Date on Master's Thesis/Doctoral Dissertation

8-2018

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Kantardzic, Mehmed

Committee Co-Chair (if applicable)

Elmaghraby, Adel

Committee Member

Elmaghraby, Adel

Committee Member

Lauf, Adrian

Committee Member

Imam, Ibrahim

Committee Member

Lewis, James

Author's Keywords

crowdsourcing; bias; spam; complex tasks; worker filtering

Abstract

Crowdsourcing is proposed as a powerful mechanism for accomplishing large scale tasks via anonymous workers online. It has been demonstrated as an effective and important approach for collecting labeled data in application domains which require human intelligence, such as image labeling, video annotation, natural language processing, etc. Despite the promises, one big challenge still exists in crowdsourcing systems: the difficulty of controlling the quality of crowds. The workers usually have diverse education levels, personal preferences, and motivations, leading to unknown work performance while completing a crowdsourced task. Among them, some are reliable, and some might provide noisy feedback. It is intrinsic to apply worker filtering approach to crowdsourcing applications, which recognizes and tackles noisy workers, in order to obtain high-quality labels. The presented work in this dissertation provides discussions in this area of research, and proposes efficient probabilistic based worker filtering models to distinguish varied types of poor quality workers. Most of the existing work in literature in the field of worker filtering either only concentrates on binary labeling tasks, or fails to separate the low quality workers whose label errors can be corrected from the other spam workers (with label errors which cannot be corrected). As such, we first propose a Spam Removing and De-biasing Framework (SRDF), to deal with the worker filtering procedure in labeling tasks with numerical label scales. The developed framework can detect spam workers and biased workers separately. The biased workers are defined as those who show tendencies of providing higher (or lower) labels than truths, and their errors are able to be corrected. To tackle the biasing problem, an iterative bias detection approach is introduced to recognize the biased workers. The spam filtering algorithm proposes to eliminate three types of spam workers, including random spammers who provide random labels, uniform spammers who give same labels for most of the items, and sloppy workers who offer low accuracy labels. Integrating the spam filtering and bias detection approaches into aggregating algorithms, which infer truths from labels obtained from crowds, can lead to high quality consensus results. The common characteristic of random spammers and uniform spammers is that they provide useless feedback without making efforts for a labeling task. Thus, it is not necessary to distinguish them separately. In addition, the removal of sloppy workers has great impact on the detection of biased workers, with the SRDF framework. To combat these problems, a different way of worker classification is presented in this dissertation. In particular, the biased workers are classified as a subcategory of sloppy workers. Finally, an ITerative Self Correcting - Truth Discovery (ITSC-TD) framework is then proposed, which can reliably recognize biased workers in ordinal labeling tasks, based on a probabilistic based bias detection model. ITSC-TD estimates true labels through applying an optimization based truth discovery method, which minimizes overall label errors by assigning different weights to workers. The typical tasks posted on popular crowdsourcing platforms, such as MTurk, are simple tasks, which are low in complexity, independent, and require little time to complete. Complex tasks, however, in many cases require the crowd workers to possess specialized skills in task domains. As a result, this type of task is more inclined to have the problem of poor quality of feedback from crowds, compared to simple tasks. As such, we propose a multiple views approach, for the purpose of obtaining high quality consensus labels in complex labeling tasks. In this approach, each view is defined as a labeling critique or rubric, which aims to guide the workers to become aware of the desirable work characteristics or goals. Combining the view labels results in the overall estimated labels for each item. The multiple views approach is developed under the hypothesis that workers' performance might differ from one view to another. Varied weights are then assigned to different views for each worker. Additionally, the ITSC-TD framework is integrated into the multiple views model to achieve high quality estimated truths for each view. Next, we propose a Semi-supervised Worker Filtering (SWF) model to eliminate spam workers, who assign random labels for each item. The SWF approach conducts worker filtering with a limited set of gold truths available as priori. Each worker is associated with a spammer score, which is estimated via the developed semi-supervised model, and low quality workers are efficiently detected by comparing the spammer score with a predefined threshold value. The efficiency of all the developed frameworks and models are demonstrated on simulated and real-world data sets. By comparing the proposed frameworks to a set of state-of-art methodologies, such as expectation maximization based aggregating algorithm, GLAD and optimization based truth discovery approach, in the domain of crowdsourcing, up to 28.0% improvement can be obtained for the accuracy of true label estimation.

Share

COinS