My specific question is how NLP data from multiple human annotators should be aggregated - though general advice related to the question title is appreciated. One critical step that I've seen in research is to assess inter-annotator agreement by Cohen's kappa or some other suitable metric; I've also found research reporting values for various datasets (e.g. here), which is helpful for baselining.

How many annotators should work on each data point depends on time, personnel, and data size requirements/constraints, among other factors (I may ask a followup question for how to find optimal n). However, once n annotators have finished a dataset, how should those n datasets be unified for a "ground truth"? A couple approaches that I have seen used or seem reasonable to me:

  • Take averages over all annotators. Classification problems are sometimes hard to restate as graduated ones, although that seems necessary if an average is to be taken.

  • Express some level of uncertainty in the data for controversial labels, or even omit them from training and evaluation.

  • Add an arbitration step to unify or discard controversial labels. I am not sure this would be worth the annotators' time.

  • Choose some "principal annotator(s)" (possibly determined by IAA scores) who get the final word in conflicts.

Guidance/references for the above and any other steps I can take to make a high quality dataset are much appreciated. I am mostly interested in efficiently removing individual annotator bias even when n is low.


3 Answers 3


Most machine learning algorithms are designed with complete trust in the labels. There is no standard way to model uncertainty in data labels. Thus, create a project-specific threshold for uncertainty to omit data or labelers. For example, a trusted classification data label would require n of m ensemble voting.

One major issue is re-labeling. Systems tend to evolve over time and the definition of labels is refined. Mature data labeling system have a notion of data lineage - "Who labeled what data when with what criteria".

The book "Human-in-the-Loop Machine Learning" by Robert Munro goes into greater detail.

  • $\begingroup$ Second "Human-in-the-Loop Machine Learning" book, it's a very comprehensive resource $\endgroup$ Jan 18, 2023 at 22:32

You ideally want a copy of The Handbook of Linguistic Annotation which covers the issues you’re up against in detail.

The basic idea is:

  • Create annotation guidelines as a training tool to increase interannotator agreement as far as possible
  • Measure interannotator agreement among people using your guidelines to get an idea of the irreducible error
  • Generate as much annotated data as you can

If you have created clear annotation guidelines, you should be able to aggregate data from multiple annotators with no additional processing; you just have to have the kappa that relates to that set of guidelines as a caveat attached to the training data.


You might be interested in a new class on this exact topic I taught at MIT: https://dcai.csail.mit.edu/lectures/dataset-creation-curation/

It covers various methods to establish consensus labels from many noisy annotations (eg. from crowdsourced data labelers) and to estimate our confidence that each consensus label is actually correct.

The best methods (like the CROWDLAB algorithm described in the linked lecture notes) train a ML classifier to augment the data annotators when estimating consensus (and confidence therein). Good estimates need to properly account for:

  • how many annotations an example received,
  • how much the annotators disagreed,
  • the quality of the annotators who labeled this example,
  • the classifier's confidence estimate,
  • and the relative accuracy of the classifier vs individual annotators.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.