பகுபத உறுப்பிலக்கணம் morphology

The same subject, two vocabularies

Morphology is the study of how a word is built out of smaller meaningful pieces. Tamil grammar has had a name for that study, and a six-part analysis to go with it, for centuries. A computational linguist and a தமிழ் teacher are looking at the same thing and calling it different names.

The mapping is worth stating plainly, because it is what lets classical grammar drive modern software: a morpheme is an உறுப்பு, a lemma is the அகராதி வடிவம், an affix is a விகுதி or இடைநிலை or சாரியை, and sandhi is புணர்ச்சி. The full mapping is the project's glossary, and it will get a page of its own here.

The verse the whole layer rests on

It names all six உறுப்பு and says that every பதம் is completed by joining whichever of the six it takes. That word, ஏற்பவை, is the part software gets wrong: a word takes the parts that apply to it, so not all six appear in every word, and a system that expects six slots will invent something to fill them.

Tholkappiyam does not enumerate these six, which is why Nannūl is primary here and the answer says so. That is the whole of the fallback rule in one worked case.

One word, taken apart

மரத்தில், "in the tree". One word in Tamil, three words in English, and three உறுப்பு inside it. This is the only word example on the site, and it is here as anatomy rather than as proof.

புணர்ச்சி: திரிதல் — ம் → த் at the சந்தி, where மரம் meets அத்து. authority: Tholkappiyam, எழுத்ததிகாரம், புணரியல்

Every label on that diagram carries the authority that names it, and the tool returns them the same way. Where the analyser cannot determine a join, it leaves it unnamed. An invented split is worse than no split, because a teacher would have to know the answer already to catch it.

Why a dictionary cannot do this

Tamil is agglutinative. Pieces chain onto a root, each one keeping its own identity, so one word carries what English needs a phrase for. One verb root yields thousands of surface forms, which means a fixed word list is out of date the moment somebody inflects something.

So the core of this project is a rule engine rather than a database. That is unusual for an MCP server, and it is forced by the language: Tamil words are generated, not listed.

ThamizhiMorph does the analysis, and we did not rebuild it

The rules run on ThamizhiMorph, a finite-state morphological analyser and generator for Tamil, published by Sarveswaran, Dias and Butt in Machine Translation 35:37–70 (2021) and released under Apache-2.0. We wrap it. Building a second one would have been a worse use of the same years.

Its own measured numbers, from the paper, on a 612-word textbook corpus:

Source: Sarveswaran, Dias & Butt 2021 · code at github.com/sarves/thamizhi-morph

The guesser FSTs are switched off, deliberately

ThamizhiMorph ships guesser transducers that will produce an analysis for a word the lexicon has never seen. We do not use them. On கொடு a guesser returns the lemma கொடுத், which is a confident error, and a confident error is the one failure mode this project exists to remove. A word the analyser genuinely does not know goes to the enrichment loop and comes back as an honest gap.

The same machine can build words, and that is the long road

A finite-state transducer runs in both directions. The one analysing மரத்தில் into three pieces can also take a lemma and a set of grammatical features and produce the correct surface form. We use one direction today.

The other direction is why a Tamil model becomes thinkable later: a system that knows what to say can hand over a lemma plus features and receive correct Tamil, instead of spelling out an inflected word it has only ever seen in fragments. That is the horizon, and it is written down rather than scheduled.