Who We Are
Nisansa de Silva’s Research Group conducts research in natural language processing with a special focus on low-resource languages such as Sinhala and Tamil, data quality in multilingual corpora, and applied machine learning, based at the Department of Computer Science & Engineering, University of Moratuwa. The group was originally established in 2011, and went on semi-hiatus while Dr. de Silva was at the University of Oregon (2014–2020). It was formally re-established in 2021 upon his return to UoM.
Faculty and students explore multiple aspects of language technology, including lexical semantics, sentiment and subjectivity analysis, graph-based NLP, and multilingual text processing. Their work has appeared at major international venues such as ACL, EACL, WMT, and LoResMT, with ongoing collaborations bridging local linguistic challenges and global research frontiers.
The group has received research funding from Google and the University of Moratuwa, while also making an impact through the release of open-source tools, surveys, and models to support the wider Sinhala and Tamil NLP communities.
The group publishes at venues ranked in Google Scholar's top 20 for Computational Linguistics, including ACL, EMNLP, EACL, TACL, COLING, LREC, SemEval, WMT, and AACL-IJCNLP. A full list is on the Publications page.
Research Programme
Nine interconnected research threads, each actively pursued. Individual projects and papers are listed on the Projects and Publications pages.
Sinhala & Tamil Resources
Building the raw materials everything else depends on: corpora, lexicons, dictionaries, OCR, spelling correction, and diachronic text spanning 2,500 years.
Wahara trilingual dictionary · First free English–Sinhala word dictionary (546k pairs) · SiDiaC diachronic corpus (LREC 2026) · Sinhala WordNetMultilingual NMT
Sinhala–Tamil–English neural machine translation built on quality-first parallel data: auditing, denoising, augmentation, and systematic domain benchmarking.
EnSiTa — 277k+ sentence pairs, 10 domains, 110+ professional translators over 3 years. EACL 2024 Low Resource Paper Award.Semantic Oppositeness & Contradiction
Teaching models to distinguish contradictory claims — the subject of Dr de Silva's PhD — applied to rumour detection (EACL 2021) and 36,877 PubMed abstracts.
PhD dissertation: "Semantic Oppositeness for Inconsistency and Disagreement Detection" (Oregon, 2020).Sentiment & Social Media NLP
Aspect-based sentiment analysis going beyond positive/negative to per-aspect sentiment, plus processing of noisy, code-mixed Sinhala social text.
SalAngaBhava Sinhala market-sentiment dataset · Instruct-DeBERTa hybrid ABSA system (ICTer 2025).SigmaLaw: Legal NLP
A decade-long programme: party identification, legal sentiment, semantic retrieval, deception detection, and low-resource legal information access.
Section-weighted case retrieval (ICNLP 2026) · LawChain (ICML 2026) · "Semantics of Subterfuge" deception benchmark (ICAIL 2026).LLM Cultural Alignment
Ensuring LLMs reason appropriately in Sri Lankan cultural and social contexts — commonsense reasoning in Sinhala, and value-aligned instruction tuning.
LKValues: 205-respondent survey → 40 Sri Lankan values → 150k-instance instruction corpus. Funded by the Rs 4.5M SCALLM grant (2026–2029).LLM Multi-Agent Systems
Coordinating LLM agents for cooperation, reasoning, and real task execution — from game-theoretic simulations to a working calendar assistant.
"Evolution of Cooperation" (COINGAS 2026) — first re-run of Boyd & Richerson's 1992 cooperation model with language-model agents.Sinhala Foundation Models
Building Sinhala its own large-model infrastructure rather than relying on incidental multilingual coverage of larger models.
SinLlama — Sri Lanka's first Sinhala-native LLM, leads on 3 Sinhala benchmark tasks (10.7M training sentences, 304M tokens). HelaBERT encoder in development.Applied NLP & Sri Lankan Data
NLP applied beyond language itself: D&D narrative generation, AI for software maintenance, multi-document summarisation, and Sri Lankan institutional data.
JavaBackports benchmark (MSR 2026) · Trilingual parliamentary debate topic modelling · M2DS multilingual summarisation dataset.How We Work: Research Training
A defining feature of the group is that undergraduate final-year projects, MSc theses, and research assistant work feed directly into datasets, benchmarks, and published papers — rather than remaining isolated coursework.
See current and past members on the Team page, and open positions on the Join Us page.
Strategic Priorities
Where the group is investing effort for the next generation of Sri Lankan NLP infrastructure.
- Diachronic corpora & OCR
- Lexicons & ABSA datasets
- Spell checking & reusable benchmarks
- Sinhala & Tamil language resources
- Sri Lankan values & commonsense
- Safety & instruction tuning
- Local-language robustness
- Culturally grounded LLMs
- Legal reasoning & deception detection
- Evidence retrieval
- Interpretable evaluation
- Domain-sensitive NLP
- Parallel-data quality & denoising
- Domain-aware NMT
- Cross-lingual representations
- Efficient transfer for low-resource pairs
Research is supported by external grants from Google, the University of Moratuwa Senate Research Committee, and OpenToken. Current and past grants are listed on the Grants page.