NLP models often rely on human-labeled data for training and evaluation. Many approaches crowdsource this data from a large number of annotators with varying skills, backgrounds, and motivations, resulting in conflicting annotations. These conflicts have traditionally been resolved by aggregation methods that assume disagreements are errors. Recent work has argued that for many tasks annotators may have genuine disagreements and that variation should be treated as signal rather than noise. However, few models separate signal and noise in annotator disagreement. In this work, we introduce NUTMEG, a new Bayesian model that incorporates information about annotator backgrounds to remove noisy annotations from human-labeled training data while preserving systematic disagreements. Using synthetic data, we show that NUTMEG is more effective at recovering ground-truth from annotations with systematic disagreement than traditional aggregation methods. We provide further analysis characterizing how differences in subpopulation sizes, rates of disagreement, and rates of spam affect the performance of our model. Finally, we demonstrate that downstream models trained on NUTMEG-aggregated data significantly outperform models trained on data from traditionally aggregation methods. Our results highlight the importance of accounting for both annotator competence and systematic disagreements when training on human-labeled data.

Distributing labeling tasks among hundreds or thousands of annotators is an increasingly important method for annotating large datasets. We present a method for estimating the underlying value (e.g. the class) of each image from (noisy) annotations provided by multiple annotators. Our method is based on a model of the image formation and annotation process. Each image has different characteristics that are represented in an abstract Euclidean space. Each annotator is modeled as a multidimensional entity with variables representing competence, expertise and bias. This allows the model to discover and represent groups of annotators that have different sets of skills and knowledge, as well as groups of images that differ qualitatively. We find that our model predicts ground truth labels on both synthetic and real data more accurately than state of the art methods. Experiments also show that our model, starting from a set of binary labels, may discover rich information, such as different "schools of thought" amongst the annotators, and can group together images belonging to separate categories.

The rapid growth of storage capacity and processing power has caused machine learning applications to increasingly rely on using immense amounts of labeled data. It has become more important than ever to have fast and inexpensive ways to annotate vast amounts of data. With the emergence of crowdsourcing services, the research direction has gravitated toward putting the wisdom of crowds to better use. Unfortunately, spammers and inattentive annotators pose a threat to the quality and trustworthiness of the consensus. Thus, high quality consensus estimation from crowd annotated data requires a meticulous choice of the candidate annotator and the sample in need of a new annotation. Due to time and budget limitations, it is of utmost importance that this choice is carried out while the annotation collection is in progress. We call this process active crowd-labeling. To this end, we propose an active crowd-labeling approach for actively estimating consensus from continuous-valued crowd annotations. Our method is based on annotator models with unknown parameters, and Bayesian inference is employed to reach a consensus in the form of ordinal, binary, or continuous values. We introduce ranking functions for choosing the candidate annotator and sample pair for requesting an annotation. In addition, we propose a penalizing method for preventing annotator domination, investigate the explore-exploit trade-off for incorporating new annotators into the system, and study the effects of inducing a stopping criterion based on consensus quality. We also introduce the crowd-labeled Head Pose Annotations datasets. Experimental results on the benchmark datasets used in the literature and the Head Pose Annotations datasets suggest that our method provides high-quality consensus by using as few as one fifth of the annotations (~80% cost reduction), thereby providing a budget and time-sensitive solution to the crowd-labeling problem.

Distributing labeling tasks among hundreds or thousands of annotators is an increasingly importantmethod for annotating large datasets. We present a method for estimating the underlying value (e.g. the class) of each image from (noisy) annotations providedby multiple annotators. Our method is based on a model of the image formation and annotation process. Each image has different characteristics that are represented in an abstract Euclidean space. Each annotator is modeled as a multidimensional entity with variables representing competence, expertise and bias. This allows the model to discover and represent groups of annotators that have different sets of skills and knowledge, as well as groups of images that differ qualitatively. We find that our model predicts ground truth labels on both synthetic andreal data more accurately than state of the art methods. Experiments also show that our model, starting from a set of binary labels, may discover rich information, such as different "schools of thought" amongst the annotators, and can group together images belonging to separate categories.