A knowledge graph that shows its working

Suffeffix means suffix — and every effing affix.

Take suffix and open it between its stem and its ending: suff·ix. Drop eff into the middle — the spoken name of the letter F, and the polite way of writing the word English speakers wedge into the middle of other words for emphasis. The result, suff·eff·ix, is a suffix with an infix inside it: a small joke about the very thing this project catalogues.

English really does this. Expletive infixation (abso-bloody-lutely, fan-effing-tastic) is a documented morphological process, and McCarthy (1982) showed that where the infix lands follows the stress pattern of the host word. Our name is a playful cousin of it, not an instance: it infixes the bare eff, without the -ing, and it does not honour the stress rule. We never claimed it was grammatical.

The first half of the name says suffix; the second half is the real scope — every effing affix. Suffixes are only the commonest kind. Of the 151 affixes in v0.2, 128 are suffixes, 9 are prefixes (un-, in-, निर्-, నిర్-, சிறு-) and 14 are compound elements (such as -शाला, -శాస్త్రం and -இயல்) — 23 that the name, taken literally, leaves out. The schema has room for circumfixes too, and nothing in it assumes an affix belongs at the end of a word.

A coinage, not a finding: this is the project’s own word-play and no claim from a reference work. McCarthy, J. J. (1982). Prosodic structure and expletive infixation. Language 58: 574–590.

Telugu, Hindi, English, French, Tamil. Five languages, two families — and between the families, only theft.

The five are not a sample of convenience. Each grouping isolates one thing a lexical graph has to get right.

Telugu and Tamil — one family, two answers to Sanskrit.

Both are Dravidian and share inherited words (పేరు/பெயர் ‘name’, కన్ను/கண் ‘eye’, మూడు/மூன்று ‘three’), which the graph stores as cognates. But Telugu absorbs Sanskrit, while modern Tamil often coins from native roots: ‘library’ is గ్రంథాలయం in Telugu and நூலகம், from நூல் ‘book’, in Tamil. In this dataset 46% of Telugu words are Sanskritic, against 18% of Tamil words.

English and French — one family, one long loan.

Germanic English took much of its everyday and learned vocabulary from French after 1066: joy, beauty and merchant carry borrowing edges from Old French. Both keep a learned Latin and Greek layer (register L): 30% of the English words here and 28% of the French ones.

English, French and Hindi — distant cousins.

Three branches of Indo-European that parted thousands of years ago still share inherited roots: mother/mère/माता, name/nom/नाम, three/trois/तीन. These are cognates, each with a source.

Sanskrit and Latin — two classical donors.

Sanskrit is to Hindi and Telugu roughly what Latin is to French and English: a prestige source of learned words. Indian grammar calls a word borrowed unchanged tatsama (‘same as that’) and one worn down by inheritance tadbhava (‘born of that’). French shows the same split in its doublets: learned -ité (humanité) beside inherited -té (bonté), both from Latin -itātem. Suffeffix records both kinds of layering in one register field.

Why “theft”. Linguists say borrowing, but nothing is ever given back. Within a family, words are inherited; across the family line they can only be taken. That is the framework’s one hard rule, and the validator enforces it: none of the 127 etymology edges claims inheritance or cognacy across the boundary, while 14 borrowing edges cross it — Sanskrit pustaka into Tamil புத்தகம் and Telugu పుస్తకం among them.

Engineering The choice of five is a design decision, not a claim about the languages. The percentages describe the 564 words in this dataset, whose meanings were chosen to showcase derivation, not the vocabularies as a whole.

Telugu and Tamil are Dravidian. English (Germanic), French (Romance) and Hindi (Indo-Aryan) are Indo-European. There are no cognates across that line — only borrowing links the families, mostly through Sanskrit, Perso-Arabic vocabulary, and English. The validator rejects any cognate or inheritance edge that crosses the boundary, so the rule is enforced by the build, not promised in prose.

The Indo-European three genuinely are related, so name/nom/नाम and mother/mère/माता are stored as cognates, as are the Dravidian pairs నీరు/நீர் ‘water’ and ఊరు/ஊர் ‘village’. Each carries a source.

Every piece of linguistic content carries one of four tags, so a citable fact is never confused with a convenience of the implementation:

Established

standard linguistic knowledge, citable from reference works

Engineering

an abstraction chosen for implementation convenience — not a linguistic claim

Hypothesis

plausible and testable, not yet demonstrated

Future

out of scope for this release

Every etymology link also carries one of two statuses. Where the reference works agree, it is stored ✓ accepted. Where the literature disputes it, it is stored † contested and never silently resolved: the claim stays, with a low confidence and this badge, instead of a winner being picked for you.

Semantic atoms are an engineering interlingua, not a theory of human cognition. Atoms seeded from NSM primes are marked established as members of that inventory; the rest are engineering additions, and the total is capped at fifty.

564 concept-aligned entries, authored for this project against the reference works in its bibliography: the OED; the TLFi; Turner’s CDIAL; Burrow and Emeneau’s DEDR; the Madras Tamil Lexicon; Monier-Williams; Platts; Brown; McGregor; Krishnamurti; Watkins. Nearly everything ships as draft — authored carefully, not yet reviewed item by item. Reference-work entry numbers are deliberately absent until someone checks them against the printed volumes.

Suffeffix contains no language-model inference: no similarity search, no retrieval, no generation at read time. Explanations are assembled by named template rules from graph facts, which is why every sentence can show the nodes, rules and sources behind it.

Code is Apache-2.0; the dataset is CC BY-SA 4.0. Cite via CITATION.cff, or read the decision log.