Reading WordNet synsets

Treat word meaning as a set, not a spelling

similarity
WordNet
word senses
Learn how WordNet groups words into synsets, why one word can cover many senses, and where a curated lexical resource helps or fails.

A museum cataloguer labels a drawer bank and then hesitates. The drawer might hold river photographs, money records, or a toy building with a coin slot.

The label is a single string. The meaning depends on the sense attached to it.

A synset is a set of words that share one sense. WordNet is a hand-built lexical database that records those sets and links them to broader meanings such as hypernyms, the “is-a” parents in a hierarchy.

TipWhat you will learn

This lesson shows how to:

  • define a synset in plain language;
  • count senses for a word string;
  • compare a sense count with a simple chance baseline;
  • list synonyms inside one synset;
  • walk hypernym links up an is-a hierarchy; and
  • name what WordNet can and cannot support.

Load the WordNet extract

The lesson reads four local CSV files made from Open English WordNet 2024. No code downloads WordNet while the page renders. The metadata file records the source, licence, and SHA-256 fingerprints.

library(readr)
library(dplyr)
library(tibble)
library(stringr)
library(knitr)

source("R/permutation-null.R")

senses <- read_csv(
  "data/wordnet/wordnet-senses.csv",
  na = character(),
  col_types = cols(
    lemma = col_character(),
    part_of_speech = col_character(),
    sense_id = col_character(),
    synset_id = col_character(),
    sense_order = col_integer()
  )
)

synsets <- read_csv(
  "data/wordnet/wordnet-synsets.csv",
  na = character(),
  col_types = cols(
    synset_id = col_character(),
    part_of_speech = col_character(),
    definition = col_character(),
    example = col_character()
  )
)

members <- read_csv(
  "data/wordnet/wordnet-members.csv",
  na = character(),
  col_types = cols(
    synset_id = col_character(),
    lemma = col_character()
  )
)

relations <- read_csv(
  "data/wordnet/wordnet-relations.csv",
  na = character(),
  col_types = cols(
    synset_id = col_character(),
    relation = col_character(),
    target_synset_id = col_character(),
    target_definition = col_character(),
    target_members = col_character()
  )
)

metadata <- read_csv(
  "data/wordnet/wordnet-metadata.csv",
  na = character(),
  col_types = cols(
    field = col_character(),
    value = col_character()
  )
)

metadata_summary <- tibble(
  item = c("lemmas", "senses", "synsets", "synset membership rows", "relations"),
  value = c(
    n_distinct(senses$lemma),
    nrow(senses),
    nrow(synsets),
    nrow(members),
    nrow(relations)
  )
)

kable(
  metadata_summary,
  col.names = c("Item", "Count"),
  caption = "Size of the local WordNet extract used in this lesson",
  row.names = FALSE
)
Size of the local WordNet extract used in this lesson
Item Count
lemmas 38
senses 419
synsets 417
synset membership rows 1050
relations 2432

This slice covers 38 lemmas, not all of English. It is enough to show how a lexical database separates spelling from meaning.

Read the senses of one word

A sense is one listed meaning of a word. The string bank has 18 senses in this extract: 10 noun senses and 8 verb senses. The first noun senses split river land, finance, piles, rows, reserves, and gambling funds.

bank_senses <- senses |>
  filter(lemma == "bank") |>
  count(part_of_speech, name = "senses") |>
  arrange(part_of_speech)

bank_noun_senses <- senses |>
  filter(lemma == "bank", part_of_speech == "noun") |>
  arrange(sense_order) |>
  left_join(synsets, by = c("synset_id", "part_of_speech")) |>
  slice_head(n = 6) |>
  transmute(
    sense = sense_order,
    definition
  )

lemma_sense_counts <- senses |>
  count(lemma, name = "sense_count")

bank_sense_null <- permutation_null(
  observed = sum(senses$lemma == "bank"),
  replicate_fn = function() {
    lemma_sense_counts$sense_count[
      match(sample(lemma_sense_counts$lemma, 1L), lemma_sense_counts$lemma)
    ]
  },
  replicates = 1000L,
  seed = 5801L,
  alternative = "greater"
)

kable(
  bank_noun_senses,
  col.names = c("Noun sense", "Definition"),
  caption = "The first six noun senses of bank in the local WordNet extract",
  row.names = FALSE
)
The first six noun senses of bank in the local WordNet extract
Noun sense Definition
1 sloping land (especially the slope beside a body of water)
2 a financial institution that accepts deposits and channels the money into lending activities
3 a long ridge or pile
4 an arrangement of similar objects in a row or in tiers
5 a supply or stock held in reserve for future use (especially in emergencies)
6 the funds held by a gambling house or the dealer in some gambling games 
kable(
  bank_sense_null |>
    mutate(across(where(is.numeric), ~ round(.x, 3))),
  col.names = c("Observed", "Null mean", "Null 5th pct", "Null 95th pct", "p-value", "Replicates", "Alternative"),
  caption = "Sense count for bank compared with random lemmas in the extract",
  row.names = FALSE
)
Sense count for bank compared with random lemmas in the extract
Observed Null mean Null 5th pct Null 95th pct p-value Replicates Alternative
18 11.352 2 48 0.138 1000 greater

A word string is not a unit of meaning. If a search, model, or count treats every bank as the same thing, it has made a simplifying choice. The null table adds a second caution: this extract was built from ambiguous teaching examples, so 18 senses for bank is not unusual inside this small file.

Any lexical method will return an answer for a known word. The safer habit is to ask what answer appears after the target word is compared with other words drawn from the same source.

Look inside a synset

For the financial-institution sense, the synset contains four member lemmas. These are four names WordNet links to the same sense.

financial_bank_synset <- senses |>
  filter(lemma == "bank", part_of_speech == "noun", sense_order == 2L) |>
  pull(synset_id)

financial_bank_members <- members |>
  filter(synset_id == financial_bank_synset) |>
  arrange(lemma) |>
  mutate(role = if_else(lemma == "bank", "queried word", "other member"))

kable(
  financial_bank_members,
  col.names = c("Synset ID", "Lemma", "Role"),
  caption = "Members of the financial-institution synset for bank",
  row.names = FALSE
)
Members of the financial-institution synset for bank
Synset ID Lemma Role
oewn-08437235-n bank queried word
oewn-08437235-n banking company other member
oewn-08437235-n banking concern other member
oewn-08437235-n depository financial institution other member

Synonymy here is local to one sense. bank and banking company share the finance sense; river bank belongs elsewhere.

Walk an is-a chain

A hypernym says that one meaning is a kind of another meaning. In this extract, the professor-position sense of chair walks upward through a position synset, then through a work synset, then to activity.

first_hypernym <- function(synset) {
  next_row <- relations |>
    filter(synset_id == synset, relation == "hypernym") |>
    arrange(target_synset_id) |>
    slice_head(n = 1)

  if (nrow(next_row) != 1L) {
    stop("No hypernym recorded for synset: ", synset, call. = FALSE)
  }

  next_row
}

select_chair_synset <- function() {
  candidates <- senses |>
    filter(lemma == "chair", part_of_speech == "noun") |>
    left_join(synsets, by = c("synset_id", "part_of_speech")) |>
    filter(str_detect(definition, regex("professorship|professor", ignore_case = TRUE))) |>
    arrange(sense_order)

  if (nrow(candidates) != 1L) {
    stop("Expected one professor-position chair sense in this extract.", call. = FALSE)
  }

  candidates$synset_id[[1]]
}

walk_hypernyms <- function(start_synset, steps) {
  path <- start_synset

  for (step in seq_len(steps)) {
    path <- c(path, first_hypernym(path[length(path)])$target_synset_id)
  }

  path
}

lookup_definition <- function(id) {
  local_definition <- synsets$definition[synsets$synset_id == id]
  if (length(local_definition) > 0L) {
    return(local_definition[[1]])
  }

  external_definition <- relations$target_definition[match(id, relations$target_synset_id)]
  if (is.na(external_definition)) {
    stop("No definition available for synset: ", id, call. = FALSE)
  }

  external_definition
}

lookup_members <- function(id) {
  local_members <- members |>
    filter(synset_id == id) |>
    arrange(lemma) |>
    pull(lemma)

  if (length(local_members) > 0L) {
    return(paste(local_members, collapse = ", "))
  }

  external_members <- relations$target_members[match(id, relations$target_synset_id)]
  if (is.na(external_members)) {
    stop("No members available for synset: ", id, call. = FALSE)
  }

  external_members
}

chair_synset <- select_chair_synset()
path_ids <- walk_hypernyms(chair_synset, steps = 3L)

hypernym_path <- tibble(
  level = seq_along(path_ids) - 1L,
  synset_id = path_ids,
  members = vapply(path_ids, lookup_members, character(1)),
  definition = vapply(path_ids, lookup_definition, character(1))
)

kable(
  hypernym_path,
  col.names = c("Level", "Synset ID", "Members", "Definition"),
  caption = "A four-step hypernym path from chair to activity",
  row.names = FALSE
)
A four-step hypernym path from chair to activity
Level Synset ID Members Definition
0 oewn-00599171-n chair, professorship the position of professor
1 oewn-00587299-n berth, office, place, position, post, situation a job in an organization
2 oewn-00583425-n business, job, line, line of work, occupation the principal activity in your life that you do to earn money
3 oewn-00408356-n activity any specific behavior

The walk assumes this extract contains the chosen chair sense and three hypernym links above it. The functions stop when that assumption is false, so a missing lemma cannot produce a half-empty path that looks valid.

The hierarchy is useful for broadening a search. A reader looking for occupations might want chair, office, and job to meet at a broader node. The same hierarchy can be too coarse for a task that needs the academic rank, the furniture sense, or a named committee role.

Keep the limits attached

WordNet is a curated snapshot with a version and a licence. It is not a direct measurement of English. It covers much common vocabulary, while domain terms, slang, new coinages, and local uses can be missing.

Sense inventories also draw boundaries. A fine split that helps one annotation task can be too fine for another, so a WordNet sense should be treated as a chosen inventory entry rather than a final answer about meaning.

WordNet is still good for teaching the difference between words and meanings, expanding a controlled search, checking whether two labels share a listed sense, and building small rule-based features where coverage can be inspected.

limit_check <- tibble(
  check = c("version recorded", "licence recorded", "target lemmas recorded"),
  value = c(
    metadata$value[metadata$field == "source_version"],
    metadata$value[metadata$field == "source_licence"],
    metadata$value[metadata$field == "n_lemmas_found"]
  )
)

kable(
  limit_check,
  col.names = c("Check", "Value"),
  caption = "Provenance fields carried with the WordNet extract",
  row.names = FALSE
)
Provenance fields carried with the WordNet extract
Check Value
version recorded Open English Wordnet 2024
licence recorded https://creativecommons.org/licenses/by/4.0
target lemmas recorded 38

What to remember

  • A synset is a set of words sharing one sense.
  • bank has 18 senses in this extract, so one spelling is not one meaning.
  • A simple null check keeps the sense count from sounding stronger than it is.
  • Hypernym links make an is-a hierarchy that can broaden a search.
  • WordNet is useful when a curated inventory is acceptable and limited coverage can be inspected.

The cataloguer keeps the bank drawer only after adding sense labels to the folders inside it.

Sources