Skip to content
Last Dictionary
Plane
Evaluation & BenchmarksData & DatasetsSociety, Ethics & Economy

human evaluation

nounAlso: human eval, human rating, human judgment, manual evaluation

In plain wordsJudging an AI's outputs by asking people to rate or compare them, because for many tasks — writing, conversation, translation — no automatic score captures quality well.

  1. Assessment of system outputs by human judges rather than automatic metrics. Common protocols are absolute ratings on Likert scales (for fluency, adequacy or helpfulness), pairwise or side-by-side preferences, rankings, error annotation against a typology such as Multidimensional Quality Metrics (MQM) in translation, and task-based studies that test whether outputs actually help people. Judges may be crowdworkers, domain experts or real users, as on LMArenaLMArena, prop. n.: A public website that ranks AI chatbots by blind taste test: visitors put one question to two unnamed models, pick the reply they prefer, and millions of such votes produce the rankings.. Human evaluation is the reference against which BLEUBLEU, abbr.: An automatic score for machine translation that checks how many short runs of words in the computer's translation also appear in human reference translations., ROUGEROUGE, abbr.: An automatic score for summaries that checks how much of the wording in human-written reference summaries also appears in the computer-generated one. and model gradersLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers. are validated, but it is costly, slow and noisy: designs must control for rater expertise, fatigue, presentation order and population, and should report inter-annotator agreementinter-annotator agreement, n.: A measure of how often different people labeling the same data give the same answer. Low agreement warns that the task is unclear or the labels are unreliable.. Raters also tend to favor longer, more confident answers, which can reward fluent errors.

    Automatic metrics favored system B, but blind human evaluation by professional translators preferred A.

Origin

Early human judgments of machine translation include John B. Carroll's intelligibility and fidelity ratings for the 1966 ALPAC reportALPAC report, prop. n.: A 1966 US government report that concluded machine translation was slower, worse and costlier than human translators. It ended most American funding for the field for many years.; DARPA's translation evaluations in the 1990s standardized 'adequacy' and 'fluency' scales. Large-scale preference judgments became central with reinforcement learning from human feedbackreinforcement learning from human feedback, n.: A way of training AI using people's judgments: humans compare pairs of answers, a scoring model learns their preferences, and the AI is then trained to produce answers that score well. (2017–2022) and crowdsourced arenas (2023).

Usage note

Not automatically more valid than automatic metrics: rushed or unqualified raters can reward confident mistakes, and experts disagree. In an uplift studyuplift study, n.: An experiment that measures how much an AI really helps people do a task — often a dangerous one, like planning a biological attack — compared with people using only ordinary resources such as the internet., people are the subjects being measured rather than the judges.

Related

See also
  • LLM-as-a-judgeLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers.
  • inter-annotator agreementinter-annotator agreement, n.: A measure of how often different people labeling the same data give the same answer. Low agreement warns that the task is unclear or the labels are unreliable.
  • LMArenaLMArena, prop. n.: A public website that ranks AI chatbots by blind taste test: visitors put one question to two unnamed models, pick the reply they prefer, and millions of such votes produce the rankings.
  • win ratewin rate, n.: The share of head-to-head comparisons in which judges — people or another AI — prefer one model's answer over a rival's for the same request.
  • crowdsourcingcrowdsourcing, n.: Getting many people over the internet to each do a small piece of a big job, such as labeling pictures for a few cents apiece, with anyone free to join in.
  • ALPAC reportALPAC report, prop. n.: A 1966 US government report that concluded machine translation was slower, worse and costlier than human translators. It ended most American funding for the field for many years.
  • reinforcement learning from human feedbackreinforcement learning from human feedback, n.: A way of training AI using people's judgments: humans compare pairs of answers, a scoring model learns their preferences, and the AI is then trained to produce answers that score well.
Contrast with
  • BLEUBLEU, abbr.: An automatic score for machine translation that checks how many short runs of words in the computer's translation also appear in human reference translations.
  • ROUGEROUGE, abbr.: An automatic score for summaries that checks how much of the wording in human-written reference summaries also appears in the computer-generated one.

Sources

  1. ALPAC (1966). Language and Machines: Computers in Translation and Linguistics. National Academy of Sciences–National Research Council.
  2. Freitag, M. et al. (2021). Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation. Transactions of the ACL, 9.

Mentioned in

benchmarkbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time., evaluationevaluation, n.: Testing an AI system to find out what it can and cannot do, how well, and how safely — from exam-style quizzes to hands-on trials and expert reviews., LMArenaLMArena, prop. n.: A public website that ranks AI chatbots by blind taste test: visitors put one question to two unnamed models, pick the reply they prefer, and millions of such votes produce the rankings., BLEUBLEU, abbr.: An automatic score for machine translation that checks how many short runs of words in the computer's translation also appear in human reference translations., Fréchet inception distanceFréchet inception distance, n.: A score for image generators that compares the overall statistics of their pictures with those of real photos, as seen by an image-recognition network; lower is better., GDPvalGDPval, prop. n.: An OpenAI test that pits AI against seasoned professionals on genuine workplace assignments, from reports and spreadsheets to slide decks, in 44 occupations important to the US economy., ground truthground truth, n.: The reference answers treated as correct when checking or training an AI system — for example, the confirmed diagnoses that a medical-imaging model's guesses are compared against., image captioningimage captioning, n.: Having a computer write a short description of a picture, such as 'a dog catching a frisbee in a park', to help blind people, search engines and other AI systems., LLM-as-a-judgeLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers., ROUGEROUGE, abbr.: An automatic score for summaries that checks how much of the wording in human-written reference summaries also appears in the computer-generated one., win ratewin rate, n.: The share of head-to-head comparisons in which judges — people or another AI — prefer one model's answer over a rival's for the same request., HealthBenchHealthBench, prop. n.: An OpenAI test of how well AI chatbots handle health conversations, grading 5,000 realistic chats against checklists written by 262 doctors from 60 countries., Inception scoreInception score, n.: An early score for image generators that rewards pictures an image-recognition network can label confidently and a collection that covers many different labels; higher is better., inter-annotator agreementinter-annotator agreement, n.: A measure of how often different people labeling the same data give the same answer. Low agreement warns that the task is unclear or the labels are unreliable., mean opinion scoremean opinion score, n.: A quality score from 1 (bad) to 5 (excellent) obtained by asking many listeners to rate audio, such as a synthetic voice or a phone call, and averaging their ratings., MT-BenchMT-Bench, prop. n.: A set of 80 two-part questions — in writing, reasoning, math, coding and more — whose answers are graded by a strong AI model, used to compare chat assistants quickly., sandwichingsandwiching, n.: A way of testing oversight methods today: non-experts, helped by an AI, try to get expert-quality answers out of it, and real experts then check whether they succeeded.