Spashta
Grammar-Aware Romanized Nepali Normalization: Rule-Bounded LLM Generation and Deterministic Verification
Project Originator, Architecture & Research LeadOverview
Spashta started as a hackathon project addressing a problem with Romanized Nepali writing. Romanized Nepali is highly inconsistent, and direct word-level conversion does not always preserve the grammatical context needed to generate natural Devanagari Nepali. The project explored whether an LLM could perform context-aware normalization while being constrained by deterministic rules so that generation would be more reliable. The project later won the hackathon and was developed further into a research study.
Research Question
Can an LLM normalize noisy, inconsistent Romanized Nepali into contextually correct Devanagari when its generation is bounded by deterministic linguistic rules and verified at the output, rather than relying on unrestricted generation?
My Contribution
- Originated the core project idea from the problem of reliably converting noisy Romanized Nepali into contextually correct Devanagari Nepali.
- Designed the initial concept of combining LLM generation with deterministic linguistic constraints and verification rather than unrestricted generation.
- Researched related approaches and references used to frame the system and the research problem.
- Developed the initial research framing, architecture, and experimental direction.
- Worked collaboratively with Rajan Raj Acharya, Krishma Paudel, and Satish Kumar Karna, who contributed to other parts of the system and research.
- Took a major role in structuring the manuscript and performed the final technical screening and consistency review before completion.
Methodology
- Formalized normalization as a constrained generation task in which an LLM proposes candidate Devanagari output.
- Introduced deterministic linguistic rules that bound the generation space to contextually consistent forms.
- Added a deterministic verification step to check candidate output against the rules before accepting it.
- Evaluated the rule-bounded pipeline against unrestricted LLM generation and word-level baselines on Romanized Nepali input.
Key Findings
- Bounding LLM generation with deterministic linguistic constraints produced more reliable, contextually consistent Devanagari than unrestricted generation.
- Deterministic verification helped catch and reject grammatically inconsistent outputs that a purely generative approach would emit.
- The constrained approach made normalization more reproducible and less dependent on the model's raw generation quality.
What I Learned
- How to make LLM output more reliable by constraining it with deterministic rules.
- The challenges of low-resource NLP and Romanized Nepali in particular.
- How to construct constrained-generation and verification pipelines in practice.
Limitations
- The rules cover a defined set of Romanized Nepali patterns and may not generalize to every dialect or exceedingly noisy input.
- The verification layer is limited to the rules it can encode and cannot catch every contextual nuance.
Paper / Code / Artifacts
Paper, code, and other artifacts are not linked yet.