Corpus Keyness Analysis

Description

Statistically compares word frequencies between a target corpus and a reference corpus to find words and collocations that are unusually frequent, revealing distinctive concerns.

Corpus linguistics provides robust statistics for comparing language across groups. Keyness compares each word's frequency in a target corpus (e.g. churned users' feedback) with a reference corpus (e.g. retained users' feedback) using log-likelihood or chi-square, with effect sizes such as log ratio. Collocation analysis (MI, t-score, log-Dice) shows which words co-occur, and concordance lines show words in context so interpretation stays grounded. Unlike word clouds or sentiment scores, keyness gives significance and effect size for differences and links back to actual utterances. It sits between qualitative reading and topic modelling, and works well with small to mid-size corpora.

Attributes

Loss type
Unmigrated
Data
Mixed
Effort
Days

Steps

  1. Assemble a target corpus and an appropriate reference corpus, with comparable text types.
  2. Tokenise and lemmatise consistently; decide on handling of case and stop words.
  3. Compute keyness (log-likelihood) and effect size (log ratio) for each word, applying minimum frequency and dispersion thresholds.
  4. Examine collocates of key words and read concordance lines to interpret usage.
  5. Group key words into semantic fields and relate them to user concerns.
  6. Report keyness tables with statistics and illustrative concordances.

Sources

  • Baker, Using Corpora in Discourse Analysis (2006)
  • Rayson & Garside, Comparing corpora using frequency profiling (2000), Workshop on Comparing Corpora
  • Scott & Tribble, Textual Patterns: Key Words and Corpus Analysis in Language Education (2006)

Also known as

  • Keyword analysis (corpus linguistics)
  • Keyness
  • Log-likelihood keyword analysis
  • Collocation analysis

Categories

  • Content
  • Linguistics, Semiotics & Discourse