The Nepali language has a rich and complex morphology. Existing lemmatization research focuses on traditional rule-based or TRIE-based approaches. These methods often fail when encountering out-of-vocabulary or misspelled words. This paper investigates neural lemmatization for the under-resourced Nepali language using multilingual transformer models. We formulate lemmatization as a text-to-text generation problem and evaluate its impacts on downstream tasks by finetuning mBART-large-50, mT5-base, and mT5-small. The models were trained on a combination of publicly available and human-annotated word-lemma pair (8,000 instances) dataset. The performance is evaluated using Character Error Rate (CER), accuracy, character-level Bilingual Evaluation Understudy (BLEU), and morphological coverage. The mT5-base model achieved the highest overall performance. The model achieved 96.1% accuracy and a 1.1% CER using a learning rate of 5 × 10−4. However, it showed slightly weaker performance in handling complex morphological variations. The mBART-large-50 model followed closely with 96.0% accuracy and 0.970 morphological coverage. To assess the efficacy of these models, we applied lemmatization to downstream tasks. In Hindi-Nepali cross-lingual alignment, performance improved significantly from 12.86% to 41.61% using mBART model. In information retrieval, the Mean Average Precision (MAP)@1 using binary index increased from 0.71 to 0.90 using mBART model. These results demonstrate that multilingual transformers effectively learn morphological transformations for low-resource languages through text-to-text generation.

In this paper, we describe our participation in the Shared Task on Multimodal Identification of Vaccine Critical Content on Social Media (VaxMeme) of EEUCA 2026, a satellite of ACL 2026. We tackle the classification of Twitter-based vaccine memes into anti-vaccine, neutral, and pro-vaccine categories using the VaxMeme dataset with 8,195 train, 1,024 val, and 1,025 test samples. We experiment with two different architecture families: (i) Multimodal hybrids: CLIP ViT-B/32 for images + BERT-based models for texts (BERT-base-uncased, ModernBERT) with late fusion strategy based on concatenation of L2-normalized feature vectors and (ii) Text-only: pre-trained models for texts (BERT-base-uncased, RoBERTa-base, ModernBERT-base, DistilBERT-base, Deberta-v3-base) for post_text. In both cases, we use a three-layer feed-forward network with GELU activation for classification. We use class-weighted cross-entropy loss, differential learning rates, AdamW optimizer, gradient accumulation, OneCycleLR scheduler, and early stopping on the val set for optimization. Data augmentation is applied for the multimodal CLIP-based approach only. The winning approach among those tested is the text-only BERT-base-uncased with a macro-F1 of 0.8102 which is ahead of the performance of the CLIP + BERT-base hybrid model, which achieves a test macro-F1 of 0.7603.

The detection of toxic behavior in online gaming communities is crucial for maintaining safe digital spaces, yet remains challenging due to subtle context-dependent and intent-driven language. The GameTox dataset consists of around 53K World of Tanks chat utterances annotated across six categories: Non-toxic, Insults and Flaming, Other Offensive Texts, Hate and Harassment, Threats, and Extremism (CITATION). Our best performing approach, across multiple transformer-based architecture experimentations, is based on the multilingual BERT variant mmBERT-base fine-tuned with class-weighted cross-entropy loss. The best mmBERT-base model achieved a Macro F1 of 0.5882 during validation and an official test Macro F1 of 0.5104 on the shared task leaderboard. An internal held-out evaluation on a development split yielded 0.4282, which we analyze to understand distributional sensitivity to gaming slang and class imbalance. The code is available at: https://github.com/sunilRegmi-ai/eeuca-toxicity-detection.

The increased sharing of memes on social media creates serious challenges for automated moderation, especially in low-resource and code-mixed languages such as Nepali. In this paper, we present our system for the CHiPSAL 2026 Shared Task on Multimodal Hate and Sentiment Understanding in Low-Resource Memes. We propose a late-fusion hybrid architecture that combines OpenAI’s Vision Transformer (CLIP ViT-B/32) with a domain-specific Nepali language model (NepBERTa) to capture both visual features and linguistic information. To address data scarcity, we introduce a cross-task label mapping and data augmentation strategy between the hate speech and sentiment datasets. By applying controlled hyperparameter settings and balanced loss optimization, our framework achieved a Macro F1 score of 0.8052 on Subtask A (Hate Speech Detection) and 0.6881 on Subtask B (Sentiment Analysis) in the official CodaBench evaluation, demonstrating the effectiveness of the proposed multimodal approach.

Fine tuning speech recognition models on noisy real world data is tricky. The model has no way of knowing which training samples are reliable and which are not, so it ends up learning from bad examples just as readily as good ones. This is a real problem for Nepali, where most available training data comes from YouTube videos with automatically generated subtitles that are often inaccurate. In this work, we tried a simple fix. Instead of feeding everything to the model, we first asked humans to rate the quality of a sample of transcriptions, trained a small Random Forest classifier on those 2,000 ratings, and used it to filter out the bad samples before each retraining round. The classifier uses four automatically computable features, Word Error Rate (WER), Character Error Rate (CER), length ratio, and length difference, and achieves 81% accuracy on a held out set. Running two filtering and retraining cycles on a 40,000 clip training subset drawn from a 68.4 hour corpus improves substantially over our own standard fine tuning baseline of 5.60% WER and 5.10% CER, reaching 4.89% WER and 4.52% CER, which corresponds to an 11 to 13% relative gain. The approach is much lighter than full Reinforcement Learning from Human Feedback but still uses real human judgment to guide training.

This paper presents a comparative analysis of FastText and BERT-based approaches for Natural Language Understanding (NLU) tasks in Devanagari script languages. We evaluate these models on three critical tasks: language identification, hate speech detection, and target identification across five languages: Nepali, Marathi, Sanskrit, Bhojpuri, and Hindi. Our experiments, although with raw tweet dataset but extracting only devanagari script, demonstrate that while both models achieve exceptional performance in language identification (F1 scores > 0.99), they show varying effectiveness in hate speech detection and target identification tasks. FastText with augmented data outperforms BERT in hate speech detection (F1 score: 0.8552 vs 0.5763), while BERT shows superior performance in target identification (F1 score: 0.5785 vs 0.4898). These findings contribute to the growing body of research on NLU for low-resource languages and provide insights into model selection for specific tasks in Devanagari script processing.

Large Language Models (LLMs) have made significant advancements in Natural Language Processing (NLP) by excelling in various NLP tasks. This study specifically focuses on evaluating the performance of LLMs for Named Entity Recognition (NER) and Part-of-Speech (POS) tagging for a low-resource language, Nepali. The aim is to study the effectiveness of these models for languages with limited resources by conducting experiments involving various parameters and fine-tuning and evaluating two datasets namely, ILPRL and EBIQUITY. In this work, we have experimented with eight LLMs for Nepali NER and POS tagging. While some prior works utilized larger datasets than ours, our contribution lies in presenting a comprehensive analysis of multiple LLMs in a unified setting. The findings indicate that NepBERTa, trained solely in the Nepali language, demonstrated the highest performance with F1-scores of 0.76 and 0.90 in ILPRL dataset. Similarly, it achieved 0.79 and 0.97 in EBIQUITY dataset for NER and POS respectively. This study not only highlights the potential of LLMs in performing classification tasks for low-resource languages but also compares their performance with that of alternative approaches deployed for the tasks.

In this era of AI and Deep Learning, Speech Recognition has achieved fairly good levels of accuracy and is bound to change the way humans interact with computers, which happens mostly through texts today. Most of the speech recognition systems for the Nepali language to date use conventional approaches which involve separately trained acoustic, pronunciation and language model components. Creating a pronunciation lexicon from scratch and defining phoneme sets for the language requires expert knowledge, and at the same time is time-consuming. In this work, we present an End-to-End ASR approach, which uses a joint CTC- attention-based encoder-decoder and a Recurrent Neural Network based language modeling which eliminates the need of creating a pronunciation lexicon from scratch. ESPnet toolkit which uses Kaldi Style of data preparation is the framework used for this work. The speech and transcription data used for this research is freely available on the Open Speech and Language Resources (OpenSLR). We use about 159k transcribed speech data to train the speech recognition model which currently recognizes speech input with the CER of 10.3%.