This paper presents the findings of the second workshop on Challenges in Processing South Asian Languages (CHiPSAL 2026), held as part of LREC 2026. South Asia is one of the most linguistically diverse regions in the world, yet its languages remain severely underrepresented in language resources and technologies, particularly in the era of large language models (LLMs). The workshop brings together research addressing key challenges in this space, including data scarcity, morphological complexity, code-mixing, script diversity, and the lack of culturally grounded evaluation benchmarks. The workshop received 57 submissions, covering a wide range of languages, tasks, and modalities, including both widely spoken languages (e.g., Bengali, Hindi, Tamil, and Urdu) and extremely low-resource and endangered languages such as Burushaski, Limbu, and Nepal Bhasha (Newari). Several contributions introduce arguably first-of-their-kind resources and benchmarks for these languages, spanning both text and speech domains, and focusing on linguistically informed and culturally grounded data creation. In addition to the main track, the workshop hosted a shared task on Multimodal Hate and Sentiment Understanding in Low-Resource Memes for Nepali, attracting strong community participation. The results highlight the effectiveness of multimodal approaches while also revealing persistent challenges in modelling culturally nuanced and low-resource data. Across the accepted papers and shared task, key insights include the central role of high-quality data, the limitations of current multilingual models in low-resource settings, and the need for culturally aware and data-centric approaches. Overall, CHiPSAL 2026 demonstrates the growing momentum in South Asian language processing and highlights the importance of sustained, community-driven efforts to build inclusive and representative language technologies.

Burushaski is a language isolate spoken in northern Pakistan with a predominantly oral tradition, limited standardized orthography, and virtually no existing speech technology infrastructure. These characteristics make conventional text-centric NLP pipelines unsuitable and position speech data collection as the primary scientific challenge. This paper introduces an audio-first, linguistically informed methodology for developing a Burushaski–English speech translation resource. Rather than prioritizing model architecture, we focus on principled corpus design tailored to the language’s morphological complexity, ergative-absolutive alignment, and four-gender agreement system. The dataset combines structured elicitation targeting high-frequency and morphologically diverse constructions, functional and formulaic speech, and oral narratives that capture discourse-level phenomena. We describe the design of a custom data collection application, community-embedded crowdsourcing strategy, and translation-aligned workflow for generating parallel speech–English data. The resulting pilot corpus comprises approximately 10 hours of curated audio from 42 speakers across controlled and naturalistic settings. While we present the results of preliminary translation experiments using Whisper, the primary contribution of this work is methodological: a scalable framework for speech-first corpus development in morphologically rich, under-resourced, and predominantly oral languages. We argue that for languages lacking stable orthography and large textual corpora, data design, not model selection constitutes the central research problem

Despite negation being one of the core element in any language, it remains a challenging phenomenon for modern Large Language Models(LLMs). Recently, there have been growing efforts to evaluate how models handle negation. However, the existing probing datasets are mostly English-centric. To facilitate evaluation for Indian Languages especially Telugu, which has complex morphological features, we present NEGTEG benchmark. This benchmark is a test suite that contains 5 tasks: Negation Detection, Negation Translation, Paraphrase Detection, Sentiment Analysis and Polarity Flipping. The test suite is designed based on strong linguistic analysis and includes annotations of different negation types. This helps us evaluate how models perform across various forms of negation. We use the benchmark to probe the negation handling capabilities of multilingual language models at different levels and our evaluation reveals that most of the models struggle significantly with Telugu negation across all tasks.

This paper presents the first ever morphological transducer for the Limbu language, also known by its endonym Yakthung Pan, an endangered Sino-Tibetan language primarily spoken in the area known as Limbuwan/Koshi Province in Eastern Nepal, with a minority population in the Sikkim state of India, where Limbu enjoys official status. Using a corpus of Limbu text produced by field interviews (Michailovsky, 1977), and a translation of the Holy Bible into the Limbu language, the paper presents various elements of the morphology of the language, how they were implemented into the transducer, and an evaluation of the transducer against identified corpora and a gold standard. With a relatively small lexicon, the transducer was found to have reasonable coverage, with high precision but low recall. The paper discusses future expansion through further involvement with the community, which can help in the maintenance and revitalisation of the endangered language.

Medical named entity recognition (NER) is a crucial task in natural language processing (NLP) for extracting meaningful entities such as diseases, symptoms, medications, body parts, and treatments from clinical text. However, NER in low-resource languages like Urdu remains underexplored due to limited annotated datasets. In this study, we evaluated the performance of two state-of-the-art large language models (LLMs), ChatGPT-4o and LLAMA 3.2, on Urdu medical NER using a dataset of 2,057 health-related Urdu news headlines manually annotated across five entity categories. Both models were evaluated using precision, recall, and F1-score. It was found that both models exhibited low precision and moderate recall. ChatGPT-4o achieved the highest F1 for Disease (0.35) while LLAMA 3.2 reached slightly lower F1 scores for Disease (0.33). Both models performed poorly on treatment-related terms, with F1 scores of 0.036 (LLAMA 3.2) and 0.011 (ChatGPT-4o). Micro-average F1-scores were 0.187 for ChatGPT-4o and 0.183 for LLAMA 3.2, indicating comparable overall performance. These findings highlight the challenges of medical NER in low-resource languages and underscore the need for domain-specific fine-tuning, transfer learning, few-shot learning, and prompt engineering to improve performance.

Retrieval-Augmented Generation performs well on English QA benchmarks, but degrades considerably in morphologically rich, low-resource languages. Urdu presents a particularly challenging case: heavy inflectional morphology, Nastaliq script inconsistencies, and limited training data produce a systematic mismatch between query representations and indexed document content that standard retrieval architectures cannot bridge. We propose DR-RAG (Dual-Representation Retrieval-Augmented Generation), which addresses this through dual indexing. Each document is represented as overlapping text chunks and as automatically generated question-answer pairs. Queries are first matched against the QA index, which aligns more reliably with natural query phrasing than declarative document chunks. When retrieval confidence falls below τ = 0.80, the system falls back to chunk-based retrieval, maintaining coverage without sacrificing precision. Evaluated on Urdu UQA and English SQuAD 2.0, DR-RAG improves Urdu METEOR by 38×, ROUGE-1 by 140%, and reduces generation latency by 43%. LLM-as judge scores show higher faithfulness (3.03 vs 1.93) and overall quality (2.99 vs 2.21) over MultiVector. English performance remains competitive throughout. These results indicate that representation-level alignment between queries and indexed content, rather than increased model complexity, is the critical factor for reliable retrieval in underserved South Asian languages.

Speech Emotion Recognition (SER) is an important part of human–computer interaction, but most existing research focuses on high-resource languages, with very limited work on regional languages such as Punjabi. This paper focuses on detecting emotions from Punjabi speech using machine learning and deep learning techniques. We curated our own Punjabi speech emotion dataset using volunteer recordings and real-world sources, covering four emotion classes: angry, happy, sad, and neutral. The data was preprocessed for consistency and evaluated using a multi-strategy framework (E1–E4) to test domain generalization. Three models were evaluated: CNN, ResNet-34, and the transformer-based Wav2Vec 2.0. Among these, the ResNet-34 model performed the best in the combined-domain strategy (E4), achieving a test accuracy of 96%. While cross-corpus evaluations (E2, E3) highlighted challenges in generalizing to neutral emotions, the model achieved perfect scores for happy and sad classes in E4. These results demonstrate the effectiveness of residual networks and combined-domain training for emotion recognition in low-resource languages and highlight the potential for further work on Punjabi SER.

SiPaKosa is a comprehensive corpus of Sinhala and Pali doctrinal texts comprising approximately 786K sentences and 9.25M words, incorporating 16 copyright-cleared historical Buddhist documents alongside the complete web-scraped Tripi t. aka canonical texts. The corpus was created through high-quality OCR using Google Document AI on historical manuscripts, combined with systematic web scraping of canonical repositories, followed by rigorous quality control and metadata annotation. The corpus is organised into language-specific subcorpora: Sinhala and Mixed Sinhala-Pali. We evaluate the performance of language models using ten pretrained models, with perplexity scores ranging from 1.09 to 189.67 on our corpus. This analysis shows that proprietary models significantly outperform open-source alternatives by factors of three to six times. This corpus supports the pretraining of domain-adapted language models, facilitates historical language analysis, and aids in the development of information retrieval systems for Buddhist scholarship while preserving Sinhala cultural heritage.

Despite the growing progress in Natural Language Inference (NLI) research, resources for the Bengali language remain extremely limited. Existing Bengali NLI datasets exhibit several inconsistencies, including annotation errors, ambiguous sentence pairs, and inadequate linguistic diversity, which hinder effective model training and evaluation. To address these limitations, we introduce BNLI, a refined and linguistically curated Bengali NLI dataset designed to support robust language understanding and inference modeling. The dataset was constructed through a rigorous annotation pipeline emphasizing semantic clarity and balance across entailment, contradiction, and neutrality classes. We benchmarked BNLI using a suite of state-of-the-art transformer-based architectures, including multilingual and Bengali-specific models, to assess their ability to capture complex semantic relations in Bengali text. The experimental findings highlight the improved reliability and interpretability achieved with BNLI, establishing it as a strong foundation for advancing research in Bengali and other low-resource language inference tasks. The link to the BNLI dataset: https://github.com/FarahHaque/BNLI-Dataset.git

Text classification in low-resource languages has become increasingly important due to the rapid growth of user-generated digital content. While multitask learning has long been studied in NLP, the use of LLMs for multitask text classification in low-resource languages such as Bengali remains underexplored. Although LLMs are inherently multilingual and multitasking, their effectiveness in structured multitask classification settings for Bengali has not been systematically evaluated. In this work, we investigate how LLMs can be leveraged for multitask Bengali text classification across five domains: sentiment analysis, aggressive text detection, fake news detection, news categorization, and emotion analysis. We compare in-context learning strategies—including zero-shot, one-shot, and chain-of-thought prompting—with parameter-efficient fine-tuning approaches. Our findings show that CoT prompting does not consistently improve performance and often degrades performance, highlighting the instability of prompt-based adaptation in low-resource settings with limited pretraining exposure. Moreover, reasoning-optimized models such as DeepSeek-R1 exhibit substantial performance drops, indicating that enhanced reasoning capabilities alone cannot overcome the challenges posed by low-resource settings. Among the evaluated mLLMs, Gemma-3-4B demonstrates the most stable and balanced cross-task performance under both in-context learning and parameter-efficient fine-tuning, making it a strong backbone candidate for multitask Bengali text classification. These results provide empirical evidence on the limitations of prompting and the advantages of lightweight fine-tuning for low-resource multilingual NLP.

Nepal Bhasha (Newari), an endangered language of the Kathmandu Valley, remains digitally marginalized due to the severe scarcity of annotated speech resources. In this work, we introduce Nwāchā Munā, a newly curated 5.39-hour manually transcribed Devanagari speech corpus for Nepal Bhasha, and establish the first benchmark using script-preserving acoustic modeling. We investigate whether proximal cross-lingual transfer from a geographically and linguistically adjacent language (Nepali) can rival large-scale multilingual pretraining in an ultra-low-resource Automatic Speech Recognition (ASR) setting. Fine-tuning a Nepali Conformer model reduces the Character Error Rate (CER) from a 52.54% zero-shot baseline to 17.59% with data augmentation, effectively matching the performance of the multilingual Whisper-Small model despite utilizing significantly fewer parameters. Our findings demonstrate that proximal transfer from Nepali language serves as a computationally efficient alternative to massive multilingual models. We openly release the dataset and benchmarks to digitally enable the Newari community and foster further research in Nepal Bhasha.

Romanized Nepali is the dominant medium of social media communication in Nepal, yet most multilingual NLP models are trained on Devanagari, creating a noticeable drop in performance in informal settings. To address this script mismatch, we develop NepaliXlit, a transliteration model fine-tuned from IndicXlit to better handle the phonetic variability of Romanized Nepali. Trained on 2,943 informal word pairs and evaluated on 736 held-out pairs, NepaliXlit improves transliteration accuracy by 8% and reduces character error rate by 11%. We use sentiment analysis as a testbed to understand whether transliteration actually helps downstream NLP tasks. We curate over 6,500 Romanized social media comments and construct a balanced subset of 1,518 manually annotated instances. Baseline experiments show that multilingual encoder models struggle with Romanized input; however, transliterating text into Devanagari using NepaliXlit consistently improves sentiment classification accuracy with mBERT and MuRIL. Comparative evaluation against large language models (LLMs) further reveals that generative models such as Gemini and GPT variants exhibit strong cross-script generalization and outperform encoder-based baselines. Our results indicate that adaptive transliteration enhances conventional multilingual models, while modern LLMs offer a better alternative for multi-script, low-resource settings.

Low-Rank Adaptation (LoRA) enables efficient fine-tuning of large language models, yet why it works particularly well for code-mixed text remains unexplained. We propose that LoRA’s efficiency stems from geometric structure in multilingual pre-trained models: code-mixed embeddings concentrate in low-dimensional cross-lingual subspaces. Through spectral analysis of mBERT and MuRIL on Hindi-English (Hinglish) data, we establish that pre-trained attention weights have effective ranks of 437–441, while LoRA updates (r = 4,8,16) exhibit ranks of 2.1–5.9—a 136× average compression. Cross-lingual geometry measured via Centered Kernel Alignment shows Hinglish embeddings align strongly with Hindi (CKA=0.279) but weakly with English (0.093), compared to a monolingual baseline of 0.074. Statistical tests (Wilcoxon p < 10−19) and permutation ablations confirm these differences are robust. We interpret the convergence of geometric overlap (3.77× baseline) and empirical compression (136×) as evidence that low-rank adaptation exploits pre-existing multilingual structure. Findings are demonstrated on token-level language identification; extensions to other language pairs and tasks remain open questions.

Informal Hindi text frequently contains multi-token slang and idiomatic expressions whose correct identification requires consistent span boundaries. Transformer-based token classifiers, despite strong contextual representations, often produce fragmented or structurally invalid BIO sequences due to largely local predictions. We propose Hi-SEMFLOW, a Lie algebra–based semantic flow framework that models span consistency as a continuous refinement process over label logits. Instead of discrete structured decoding (e.g., CRFs), Hi-SEMFLOW learns context-dependent transition operators derived from antisymmetric generators and propagates structural information through smooth, fully differentiable transformations. This formulation integrates structural bias directly into end-to-end training without requiring dynamic programming or hard decoding constraints. Experiments on the HiSlang-4.9k benchmark show that Hi-SEMFLOW improves span-level F1 by up to 2–3 absolute points and yields consistent macro-F1 gains across Hindi-pretrained encoders. Extensive ablations demonstrate that continuous geometric refinement provides a flexible and effective alternative to discrete structured decoding for span-centric sequence labeling.

Large language models perform strongly on standard evaluations, yet these benchmarks prioritize high-resource languages and culturally dominant knowledge, leaving culture-specific commonsense underexamined. In low-resource languages such as Nepali, everyday communication depends on culturally embedded cues, including kinship hierarchies, ritual practices, food systems, idioms, and honorific distinctions that literal translation often fails to capture. As a result, models that appear competent on global metrics can perform poorly in local contexts. To address this gap, we introduce NeCCo, a curated multiple-choice benchmark for culturally situated reasoning across five domains: kinship and social hierarchy; festivals, rituals, and geography; idioms, proverbs, and metaphors; commonsense and daily life; and gastronomy, agriculture, and nature. The dataset was created through structured authoring, cross-review, and normalization, and is released in Devanagari, English, and Romanized formats. We evaluate multiple state-of-the-art LLMs using standardized prompting and controlled decoding. Results show substantial variation: models perform better on globally documented knowledge such as geography, but struggle with relational and linguistically implicit tasks, including extended kinship reasoning and proverb interpretation. The most culturally dense categories expose brittleness and increased hallucination. These findings suggest that multilingual competence requires more than translation coverage and highlight the need for culturally grounded benchmarks and training signals.

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.

Large Language Models (LLMs) have shown strong generalization across tasks in high-resource languages; however, their linguistic competence in low-resource and morphologically rich languages such as Tamil remains largely unexplored. Existing multilingual benchmarks often rely on translated English datasets, failing to capture the language specific linguistic and cultural nuances of the target language. To address this gap, we introduce ILAKKANAM, the first Tamil-specific linguistic evaluation benchmark manually curated using 820 questions from Sri Lankan school-level Tamil subject examination papers spanning Grades 1–13. Each question is annotated by trained linguists under five linguistic categories and a factual knowledge category. We evaluate both closed-source and open-source LLMs using a standardized evaluation pipeline. Our results show that Gemini 2.5 achieves the highest overall performance, while open-source models lag behind, highlighting the gap in linguistic grounding. Category- and grade-wise analyses reveal that all models perform well on lower-grade questions but show a clear decline as the grade level and the linguistic complexity of the questions increase. Further, no strong correlation is observed between a model’s overall performance and its ability to identify linguistic categories, suggesting that performance may be driven by exposure rather than genuine understanding. The code and dataset used in this study are publicly available in our repository, where the dataset consists only of extracted examination questions to mitigate potential data leakage. Keywords: Tamil, Linguistic Benchmark, Linguistic diagnostics, Low-resource language

Detecting online toxicity in morphologically rich, low-resource languages like Tamil remains a major computational challenge. Standard transformer models often struggle with sub-word fragmentation, which can dilute the semantic intensity of regional insults and out-of-vocabulary slang. To mitigate this limitation, we train a multi-layer hybrid framework that fuses the deep contextual representations of L3Cube-TamilBERT with the character-level robustness of FastText embeddings. Our architecture leverages Last-4 Layers averaging and a dual pooling strategy (Mean + Max) to capture both global sentence intent and extract high-activation spikes of offensive cues typically lost in single layer representations. Experiments show that this hybrid model achieves a Macro-F1 of 0.7883, notably enhancing Hate Recall (0.7503) for detection of offensive content. Additionally, as reported by other studies, stacking ensemble achieves peak hate precision (0.9296), providing a high accuracy alternative for moderation scenarios requiring minimal false positives. By combining deep contextual hidden states with FastText embeddings, the proposed feature-fusion ensemble approach with multi-layer hybrid framework approach establishes a new benchmark for hate speech detection for Tamil.

Tokenization is crucial in NLP, influencing performance for morphologically rich, low resource languages like Tamil. This study comprehensively analyzes WordPiece, SentencePiece, and Byte-Level Byte Pair Encoding (BBPE) for Tamil text classification. We assess tokenization efficiency using metrics including token count, fragmentation, OOV rate, and compression ratio. Additionally, we analyze downstream impact through Tamil news title classification using a custom lightweight BERT based Transformer architecture. Tokenizers were pretrained on a 5.45 GB Tamil Corpus and evaluated on a Kaggle Tamil News Dataset. Results indicate WordPiece and SentencePiece outperform BBPE in efficiency and accuracy. While BBPE eliminates OOV words, excessive fragmentation hinders model learning. Increasing vocabulary size improves WordPiece and SentencePiece but not BBPE. Misclassification analysis highlights overfragmentation challenges. This study contributes to Tamil NLP by comparing tokenizers, aiding researchers in selecting appropriate strategies for agglutinative languages.

The proliferation of health misinformation in Low-Resource Languages (LRLs) poses a severe threat to public health, yet automated detection remains critically under-studied due to the scarcity of high-quality benchmarks. We address this gap by introducing Nep-Health-Misinfo, a novel human-verified corpus for health misinformation identification in Nepali. The dataset was developed by adapting four foundational benchmarks (Monkeypox-V1, Monkeypox-V2, COVID-19, and CoAID) through a systematic Machine Translation Post-Editing (MTPE) protocol involving native experts. Our evaluation of Neural Machine Translation (NMT) systems reveals a significant translation asymmetry: while state-of-the-art (SOTA) systems achieve a BLEU score of 43.21 on factual health data, performance degrades sharply on deceptive narratives, with BLEU and TER scores dropping to 19.11 and 62.42, respectively. To establish robust baselines, we benchmark seven recent open-weight Large Language Models (LLMs), including Qwen2.5-7B-Instruct, Gemma-3-4B-IT, and Ministral-8B-Instruct, across zero-shot and few-shot settings. For the few-shot evaluation, we compare stochastic sampling against a K-means centroid-based approach for semantically representative exemplar selection. Experimental results indicate that Qwen2.5-7B-Instruct achieves a peak Macro F1-score of 0.8488, improving over its zero-shot performance (0.7188) on the same dataset. Our findings demonstrate that while few-shot prompting effectively mitigates distribution shifts in low-resource medical contexts, performance remains highly sensitive to the semantic density of exemplars. This work provides the first human-verified Nepali health misinformation corpus. All code and resources are available at https://github.com/SUJAL390/Nep-Health-Misinfo-CHIPSAL-LREC.

This paper presents an overview of the Shared Task on Multimodal Hate and Sentiment Understanding in Low-Resource Memes, organized as part of the Second Workshop on Challenges in Processing South Asian Languages (CHiPSAL 2026) at LREC 2026. The task addresses automated content understanding in low-resource settings by focusing on monolingual Nepali memes written in Devanagari script. Built upon the NeMeme dataset, the task comprises two subtasks: (1) binary hate speech detection and (2) three-class sentiment analysis. The competition attracted 23 teams for hate detection and 13 teams for sentiment analysis. Participating teams employed diverse strategies, including late-fusion multimodal architectures combining multilingual text encoders with vision models, caption-based approaches using large vision-language models, and ensemble techniques. The top-performing system achieved macro-F1 scores of 80.52% on hate detection and 68.81% on sentiment analysis using a late-fusion hybrid architecture with discriminative learning rates. Our analysis reveals that multimodal fusion consistently outperforms unimodal baselines, sentiment analysis poses greater challenges than hate detection due to increased semantic nuance, and the scarcity of Devanagari-centric pretrained models remains a significant bottleneck. This shared task establishes a benchmark for multimodal understanding in low-resource South Asian languages and provides insights for developing inclusive content moderation systems.

Internet memes pose significant challenges for automatic content moderation due to the interaction of visual and textual cues, sarcasm, and cultural context. In this work, we participate in the CHiPSAL 2026 shared task on multimodal hate and sentiment understanding in Nepali memes. The task consists of two subtasks: binary hate speech detection and three-class sentiment classification. We investigate both early-fusion and late-fusion multimodal architectures. Our primary system employs a late-fusion dual-encoder architecture combining XLM-RoBERTa for multilingual text representation and CLIP for visual encoding. We further evaluate an early-fusion ViLT-based joint vision–language transformer using NepBERTa tokenization as a baseline. Experimental results show that late-fusion models consistently outperform early-fusion architectures, particularly for code-mixed memes containing Devanagari Nepali and Roman-script English text. Our best system achieves a Macro-F1 of 0.6564 for hate speech detection and 0.4859 for sentiment classification. We provide analysis highlighting the challenges of multilingual code-mixing, sarcasm, and implicit sentiment in low-resource multimodal settings.

Memes are widely used for communication on social media but are increasingly exploited to spread hate and harmful stereotypes. Detecting hate speech in memes is particularly challenging because meaning is conveyed jointly through images and embedded text, and the problem becomes more complex in low-resource languages such as Nepali. In this work, we participate in Subtask A of the CHiPSAL 2026 Shared Task, focusing on hate speech detection in Nepali-only memes. We benchmark three multimodal vision language backbones, ViT-B-32 (OpenCLIP), AltCLIP, and BLIP2+mT5, under controlled preprocessing and augmentation settings. Our best-performing system uses AltCLIP to extract aligned text and image representations, followed by a late-fusion classifier trained with stratified 5-fold cross-validation to address class imbalance. The proposed model achieves a macro F1-score of 0.66 on the validation set. Experimental results highlight the effectiveness of aligned vision language representations and demonstrate that preprocessing and augmentation strategies have model-dependent effects in low-resource multimodal hate speech detection.

The analysis of internet memes in the Nepali language is complicated by frequent code-mixing and a lack of established baseline resources. While memes inherently combine visual and textual elements, this study focuses on a text-centric approach by extracting embedded text using an OCR layer and modeling it with Transformer-based architectures. We evaluate six distinct models and investigate the comparative effectiveness of Hard and Soft Voting ensemble strategies across two tasks: binary hate speech detection and three-class sentiment analysis. Experimental results show that a standalone decoder-only model achieved the highest performance for binary classification, whereas the Soft Voting ensemble performed best for the multi-class sentiment task, yielding a 15.8% relative improvement in Macro F1-score over the strongest standalone baseline. These findings suggest that ensemble strategies behave differently across binary and multi-class tasks, highlighting the importance of selecting aggregation methods suited to the classification objective.

Memes have emerged to be an essential medium of online expression, where the sentiment is determined by the interaction of text and image. Sentiment analysis of memes is particularly challenging when the language is low-resource, such as Nepali, due to the lack of resources and the complex relationships between text and image modalities. In this paper, we report our submission to Subtask B of CHiPSAL 2026, where the task was sentiment analysis of Nepali text-embedded memes for three sentiment classes: Negative, Neutral, and Positive. Through this submission, we present a late fusion multimodal framework that encompasses lexical, semantic, and visual models through a cross-validated stacking approach. Our submission to the shared task competition received a Macro F1 of 0.5045 on the official test set, achieving 6th place in the leaderboard. This demonstrates the strength of well-structured late fusion approaches to multimodal sentiment analysis of text-embedded memes.

Memes have become ubiquitous on social media platforms blending text and imagery to express complex and culturally nuanced messages. While a high degree of automation in meme moderation has been achieved for high-resource languages, low-resource languages, such as Nepali, still remain largely neglected. In this paper, we describe our system submission to the CHiPSAL 2026 Shared Task on Multi-modal Hate and Sentiment Understanding in Low-Resource Nepali Memes, which features two main sub-tasks: (1) Detection of HateSpeech as binary classification and (2) Sentiment Analysis as multi-class classification in Nepali memes. We perform a comprehensive analysis of the following models: uni-modal textual models (mBERT, XLM-RoBERTa,MuRIL), uni-modal visual models (ResNet, ConvNeXt, ViT), nine different late-fusion multimodal models, and the vision-language foundation model, SigLIP. Among all models, the ViT model achieved the best macro F1-score(0.6278) for the hate speech detection task, while SigLIP achieved the best score (0.5481) for the sentiment analysis task. We hypothesize that the under-performance of fusion models may be attributed to OCR noise and inadequate low-resource textual representations that act as a bottleneck when paired with more advanced visual encoders. These results highlight the unique challenges of multimodal meme comprehension in low-resource contexts and underscores the requirement for culturally grounded, noise-robust approaches to content moderation in Nepali.

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.

This paper presents our system for the CHiPSAL 2026 shared task on multimodal hate speech and sentiment detection in Nepali memes. We address both subtasks: binary hate speech classification and three-class sentiment analysis. Our approach adapts the Robust Adaptation of Hateful Meme Detection (RA-HMD) framework using Qwen3-VL-8B-Instruct, a state-of-the-art vision-language model with native Devanagari support. We employ a two-stage training pipeline: (1) LoRA fine-tuning with an MLP projection head for generative classification, and (2) contrastive backbone fine-tuning with supervised InfoNCE loss. We handle class imbalance through minority oversampling, image augmentation, and focal loss. At inference, we ensemble Stage 1 token probabilities with Stage 2 classifier scores using validation-tuned weights. Our end-to-end approach eliminates error propagation from separate OCR and translation pipelines by leveraging the model’s native Devanagari understanding. Our system achieved 2nd place on hate speech detection (F1: 0.797) and 4th place on sentiment analysis (F1: 0.518). We provide detailed ablations, error analysis, and insights into adapting large vision-language models for low-resource South Asian languages.

With the proliferation of multimodal content on various social media platforms, automated hate speech detection has emerged as a challenge, especially in meme-based communication, where meaning arises from interactions between text and images. In these situations, unimodal techniques are inadequate in capturing semantics. In order to address such issues, a late-fusion-based multimodal hate speech detection framework has been proposed and implemented for the CHiPSAL shared task. In the proposed framework, multimodal content is processed by utilizing XLM-RoBERTa for multilingual text representation and a Vision Transformer (ViT) for visual representation. Both modal representations are fused using a fully connected classification head and are used for binary hate speech detection. The findings suggest that multimodal content effectively captures features from individual modalities and helps improve hate speech detection accuracy by obtaining a Macro F1-score of 0.66 and ranking 5th on the leaderboard. Also, transformer-based multimodal fusion performs effectively and acts as a reliable baseline for hate speech detection in low-resource multilingual meme-based communication scenarios.

Memes serve as a method to express feelings such as humor, sarcasm, and diverse viewpoints. The task of identifying sentiment in memes is becoming increasingly complex, particularly in low-resource languages like Nepali where memes often combine images, texts, and code-mixed language. However, multimodal methods for sentiment analysis in Nepali memes seem to be insufficient. In this paper, we present our system for the Subtask B(Sentiment Analysis) for Shared Task on Multimodal Hate and Sentiment Understanding in Low-Resource Memes@CHiPSAL LREC 2026. We implement various unimodal models, such as XLM-RoBERTa-large,MuRIL-base, Twitter-XLM-R for text. Moreover, we incorporate BLIP-2 captions to enhance visual-text understanding and adopted a multimodal approach that fuses textual embeddings, image embeddings, caption embeddings, and similarity scores. The fused features process through cross-attention and a dense neural network for classification, with focal loss and class weighting used to improve performance. Our approach achieved a macro F1 score of 0.50 securing 7th place and highlighting the importance of cross-modal interaction and large-scale pretrained vision-language models for robust meme understanding in sentiment analysis.

Hate speech detection in Devanagari-scripted social media memes presents compounded challenges: multimodal content structure, script-specific linguistic complexity, and extreme data scarcity in low-resource settings. This paper presents our system for the CHiPSAL 2026 shared task, addressing both Subtask A (binary hate speech detection) and Subtask B (three-class sentiment classification: positive, neutral, negative). We propose a hybrid cross-modal attention fusion architecture that combines CLIP (ViT-B/32) for visual encoding with BGE-M3 for multilingual text representation, connected through 4-head self-attention and a learnable gating network that dynamically weights modality contributions on a per-sample basis. Systematic evaluation across eight model configurations demonstrates that explicit cross-modal reasoning achieves a 5.9% F1-macro improvement over text-only baselines on Subtask A, while uncovering two unexpected but critical findings: English-centric vision models exhibit near-random performance on Devanagari script, and standard ensemble methods catastrophically degrade under data scarcity (N ≈ 850 per fold) due to correlated overfitting. Code and implementation details are available at a repository that has been anonymized for the review process and will be fully disclosed in the final version

This paper presents a system for hate speech detection in low-resource Nepali memes, submitted as part of Subtask A of the Shared Task on Multimodal Understanding at CHiPSAL 2026. Detecting hateful memes is particularly challenging due to the combination of images, text, and emojis used to portray humor, satire, or sociopolitical commentary, as well as the low-resource nature of the Nepali language. We investigate a range of unimodal and multimodal modeling strategies, including text-only, vision-text, and caption-based approaches. For caption generation, the Gemini family of models (Gemini 2.X and Gemini 3.X) was used to produce contextually rich captions, which are publicly released as NeMeme-CAP on Hugging Face. Caption-based modeling leverages stochastic caption augmentation to address class imbalance and Test-Time Augmentation (TTA) to reduce prediction variance and improve model robustness. The best-performing system fine-tunes an encoder-only transformer model, RoBERTa-base, on the generated captions, achieving third place on the official leaderboard with a macro-averaged F1-score of 0.7397. The code is publicly available at https://github.com/thapaliya123/LREC-CHiPSAL-2026.

Memes have become a popular way for people to share opinions and emotions on social media, but they are also often used to spread hate and negative sentiments. In this paper, we present our multimodal approach to the CHiPSAL 2026 shared task on multimodal hate and sentiment detection in Nepali memes, which includes two subtasks: hate detection and sentiment analysis. Since memes usually combine both text and images, we first experimented with different unimodal models for text and images separately. After identifying the top two best-performing text and image models, combined them using different fusion techniques. The results show that multimodal models outperform unimodal ones, highlighting that both textual and visual information are important for understanding the context of memes. The multi- modal model, which combines sentence-transformers/LaBSE for text and ResNet-18 for image using weighted Fusion technique, achieved a macro F1 score of 0.6614 for Subtask A and sentence-Transformers/LaBSE for text and deit- Base for image using simple Fusion technique, achieved a macro F1 score of 0.4839 for SubTask B, on the test dataset.