TamilSandhi: A Neuro-Symbolic AI Toolkit for Correcting Sandhi Errors in Tamil

TY - GEN

T1 - TamilSandhi

T2 - 45th SGAI International Conference on Artificial Intelligence, AI 2025

AU - Yazhmozhi, V. M.

AU - Waller, Annalu

AU - Visser, Jacky

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

PY - 2026

Y1 - 2026

N2 - Tamil is an agglutinative and highly inflectional language with rich morphology. Sandhi errors are a class of spelling mistakes in Tamil that occur at word boundaries due to incorrect insertion or omission of hard consonants. This paper presents TamilSandhi, an open-source neuro-symbolic framework for identifying and correcting such errors. It combines rule-based logic for simple Sandhi rules with neural models for one complex rule. The rule-based system implements 12 Vallinam (hard consonant) addition rules and 8 deletion rules, derived from classical Tamil grammar texts such as Tolkappiyam and Nannool. These rules were packaged into a PyPi library (https://pypi.org/project/tamilsandhi-toolkit/) and validated using a suite of 300 unit tests (235 for addition, 65 for deletion). To address a complex rule not covered by the rule-based logic, a neural sequence-to-sequence approach was used. A supervised corpus of 10,434 manually annotated sentence pairs, each consisting of an incorrect and corrected version, was created. Three multilingual transformer models—mBART, mT5, and NLLB—were fine-tuned. Among them, mBART achieved the best performance, with a BLEU score of 99.9 and exact match accuracy of 97.9%.TamilSandhi is released as an open-source project on GitHub (https://github.com/TamilGeekGirl/TamilSandhiNeuroSymbolicAI/). Due to its modularity, reproducibility, and linguistic validity, TamilSandhi constitutes a significant contribution to NLP research in a widely used but low-resource language.

AB - Tamil is an agglutinative and highly inflectional language with rich morphology. Sandhi errors are a class of spelling mistakes in Tamil that occur at word boundaries due to incorrect insertion or omission of hard consonants. This paper presents TamilSandhi, an open-source neuro-symbolic framework for identifying and correcting such errors. It combines rule-based logic for simple Sandhi rules with neural models for one complex rule. The rule-based system implements 12 Vallinam (hard consonant) addition rules and 8 deletion rules, derived from classical Tamil grammar texts such as Tolkappiyam and Nannool. These rules were packaged into a PyPi library (https://pypi.org/project/tamilsandhi-toolkit/) and validated using a suite of 300 unit tests (235 for addition, 65 for deletion). To address a complex rule not covered by the rule-based logic, a neural sequence-to-sequence approach was used. A supervised corpus of 10,434 manually annotated sentence pairs, each consisting of an incorrect and corrected version, was created. Three multilingual transformer models—mBART, mT5, and NLLB—were fine-tuned. Among them, mBART achieved the best performance, with a BLEU score of 99.9 and exact match accuracy of 97.9%.TamilSandhi is released as an open-source project on GitHub (https://github.com/TamilGeekGirl/TamilSandhiNeuroSymbolicAI/). Due to its modularity, reproducibility, and linguistic validity, TamilSandhi constitutes a significant contribution to NLP research in a widely used but low-resource language.

KW - Neuro-symbolic AI

KW - NLG

KW - Sandhi errors

KW - Tamil

UR - https://www.scopus.com/pages/publications/105023502359

U2 - 10.1007/978-3-032-11402-0_18

DO - 10.1007/978-3-032-11402-0_18

M3 - Conference contribution

AN - SCOPUS:105023502359

SN - 9783032114013

T3 - Lecture Notes in Computer Science

SP - 243

EP - 256

BT - Artificial Intelligence XLII - 45th SGAI International Conference on Artificial Intelligence, AI 2025, Proceedings

A2 - Bramer, Max

A2 - Stahl, Frederic

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 16 December 2025 through 18 December 2025

ER -