A Quantitative Overview of Modern Tamil Lexicon and Morphology

How many words are there in the modern Tamil language? How many and what classes, sub-classes and varieties are there? What are the rules for generating all the inflected forms?

As any Tamil speaker knows, the language is extremely morphologically rich. A single root word can be modified using suffixes to produce a large number of inflected forms that contain layers of meaning expressing tense, number, case, person, voice, emphasis, etc. The noun for tree, மரம், can produce மரங்களிலிருந்து which means 'from the trees', or மரத்தால் which could mean 'because of/by the tree' and so on. Verbs can be chained together (கொண்டுவா, முடித்துவிடு). Various cases like accusative (மரத்தை), dative (மரத்துக்கு), locative (மரத்தில்), vocative (மரமே) and many more are expressed through suffixes, often with some modification of the root lemma as well.

As part of some experiments in training Tamil Language Models, with extensive use of ChatGPT 5.6 Sol, I attempted to build a full rule-based morphological model of the Tamil lexicon. Starting with word lists obtained from Tamil Wiktionary and the University of Madras Tamil Lexicon as well as other modern corpus sources, I built Finite State Transducer models that can parse any valid Tamil word into its semantic components as well as generate all valid inflected forms from a root word (lemma). These FSTs were built on top of a core set of FSTs published as Thamizhi-Morph by academics at University of Moratuwa, but with a greatly expanded, corrected and modified class structure and inflection rules as well as root lemma vocabulary.

Here is a quantitative overview of the resulting modern Tamil lexicon. Note that the coverage of these models is far from complete, especially on Classical Tamil. The models were built almost exclusively on modern (~last 100 years) Tamil sources. The models are certain to still contain errors and omissions.

How large is the modern Tamil lexicon?

The model generates 2,130,878,180 distinct written forms from the model's vocabulary and grammatical rules. This does not mean that Tamil has 2.131 billion ordinary dictionary words. Most of these are inflected, compound, alternative or context-dependent forms. Of this total, the noun, verb and other basic models together generate about 205.5 million forms, while the productive multi-verb model generates more than 1.9 billion.

The current model contains 139,895 root lemmas. The two largest inflectional groups are:

• 127,311 noun roots
• 11,399 verb roots or stems

These numbers overlap slightly because the same spelling can sometimes function as both a noun and a verb.

The noun vocabulary is much larger, but verbs produce richer families of forms. A noun mainly combines with number, case and particles. A verb can express tense, person, number, gender, mood, negation, voice and several kinds of participles. It can also combine with other verbs.

A grammatical analysis is the model's description of how a written word is built and what grammatical information it carries. For example,

• மரங்களை → மரம் + noun + plural + accusative
• படித்தான் → படி + verb + past + third-person singular masculine

Spelling and grammatical analysis do not have a one-to-one relationship. மரத்தால் can be read either as a noun form—மரம் + “by/with”—or as a conditional verb form from மர, depending on context. In the other direction, the single analysis சுவர் + noun + accusative can produce either சுவரை or சுவற்றை.

The billions of generated spellings reuse a much smaller collection of grammatical patterns, called semantic signatures, where the lemma is excluded (e.g. noun + plural +accusative, verb + past + third-person singular, etc)

The important point is that Tamil can generate billions of forms from fewer than 140,000 basic lexical entries. The size of the language lies not only in the number of roots, but in the productive rules that combine and modify them.

Noun Varieties

Tamil noun inflection attaches suffixes to root words while also often changing the root stem, with the precise rewrite-rule depending on the type (class/subclass/word-ending variety) of the root word.

The model requires 23 noun classes and subclasses to represent these different behaviours.

The noun written-form count is the output of the noun model by itself, before final-sandhi forms are removed.

The complete set of noun paths is shown below. “Class” here means a path in the finite-state model. Closely related paths are kept separate when they permit different written forms. As in the verb table below, the entry counts are class assignments rather than additional unique nouns: a spelling can have more than one lexical analysis and therefore occur in more than one class.

The ending of a noun is therefore useful evidence, but it does not determine the class by itself. The model also uses a 38-family ending audit. Each source lemma is assigned to its longest matching ending, so a word in -ியம் is counted there rather than again under the shorter -ம் family. “Coverage” means that the source spelling receives some analysis from the released FST system. The final column shows the noun classes of the explicitly listed roots; some additional source words are recognized as derived forms, compounds or other word types.

The long, specific families are usually the most predictable. Every audited -ட்டை, -ிப்பு, -ைப்பு, -ந்தை, -ச்சை, -ங்கை, -பூ and -வியல் spelling is recognized. Short endings are less informative because unrelated words can share them: coverage is about 75% for -ி and 73% for short -ு, and each can contain several lexical types.

Verb Varieties

The 11,399 verb roots and stems are divided into 24 verb classes and subclasses.

The multi-verb row counts productive combinations rather than additional dictionary roots. These individual-model figures are before final-sandhi removal, and they overlap with one another, so they should not be added to obtain the complete-model total.

The classes are based on the form taken by a root before tense and person endings. The table gives masculine third-person singular forms so that the stem changes can be compared directly. The entry counts are class assignments, not additional unique verbs: a spelling can have more than one lexical analysis, and the four modern irregular paths overlap older numbered paths where appropriate.

Verb endings are less predictive than the longer noun-ending families. For example, உ-ending verbs occur in classes 1, 5, 6.1, 6.2 and 6.3; இ-ending verbs split between the weak class 4 and strong class 11; ள்-ending verbs occur in classes 2 and 9; and ல்-ending verbs occur in classes 3 and 10. The lexical class is therefore needed even when the final letter is known.

Tamil verbs are especially productive because verbs can also combine with other verbs:

• படித்துக்கொடுத்தான்
• செய்துகொண்டிருக்கிறான்
• முடித்துவிட்டான்

The model uses 20 common auxiliary or light verbs and 70 continuation patterns to represent such verb combinations. Before tense and person endings are even added, it generates nearly 4.9 million possible root-and-auxiliary combinations.

Productive verb chaining is responsible for most of the enormous number of generated Tamil forms. As mentioned before, the noun, verb and other basic models together generate about 205.5 million forms, while the productive multi-verb model generates more than 1.9 billion.

Adjectives, Adverbs, Pronouns and Other Word Types

The model also has separate inventories for adjectives, adverbs, pronouns, particles and other function words.

The written-form counts in this table are the outputs of the individual models before final-sandhi forms are removed. They should not be added to the overall total because the same spelling can appear in more than one model.

Adjectives

The standalone adjective model contains:

• 648 bases with adjective-leading readings, including words such as நல்ல, பெரிய, சிறிய and புதிய.
• 14 other bases whose readings begin with a demonstrative, adverb, intensifier or particle.
• No noun-leading paths. Noun-derived forms such as வலிமை → வலிமையான and மாசு → மாசற்ற are owned by the noun model.

These groups produce 662 distinct bases, 3,081 grammatical paths and 3,080 distinct written forms in the standalone adjective model.

This standalone adjective count is not the main measure of adjective-like coverage. Tamil productively forms modifiers from nouns, so the model places the general rule in the noun classes that already know how each stem changes. Seventeen eligible singular noun routes and twenty plural routes form reversible copular expressions in -ஆன, privatives in -அற்ற, and their nominalized forms in -ஆனது, -அற்றது, -ஆனவை and -அற்றவை. The 148 class-level continuations add about 1,011,524 adjective-like grammatical paths to the noun model.

This division is important. A word such as நல்ல is a lexical adjective and belongs in the adjective model. A form such as மரமற்ற is built productively from the noun மரம் and belongs to the noun model.

Examples of adjective-like forms are:

• நல்ல → நல்ல + adjective
• வலிமையான → வலிமை + ஆன + adjective
• மாசற்ற → மாசு + அற்ற + adjective
• மரமற்ற → மரம் + noun + அற்ற + adjectival participle + adjective
• மரங்களான → மரம் + noun + plural + ஆன + adjectival participle

Adverbs

The adverb model contains 771 bases with direct adverb-leading readings. When demonstrative, noun, adjective, pronoun and verb bases that also produce adverb readings are included, the model contains 932 distinct bases. Of these, 130 have noun-based readings and 15 have adjective-based readings.

It represents 4,592 grammatical paths and 4,584 distinct written forms. The inventory includes:

• ordinary lexical adverbs: மிகவும், தவறாமல்
• compositional manner forms: முதல்முறையாக, பிஸியாக
• distributive forms: மாதந்தோறும், வீடுதோறும்
• time, place and situation words: இப்போது, எப்போது
• forms with emphasis, addition or questioning: இப்போதுதான், இப்போதும், எப்போதோ

For example:

• இப்போது → this time / now
• இப்போதுதான் → இப்போது + emphasis
• இப்போதும் → இப்போது + “also/even”
• எப்போதோ → எப்போது + question or indefinite ending
• முதல்முறையாக → முதல்முறை + noun + translative + adverb
• மாதந்தோறும் → மாதம் + noun + time + distributive + adverb

The model adds adverbs carefully. It does not assume that every word ending in -ஆக is an adverb. For example, words describing someone’s role (விமானியாக), expressions meaning “for” or “for the purpose of,” (-ற்காக) and certain verb-related forms (-தாக) are identified separately.

Pronouns

The pronoun model begins with 44 base forms and generates 1,723 grammatical analyses representing 1,594 distinct written forms.

The base forms include overlapping groups:

These figures overlap because one base can carry more than one property. The demonstrative system itself distinguishes 6 near-speaker forms, 6 distant forms and 7 medial forms.

Pronouns combine with case and number much like nouns, while also retaining information about person, distance, inclusiveness and respect:

• நான் → nominative
• என்னை → accusative
• என்னால் → instrumental
• எனக்கு → dative
• என்னுடைய → genitive
• நாம் / நாங்கள் → inclusive and exclusive first-person plural
• இவர் / அவர்கள் → demonstrative and respectful or plural forms
• எது / எவை → interrogative singular and plural forms

The wider system also covers productive forms such as இவர்களிடமிருந்து and எதுவாகும்.

Particles and Other Function Words

The remaining function-word model contains 414 distinct bases, 2,568 grammatical analyses and 2,471 distinct written forms.

Its main groups are:

These rows overlap because a spelling can have more than one function. For example, ஆம் can be analyzed as a conjunction, a particle or an affirmative response.

Representative examples include:

• ஆனால் — conjunction
• மட்டும் — particle or postposition
• எத்தனை — interrogative quantifier
• ஆயிரம் — cardinal number, with case and other extensions
• ஐந்தாவது — ordinal number
• ஆல் — case marker

What do these numbers tell us?

The main lesson is that Tamil does not have a simple boundary between “word” and “grammar.”

The model contains about 140,000 basic lexical entries, but these roots participate in tens of thousands of reusable grammatical patterns and more than 2.13 billion written forms.

It identifies:

• 23 noun classes and subclasses
• 24 verb classes and subclasses
• 38 major word-ending families
• 20 productive auxiliary verbs
• 46,846 distinct grammatical patterns, or 36,058 after patterns containing final-sandhi markers are excluded

These exact numbers belong to this model and may change as the vocabulary and rules improve. Another grammatical analysis might combine some classes or divide others more finely.

But the broad picture is clear. Tamil achieves an enormous vocabulary of possible forms using a relatively compact set of root words and highly productive grammatical rules. A long Tamil word is often not an indivisible dictionary item. It is a structured combination of a root, several layers of grammar and sometimes more than one verb.