Skip to main content

How Often Is This Word Actually Used?

A dictionary is deliberately democratic: a word attested once in a single late commentary gets an entry formatted exactly like ca or bhū. That is what makes dictionaries complete — and what makes them silent about usage. This page adds the missing signal: corpus attestation counts from Oliver Hellwig's Digital Corpus of Sanskrit (DCS, CC BY), a lemmatized, human-validated corpus. The feed used here counts 4,550,704 tokens over 59,282 lemmas.

Sanskrit is extremely top-heavy​

A small core of lemmas does most of the work in real texts:

Most frequent…Tokens coveredShare of the counted corpus
100 lemmas1,605,83635.3%
500 lemmas2,534,20255.7%
1,000 lemmas3,022,01566.4%
2,000 lemmas3,503,50877.0%

Of 4,550,704 tokens counted over 59,282 lemmas (Digital Corpus of Sanskrit, whole-corpus counts).

Two thousand lemmas — roughly one percent of Monier-Williams' 185,803 entries — account for over three quarters of all corpus tokens. For a learner this is the strongest prioritization signal there is: the MW reading guide teaches you how to read an entry; this layer tells you which entries you will actually meet.

The most frequent lemmas​

RankLemmaSLP1 keyPOSTokensPer 10,000 tokens
1cacaind155,088340.8
2tadtadpron151,248332.4
3nanaind53,981118.6
4madmadpron49,367108.5
5evaevaind45,884100.8
6yadyadpron45,46699.9
7itiitiind44,37997.5
8tvadtvadpron37,26381.9
9idamidampron36,78780.8
10tutuind35,00476.9
11apiapiind34,02274.8
12kṛkf2.Ā.33,25173.1
13bhūBU1.Ā.32,61271.7
14sarvasarvapron31,37768.9
15vacvac2.P.30,97068.1
16mahatmahatadj24,95754.8
17asas2.P.24,56754.0
18etadetadpron24,23953.3
19tathātaTAind22,74150.0
20tatastatasind21,96048.3
21hihiind19,45242.7
22vāvAind19,24942.3
23rājanrAjanm17,08237.5
24ivaivaind16,24035.7
25gamgam6.Ā.15,95035.0

POS tags are the DCS grammar codes (ind indeclinable, pron pronoun, m/f/n noun by gender, adj adjective, 1.P.-style codes for verb roots by class and pada). Every lemma is keyed in SLP1, so it joins directly against Cologne headword keys.

Most headwords are corpus-rare​

Only 59,282 lemmas carry a whole-corpus count (83,277 appear in the source layer at all) — against 185,803 MW entries. The gap is not an error: large dictionaries carry enormous tails of words inherited from the indigenous lexicographic tradition (MW's "L." = lexicographers marker; see the SKD and VCP kośa pages), attested rarely or only in lexica. When a dictionary entry cites only lexicographers and the corpus count is zero, you are looking at a word of the dictionary tradition rather than of surviving usage — a distinction no dictionary layout shows you, but the corpus layer does.

Words with a history​

The feed carries a per-period vector for each lemma (DCS's chronological buckets — numbered slots ending at an approximate year, e.g. 1 -800 ≈ texts up to 800 BCE, 5 1200 ≈ up to 1200 CE; the labels 3200 and 4700 are slots 3 and 4 — up to 200 CE and 700 CE — with the space lost in the upstream export; plus the undatable genre buckets 9 Vedic, 11 Epic, 12 Classic). Frequency profiles differ dramatically:

LemmaRank9 Vedic1 -8002 -300320047005 12006 17007 190011 Epic12 Classic
ūti166749%49%0%0%·0%···1%
rayi187449%47%2%·0%0%··0%0%
vṛṣan125848%48%1%1%0%0%··1%0%
vaiśampāyana559···49%0%1%··49%1%
vaidehī16910%··49%1%0%0%·49%1%
pārada788···0%0%18%22%8%1%50%
gandhaka569···0%·20%28%2%0%51%
abhraka1199·····22%26%2%·50%

Share of each lemma's period-datable tokens per DCS period bucket (rows sum to 100%;  ·  = no tokens in that bucket).

  • ūti "help, favour", rayi "wealth", vṛṣan "bull, mighty" — Vedic workhorses that almost vanish afterwards. For these, Grassmann (a dedicated Rigveda dictionary with complete attestation) serves you better than any general-purpose dictionary.
  • Vaiśampāyana, Vaidehī — epic names concentrated in the 11 Epic bucket: Mahābhārata/Rāmāyaṇa reading is PWG and MW territory, both of which cite the epics massively (see what each dictionary quotes).
  • pārada "mercury", gandhaka "sulphur", abhraka "mica" — alchemical (rasaśāstra) vocabulary whose counts sit in the late buckets. Late technical Sanskrit is real Sanskrit too, and its vocabulary barely overlaps the poetic core.

The same lemma-level profiles power the corpus-era view in the dictionary landscape page's date reasoning: which dictionary to open depends on when your text was written.

Trust block
  • Evidence: src/data/corpus-frequency.json (top 2,000 lemmas by corpus rank + whole-corpus stats), regenerated by scripts/build-corpus-frequency.mjs from kosha/data/frequency/lemma_frequency.tsv (83,277 lemmas), itself built from the VisualDCS M9 archive of DCS data. Upstream: Digital Corpus of Sanskrit (Oliver Hellwig), CC BY. Every table and percentage on this page is computed from the committed feed at build time — nothing is hand-typed.
  • Limitations: the page's tables see the top-2,000 slice, not all 83,277 lemmas. DCS is genre-skewed (its later periods are dominated by śāstric and alchemical texts, which is why the "late-heavy" examples are rasaśāstra terms — a corpus fact, not a claim about all late Sanskrit). Period bucket boundaries are DCS-coded and approximate (the upstream QL layer README flags the label-to-boundary mapping as not verified cell-by-cell). Counts are lemmatized tokens from DCS's analyses — a different corpus would give different numbers.
  • Rights: DCS is CC BY; the feed is a derived aggregation with attribution (evidence record §1 in NON_COLOGNE_SOURCES.md).
  • Owner repos: DCS (upstream) → VisualDCS → kosha (frequency layer) → this repo (rendering).
  • Built by: Fable 5 (claude-fable-5), 07-07-2026, handoff H280.

See also​