Solthiruthi – A Multi-pass spell checker for Tamil language (Draft-1)

Author: Muthiah Annamalai,  edited: April, 21, 2015

Introduction

Following compiler construction theory, we may write a multi-pass spell checker for Tamil language. Tamil falls under a linguistic classification  as a agglutinative language, and has a phonetic script.

In this article I will present a conceptual framework, a proposal, for writing a multi-pass solthiruthi spell-checker. This effort is variation on theme of other spell checkers for Tamil created by projects myspell, hunspell, navi/vani, and languagetool respectively.

N.B.: I invite interested individuals to contribute to this document, and collaborate with me, under usual open-source credit/attribution terms.

Problem

1. Given a Tamil text, urai, we want to find out all spelling mistakes in the text
2. For words which are in error, we want to further provide alternatives
3. For Tamil language we have to apply santhi rules, puranchi vithigal, from grammar.
4. We also would like to provide context sensitive help like A.I.

Solution Overview

I propose a multi-pass spell checker comprised of N-levels. Each level upto the penunltimate (Level N-1) will continue to refine the words which are found in error; we call this analysis levels. The final level will form the suggestions for the words in error; we call this synthesis level. Such an algorithm can be built from 2-Level and upwards in the complexity.

We use a dictionary of valid Tamil words, to validate the Incoming text is split into the form of words, and each word goes through the N-steps.

Typical levels for analysis include, but not limited to,

1. Dictionary based white-list analysis for incorrectly spelled words
2. Morphological analysis via stemmer – this is a requirement of agglutinative languages
3. Clean up proper nouns words, peyarechcham,  using names of places, people, things
4. Bayesian classification of errors using bigram, trigram, N-gram letter probabilities
5. Context sensitive word suggestion by Bayesian classification using N-gram word probabilities
6. Keyboard models, Bamini, Tamil99, Anjal, layouts and common misspellings in these
7. Edit-distance based analysis for mistyped words
8. Santhi checker
9. Punarchi checker
10. Detect transliterated English words in Tamil and ignore

Synthesis step is usually made from the following,

1.  Norvig algorithm for spelling suggestions modified for Tamil arichuvadi
2.  For Tamil we need to combine Santhi, and Punarchi rule based pruning of suggestions

Pizhai : 2-Level spell checker

Pizhai spell checker algorithm uses the 2 analysis steps of whitelisted dictionary and the bigram/trigram letter analysis to flag erroneous words. Finally we suggest word alternates to the user during synthesis.

Solthiruthi : 3-Level spell checker

Solthiruthi spell checker algorithm uses the 3 analysis steps of whitelisted dictionary and the bigram/trigram letter analysis to flag erroneous words, and includes a stemmer. Finally we suggest word alternates to the user during synthesis.

Code

All of these ideas can be implemented in Python using open-tamil project, today. We welcome enthusiastic contributors – linguists, language enthusiasts and developers.