Il prossimoProssima guida
Precisione e richiamo
Fondamenti
GUIDA TECNICA
Recall stima il numero di documenti reattivi trovati da una recensione; Precision stima quanti documenti etichettati come reattivi sono realmente reattivi.
Ascolta questa guidaNarrato dall'IA · ~3 min
Elusion examines responsive documents hidden in the set classified nonresponsive. Sampling supports estimates, not a guarantee of perfect production.
In technology-assisted document review, recall is the proportion of all truly responsive documents that the process identifies. Precision is the proportion of documents identified as responsive that are truly responsive. A confusion matrix separates true positives, false positives, false negatives, and true negatives against a defined reference classification. These measures answer different questions: high recall can come with lower precision, and vice versa. Elusion estimates how many responsive documents are present in the population the system classified as nonresponsive. It can be easier to sample than the entire responsive population, but low elusion does not always establish high recall, especially when responsive documents are rare. EDRM’s statistical sampling guide warns that inference depends on prevalence and a valid sample, and that derived recall estimates do not automatically inherit the confidence level of their component estimates. Statistical random sampling can support quantitative estimates when the population, coding criteria, sample design, and uncertainty are documented. Judgmental review may help find examples or guide training, but it does not provide the same statistical conclusions. The result is an estimate tied to its sample and assumptions, not proof that every responsive document was found. Teams should define responsiveness with counsel, compare reviewers to a defensible reference standard, and explain limitations in any validation report. The specific protocol should fit the matter and any agreements or court orders.
Counsel estimates whether the responsive set may contain important uncoded material.
A team samples documents coded nonresponsive to estimate elusion.
A reviewer compares machine coding with a documented human-coded reference set.
Parties agree on sampling method, confidence goals, and the review population.
As review platforms add analytics and generative features, defensible validation still depends on a clear population, documented criteria, and appropriate sampling. Teams may refine workflows as they learn more about the collection. Measurements help assess risk and workload, but they do not replace legal judgment or matter-specific agreements. Courts and parties may choose different protocols based on collection size, claims, and production needs. As models and search tools change, teams should explain their design and test behavior on representative data. Transparent estimates help discuss risk, but no single threshold resolves every legal dispute.
Newsletter gratuita
Tre storie verificate sull'IA ogni mattina nei giorni feriali, scritte in un inglese semplice. Gratuito per sempre, senza pubblicità.
Un'email ogni giorno feriale. Disiscriviti con un clic. Non vendiamo né condividiamo mai il tuo indirizzo.
Mettiti alla prova
Feedback immediato su ogni risposta e un certificato condivisibile con un ID verificabile una volta superato un corso.
Sostieni l'istruzione gratuita sull'IA. AI Understanding è un'organizzazione non profit 501(c)(3) — nessun annuncio, nessun paywall, mai. Fai una donazione
Recall estimates how many responsive documents a review found; precision estimates how many documents labeled responsive are truly responsive. Elusion examines responsive documents hidden in the set classified nonresponsive. Sampling supports estimates, not a guarantee of perfect production.
Precision asks what share of identified responsive documents are actually responsive.
The item is responsive in the reference coding but missed by the review.
Elusion concerns responsive items among the nonresponsive population.
EDRM cautions that low elusion may be misleading when responsive prevalence is very low.
Statistical inference requires an appropriate selection design, not convenience selection.
Continua a imparare
Altre guide selezionate per questo argomento
Il prossimoProssima guida
Precisione e richiamo
Fondamenti