Nepali Natural Language Processing (NLP) models are typically trained and evaluated on Standard Nepali, which can introduce bias against regional dialects. This study investigates the performance of Named Entity Recognition (NER) and Part-of-Speech (POS) tagging models on Achhami, a Far-Western dialect of Nepal. A parallel corpus of 300 sentence pairs was created, covering news, cultural topics, and everyday conversations. Achhami translations were produced by native speakers to preserve linguistic authenticity. The evaluation compared fine-tuned Transformer models with large language models using zero-shot prompting. Across both tasks, all models showed consistent performance degradation on the Achhami dialect. For NER, F1 scores decreased by 2.12 to 3.97 percent. Claude 3.5 Haiku achieved the best NER performance, while the monolingual NepBertA model unexpectedly outperformed multilingual alternatives, challenging assumptions about multilingual advantages. POS tagging results showed a similar pattern, with accuracy dropping notably on Achhami data. Large language models also showed comparable weaknesses, with accuracy reductions ranging from 2.9 to 7.0 percent. These findings quantify Kathmandu-centered bias in Nepali NLP and highlight the importance of dialectally diverse training data for building more inclusive and equitable language technologies.

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.

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.

Bias in Nepali NLP is rarely addressed, as the language is classified as low-resource, which leads to the perpetuation of biases in downstream systems. Our research focuses on gender bias in Nepali-English machine translation, an area that has seen little exploration. With the emergence of Large Language Models(LLM), there is a unique opportunity to mitigate these biases. In this study, we quantify and evaluate gender bias by constructing an occupation corpus and adapting three gender-bias challenge sets for Nepali. Our findings reveal that gender bias is prevalent in existing translation systems, with translations often reinforcing stereotypes and misrepresenting gender-specific roles. However, LLMs perform significantly better in both gender-neutral and gender-specific contexts, demonstrating less bias compared to traditional machine translation systems. Despite some quirks, LLMs offer a promising alternative for culture-rich, low-resource languages like Nepali. We also explore how LLMs can improve gender accuracy and mitigate biases in occupational terms, providing a more equitable translation experience. Our work contributes to the growing effort to reduce biases in machine translation and highlights the potential of LLMs to address bias in low-resource languages, paving the way for more inclusive and accurate translation systems.