概述
Cohen’s kappa is a chance-corrected statistic for two raters on categorical labels; it is one measure, not a verdict on label quality or construct validity.
深入探討
Raw percent agreement — simply counting how often two annotators chose the same label — is intuitive but flawed because it doesn't account for the possibility that annotators could agree just by chance, especially when one label dominates the dataset. Cohen's kappa addresses this by computing the difference between observed agreement and expected agreement (the agreement you'd predict from each annotator's individual label frequencies if their choices were independent), then normalizing that difference against the maximum possible improvement over chance. The formula is kappa = (observed agreement − expected agreement) / (1 − expected agreement), producing a value that is 1 for perfect agreement, 0 for agreement no better than chance, and negative when agreement is worse than chance. Commonly cited rough interpretation bands (such as below 0 as poor, 0.01–0.20 as slight, up to 0.81–1.00 as almost perfect) come from a widely used but informal scale, not a universally agreed statistical standard, and different fields apply different thresholds for what counts as "good enough" to trust a labeled dataset. Cohen's kappa applies specifically to exactly two raters; for three or more, Fleiss’ kappa is one common nominal-label alternative, not a pairwise mean. Other designs, including those with missing ratings, may use Krippendorff’s alpha. For labels with a natural order (like a 1-5 rating scale), a weighted kappa variant is preferred because it treats a one-point disagreement as less severe than a four-point disagreement, whereas standard kappa treats all disagreements equally regardless of magnitude. A common misconception is that a high percent agreement always signals a well-defined labeling task; when the label distribution is skewed, percent agreement can look high even when agreement beyond chance is limited, which kappa can help reveal. Krippendorff's alpha is a related and more flexible alternative that handles missing data and more than two raters natively, and is preferred by some researchers for that flexibility.
現實世界的實施
Two annotators labeling tweets as toxic or not toxic agree 90% of the time, but because 85% of tweets are non-toxic, a kappa calculation shows the agreement beyond chance is only moderate, not excellent.
A team comparing three or more annotators on a multi-class task uses Fleiss' kappa rather than Cohen's kappa, since Cohen's kappa is defined specifically for exactly two raters.
A project with ordinal ratings (like a 1-5 quality scale) uses a weighted kappa variant so that a disagreement of one point counts differently from a disagreement of four points.
A labeling vendor reports a kappa score of 0.55 on a new task, prompting the client to require revised guidelines and a recalibration session rather than accepting the batch as-is.
The Future of Inter-Annotator Agreement and Cohen's Kappa
Agreement reporting should become more transparent about sampling, class balance, rater training, adjudication, and uncertainty. For ordinal, multi-label, or missing-data tasks, select a statistic whose assumptions fit the design rather than stretching Cohen’s kappa. Comparing human and model labels does not establish that either is correct. Use disagreement to identify ambiguous items or instructions, then combine statistics with expert review and task-specific validity checks; avoid treating any single coefficient or conventional band as a universal quality threshold in any application.
付諸實踐
為什麼這很重要
- 成本與預算. 多年來,架構決策決定著效能和營運成本。
- 更明確的決策. 技術教育幫助團隊選擇正確的堆疊,而不僅僅是最新的堆疊。
- 品質管控. 更好的工程選擇可以減少生產中的可靠性事故。
留意
- 優化一項基準測試可以隱藏更廣泛的系統弱點。
- 基礎設施和維護成本常常被低估。
- 隨著系統變得更加複雜,安全性和可觀察性差距可能會擴大。
後續步驟
- 在實施之前定義延遲、品質和成本目標。
- 在實際負載和資料條件下進行基準測試。
- 儀器監控錯誤、漂移和使用者影響。
- 在擴展之前準備回滾和事件回應路徑。
不斷探索
常見問題
What is Inter-Annotator Agreement and Cohen's Kappa?
Inter-annotator agreement (IAA) describes how often labelers apply the same labels to the same items under defined instructions. Cohen’s kappa is a chance-corrected statistic for two raters on categorical labels; it is one measure, not a verdict on label quality or construct validity.
根據指南,為何兩位標註者之間的原始百分比一致性可能會產生誤導?
The guide explains that raw agreement doesn't correct for chance agreement, which can inflate the apparent consistency when one label dominates the data.
Cohen κ 具體修正了什麼,而原始百分比一致率卻沒有?
Kappa subtracts out expected agreement (computed from each annotator's marginal label frequencies) before normalizing, unlike raw percent agreement.
當超過兩位評分者時,常用於名義標籤的替代統計量是什麼?
Fleiss’ kappa is a common multiple-rater extension for nominal ratings; other coefficients may be preferable depending on the design.
何時應該使用加權 κ 變體而非標準 κ?
The guide describes weighted kappa as preferred for ordinal scales, treating smaller disagreements as less severe than larger ones.
當觀測一致性等於由評分者邊際分布估計的機會一致性時,Cohen 的 κ 取何值?
The guide states kappa is 0 when agreement is no better than chance, 1 for perfect agreement, and negative when agreement is worse than chance.