Keyness Analysis

The most common is the log-likelihood ratio (G²), which compares observed and expected frequencies across the two corpora and is robust to the skewed, low-count distributions typical of word frequencies. Chi-squared was used historically but is less reliable for rare items. Increasingly, analysts report log-likelihood for significance alongside an effect-size metric such as %DIFF or log ratio to rank keywords by how large, not just how reliable, the difference is.

The reference corpus defines 'normal', so it should be larger than the target and appropriate to the comparison. A general balanced corpus (such as a national reference corpus) suits questions about what is distinctive relative to language at large, while a contrasting subcorpus suits questions about what differs between two specific varieties. A mismatched reference produces keywords that merely reflect the mismatch, so the choice must be justified.

In large corpora, the significance test has so much power that even negligible frequency differences become 'significant', producing inflated keyword lists dominated by high-frequency function words. Effect-size measures such as %DIFF or log ratio quantify how much more frequent a word is, allowing analysts to rank by substantive importance rather than statistical detectability, which Gabrielatos and others argue is essential for meaningful keyness analysis.