We find that simple lexical similarity methods can recover and confirm linguistically established dialectal subgroupings. We find that in multiple cases the two metrics diverge, offering insights on contested subgroupings such as Mantauran Rukai. This work establishes a scalable methodological framework for dialectometry in low-resource languages, demonstrates the value of the ePark corpus for Formosan NLP research, and encourages future work in NLP on Formosan dialects.

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.

Resolving contradictions in biomedical literature requires more than factual recall; it demands identifying the hidden variables that explain divergent findings. Existing NLI benchmarks such as MedNLI operate at the sentence level and fail to capture document-level conflicts driven by differences in dosage, cell type, or study design. We introduce BioConflict, a benchmark of 250 expert-annotated paper pairs (500 abstracts) across ten biomedical topics, formalizing three tasks: conflict detection, contextual variable extraction, and consensus synthesis. We evaluate five general-purpose large language models and two domain-specific baselines, finding that general-purpose large language models achieve strong conflict detection (F1 up to 0.89) but exhibit brittle reasoning in synthesis, while domain-specific models lag significantly on all generative tasks. These findings highlight the need for context-aware biomedical AI capable of resolving, not merely retrieving, conflicting scientific evidence.

Large language models (LLMs) struggle with low-resource polysynthetic languages, yet the nature of their failures remains underexplored. We evaluate GPT-4o-mini, Gemma~3~27B, Llama~3.3~70B, and NLLB-200 on Spanish$\leftrightarrow$Wixárika translation using zero-shot and 5-shot prompting. All systems are unusable, scoring below 3 BLEU and 21 chrF. Qualitative analysis reveals that LLMs largely ignore source content and instead generate fluent hallucinations. Spanish outputs frequently include indigenous cultural stereotypes such as deer, deities, rain dance, and shamans, regardless of the input, while Wixárika outputs are repetitive across different inputs and morphologically implausible. Few-shot prompting yields model-dependent improvements, with Gemma and Llama improving substantially at higher shot counts while GPT-4o-mini remains flat. These results demonstrate that current LLMs are unable to represent polysynthetic morphology and instead default to exoticizing Indigenous culture and identity. We call for the development of inclusive morphological-aware modeling strategies and increased resource creation to ensure that Indigenous languages of the Americas are represented safely and accurately.

Kazakh is a Turkic language using the Arabic, Cyrillic, and Latin scripts, making it unique in terms of optical character recognition (OCR). Work on OCR for low-resource Kazakh scripts is very scarce, and no OCR benchmarks or images exist for the Arabic and Latin scripts. We construct a synthetic OCR dataset of 7,219 images for all three scripts with font, color, and noise variations to imitate real OCR tasks. We evaluated three multimodal large language models (MLLMs) on a subset of the benchmark for OCR and language identification: Gemma-3-12B-it, Qwen2.5-VL-7B-Instruct, and Llama-3.2-11B-Vision-Instruct. All models are unsuccessful with Latin and Arabic script OCR, and fail to recognize the Arabic script as Kazakh text, misclassifying it as Arabic, Farsi, and Kurdish. We further compare MLLMs with a classical OCR baseline and find that while traditional OCR has lower character error rates, MLLMs fail to match this performance. These findings show significant gaps in current MLLM capabilities to process low-resource Abjad-based scripts and demonstrate the need for inclusive models and benchmarks supporting low-resource scripts and languages.

Named Entity Recognition (NER) models trained on clean text often fail on real-world data containing orthographic noise. Work on NER for Persian is emerging, but it has not yet explored the orthographic robustness of models to perturbations often exhibited in user-generated content. We evaluate ParsBERT, ParsBERT v2.0, BertNER, and two XLM-r-based models on a subset of Persian-NER-Dataset-500k after applying eleven different perturbations, including simulated typos, code-switching, and segmentation errors. All models were competitive with each other, but XLM-r-large consistently displayed the best robustness to perturbations. Code-switching, typos, similar character swaps, segmentation errors, and noisy text all decreased F1 scores, while Latinized numbers increased F1 scores in ParsBERT. Removing diacritics, zero-width non-joiners, and normalizing Yeh/Kaf all did not have an effect on F1. These findings suggest that Persian NER models require improvement for performance on noisy text, and that the Perso-Arabic script introduces unique factors into NER not present in many high-resource languages, such as code-switching and Eastern Arabic numerals. This work creates a foundation for the development of robust Persian NER models and highlights the necessity of evaluating low-resource NER models under challenging and realistic conditions.