Toward a Robust and Universal Crowd Labeling Framework

• Main Author: Khattak, Faiza Khan
• Language: English
• Published: 2017
• Subjects: Computer science; Labels; Bayesian statistical decision theory
• Online Access: https://doi.org/10.7916/D8Q24BJ2

The advent of fast and economical computers with large electronic storage has led to a large volume of data, most of which is unlabeled. While computers provide expeditious, accurate and low-cost computation, they still lag behind in many tasks that require human intelligence such as labeling medical images, videos or text. Consequently, current research focuses on a combination of computer accuracy and human intelligence to complete labeling task. In most cases labeling needs to be done by domain experts, however, because of the variability in expertise, experience, and intelligence of human beings, experts can be scarce. As an alternative to using domain experts, help is sought from non-experts, also known as Crowd, to complete tasks that cannot be readily automated. Since crowd labelers are non-expert, multiple labels per instance are acquired for quality purposes. The final label is obtained by com- bining these multiple labels. It is very common that the ground truth, instance difficulty, and the labeler ability are unknown entities. Therefore, the aggregation task becomes a “chicken and egg” problem to start with. Despite the fact that much research using machine learning and statistical techniques has been conducted in this area (e.g., [Dekel and Shamir, 2009; Hovy et al., 2013a; Liu et al., 2012; Donmez and Carbonell, 2008]), many questions remain unresolved, these include: (a) What are the best ways to evaluate labelers? (b) It is common to use expert-labeled instances (ground truth) to evaluate la- beler ability (e.g., [Le et al., 2010; Khattak and Salleb-Aouissi, 2011; Khattak and Salleb-Aouissi, 2012; Khattak and Salleb-Aouissi, 2013]). The question is, what should be the cardinality of the set of expert-labeled instances to have an accurate evaluation? (c) Which factors other than labeler expertise (e.g., difficulty of instance, prevalence of class, bias of a labeler toward a particular class) can affect the labeling accuracy? (d) Is there any optimal way to combine multiple labels to get the best labeling accuracy? (e) Should the labels provided by oppositional/malicious labelers be dis- carded and blocked? Or is there a way to use the “information” provided by oppositional/malicious labelers? (f) How can labelers and instances be evaluated if the ground truth is not known with certitude? In this thesis, we investigate these questions. We present methods that rely on few expert-labeled instances (usually 0.1% -10% of the dataset) to evaluate various parameters using a frequentist and a Bayesian approach. The estimated parameters are then used for label aggregation to produce one final label per instance. In the first part of this thesis, we propose a method called Expert Label Injected Crowd Esti- mation (ELICE) and extend it to different versions and variants. ELICE is based on a frequentist approach for estimating the underlying parameters. The first version of ELICE estimates the pa- rameters i.e., labeler expertise and data instance difficulty, using the accuracy of crowd labelers on expert-labeled instances [Khattak and Salleb-Aouissi, 2011; Khattak and Salleb-Aouissi, 2012]. The multiple labels for each instance are combined using weighted majority voting. These weights are the scores of labeler reliability on any given instance, which are obtained by inputting the pa- rameters in the logistic function. In the second version of ELICE [Khattak and Salleb-Aouissi, 2013], we introduce entropy as a way to estimate the uncertainty of labeling. This provides an advantage of differentiating between good, random and oppositional/malicious labelers. The aggregation of labels for ELICE version 2 flips the label (for binary classification) provided by the oppositional/malicious labeler thus utilizing the information that is generally discarded by other labeling methodologies. Both versions of ELICE have a cluster-based variant in which rather than making a random choice of instances from the whole dataset, clusters of data are first formed using any clustering approach e.g., K-means. Then an equal number of instances from each cluster are chosen randomly to get expert-labels. This is done to ensure equal representation of each class in the test dataset. Besides taking advantage of expert-labeled instances, the third version of ELICE [Khattak and Salleb-Aouissi, 2016], incorporates pairwise/circular comparison of labelers to labelers and in- stances to instances. The idea here is to improve accuracy by using the crowd labels, which unlike expert-labels, are available for the whole dataset and may provide a more comprehensive view of the labeler ability and instance difficulty. This is especially helpful for the case when the domain experts do not agree on one label and ground truth is not known for certain. Therefore, incorporating more information beyond expert labels can provide better results. We test the performance of ELICE on simulated labels as well as real labels obtained from Amazon Mechanical Turk. Results show that ELICE is effective as compared to state-of-the-art methods. All versions and variants of ELICE are capable of delaying phase transition. The main contribution of ELICE is that it makes the use of all possible information available from crowd and experts. Next, we also present a theoretical framework to estimate the number of expert-labeled instances needed to achieve certain labeling accuracy. Experiments are presented to demonstrate the utility of the theoretical bound. In the second part of this thesis, we present Crowd Labeling Using Bayesian Statistics (CLUBS) [Khattak and Salleb-Aouissi, 2015; Khattak et al., 2016b; Khattak et al., 2016a], a new approach for crowd labeling to estimate labeler and instance parameters along with label aggregation. Our approach is inspired by Item Response Theory (IRT). We introduce new parameters and refine the existing IRT parameters to fit the crowd labeling scenario. The main challenge is that unlike IRT, in the crowd labeling case, the ground truth is not known and has to be estimated based on the parameters. To overcome this challenge, we acquire expert-labels for a small fraction of instances in the dataset. Our model estimates the parameters based on the expert-labeled instances. The estimated parameters are used for weighted aggregation of crowd labels for the rest of the dataset. Experiments conducted on synthetic data and real datasets with heterogeneous quality crowd-labels show that our methods perform better than many state-of-the-art crowd labeling methods. We also conduct significance tests between our methods and other state-of-the-art methods to check the significance of the accuracy of these methods. The results show the superiority of our method in most cases. Moreover, we present experiments to demonstrate the impact of the accuracy of final aggregated labels when used as training data. The results essentially emphasize the need for high accuracy of the aggregated labels. In the last part of the thesis, we present past and contemporary research related to crowd la- beling. We conclude with future of crowd labeling and further research directions. To summarize, in this thesis, we have investigated different methods for estimating crowd labeling parameters and using them for label aggregation. We hope that our contribution will be useful to the crowd labeling community.