350 likes | 367 Views
This paper introduces iCrowd, an adaptive crowdsourcing framework that addresses the challenge of quality control in crowdsourcing. By estimating worker accuracy and selectively assigning tasks, iCrowd improves the overall quality of results.
E N D
iCrowd: An Adaptive Crowdsourcing Framework Ju Fan1, Guoliang Li2, Beng Chin Ooi1, Kian-Lee Tan1, Jianhua Feng2
Crowdsourcing is very useful… • Humans are better than the machines at many complex jobs • Entity Resolution • Are “IBM” and “Big Blue” the same company? • Classification • Is an image pornography? • Others • Translation, Image Tagging, etc. • Crowdsourcing • Harness Human Intelligence Adaptive Crowdsourcing, SIGMOD 2015
But it often yields low-quality results • Crowdsourcing outsources tasks for solutions from an unknown group of people (aka workers) • Cheating: answering tasks randomly for $$$ • Attacking: providing adverse answers • Under-qualified: having low answer accuracy Adaptive Crowdsourcing, SIGMOD 2015
Quality Control in Crowdsourcing • A redundancy-based strategy • Assign a crowdsourcing task to multipleworkers • Derive result by aggregating worker answers • A simple solution: Majority Voting Not effective if there are many low-quality workers YES Final Result Same Product? NO NO • iPhone 6 • iPhone 6 Plus NO Task Assignment Answer Aggregation Adaptive Crowdsourcing, SIGMOD 2015
Considering Worker’s Quality • Offline Worker Quality Estimation • Gold-Injected: Leverage a small amount of “gold” tasks • Estimating quality as ratio of correctly completed “gold” tasks • Aggregating worker answers using probabilistic fusion • EM-Based: Exploit an Expectation-Maximization technique • Iteratively estimating worker quality and aggregating answers • Qualification test: eliminating “bad” workers • Quality-Aware OnlineTask Assignment • AskIt!, QASCA (also in this research session) Perform well in simple crowdsourcing tasks, e.g., image labeling Adaptive Crowdsourcing, SIGMOD 2015
Accuracy Diversity in Complicated Tasks • More complicated crowdsourcing tasks may require domain knowledge, e.g., • Evaluating quality of question-answers (QA) • Workers are only good at tasks in their familiardomains Does a user-generated answer address its question? • Question: Who first proposed Heliocentrism? • Answer: NicolausCopernicus Worker QA Domains Adaptive Crowdsourcing, SIGMOD 2015
Challenge & Opportunity • Challenge: Accuracy diversity makes existing solutions inadequate for producing high-quality result • Expert in some domains, spammer in others • Being good in qualification, performing bad in other tasks • Opportunity: Adaptive Crowdsourcing • Can we selectively assign tasks to workers with specific expertise to further improve the quality? Adaptive Crowdsourcing, SIGMOD 2015
Outline • Overview of Adaptive Crowdsourcing • The iCrowd Framework • Diverse Accuracy Estimation • A Graph-Based Approach • Online Task Assignment • An Adaptive Assignment Approach • Experimental Study • Conclusion Adaptive Crowdsourcing, SIGMOD 2015
iCrowd: Adaptive Crowdsourcing • Assignment is triggered by a requestfrom a worker • On-the-fly estimating worker accuracy on completed tasks • Assigning the worker a task, to which she has thehighest estimated accuracy among all online workers • Once a worker submits his answer to a task, adjusting estimation of her accuracies • Addressing the cold-start problem by leveraging some microtasks with ground truth, called qualification tasks Adaptive Crowdsourcing, SIGMOD 2015
Running Example • Crowdsourcing Tasks for Entity Resolution • Illustration of iCrowd Adaptive Crowdsourcing, SIGMOD 2015
Graph-Based Accuracy Estimation • Estimating worker’s accuracies on unassigned tasks from their similar completed tasks • Observed Accuracies : probability of answer correctness on the microtasks completed by worker • Estimated Accuracies : probability of answer correctness on the microtasks to be assigned to worker Estimation: correctly completed will correctly complete ? Adaptive Crowdsourcing, SIGMOD 2015
Graph-Based Accuracy Estimation • Estimation Optimization Objectives • Objective 1: Minimize accuracy differences of similar tasks • Worker has comparable accuracies on similar microtasks • Objective 2: Minimize difference between and • Not to deviate too much from the observed accuracies • Optimization Formulation Objective 2 Objective 1 : a parameter for balancing the two objectives Adaptive Crowdsourcing, SIGMOD 2015
Graph-Based Accuracy Estimation • Analytical Solution of Optimization • An Iterative Approach for Computing Solution • Linearity Property: avoid online iterative computation • Offline: , given , compute • Online: Simply estimating (linear sum) Challenging Adaptive Crowdsourcing, SIGMOD 2015
Adaptive Microtask Assignment • Adaptively selecting top workers with higher accuracy, as estimated accuracies are continuously updated. • Top worker set : a set of top workers with the highest accuracies on microtask Accuracy • A worker may be the top worker for multiple microtasks • However, she can only complete at ONE task per time! Top Worker Sets ...... Adaptive Crowdsourcing, SIGMOD 2015
Optimal Microtask Assignment • Microtask Assignment • Given top worker set for each microtask , compute an optimal assignment such that each worker is assigned at most ONE microtask. • Basic Idea of OptimalAssignment • Making as many microtasks to be globally completed as possible, so as to improve accuracy estimation • Complexity: NP-Hard (reduced k-set packing) • A Greedy-Based Assignment Algorithm Adaptive Crowdsourcing, SIGMOD 2015
Qualification Microtask Assignment • Qualification Microtask Assignment • Preferring microtasks from which we can infer accuracies on more other microtasks larger “influence” • Optimization: An NP-hard problem Adaptive Crowdsourcing, SIGMOD 2015
Outline • Overview of Adaptive Crowdsourcing • The iCrowd Framework • Diverse Accuracy Estimation • A Graph-Based Approach • Task Assignment • An Adaptive Assignment Approach • Experimental Study • Conclusion Adaptive Crowdsourcing, SIGMOD 2015
Experiment Setup • Datasets • YahooQA dataset: Evaluating quality of Question-Answers • ItemCompare dataset: compare two items based on a specified comparison criteria • To ensure that all approaches were compared on the same worker set, collect enough answers for each microtask from Amazon Mechanical Turk Adaptive Crowdsourcing, SIGMOD 2015
Experiment Setup • Baseline Approaches • We compared both task assignment and aggregation. • [RandomMV]:used a random strategy for task assignment and aggregated workers’ answers using majority voting. • [RandomEM]: used the random strategy for task assignmentand an EM algorithm to aggregate the answers. • [AvgAccPV]: estimated an average accuracy for each worker by injecting some “gold” microtasks and aggregated the answers using a probabilistic verification model. • Evaluation Metric • Accuracy = Ratio of correctly answered microtasks Adaptive Crowdsourcing, SIGMOD 2015
Effect of Similarity Measures • Different similarity measures did not significantly affect the performance • Cos(topic) with topic analysis achieved better performance • Similarity Measures • Jaccard • Cosine with TF-IDF weights • Cosine with “topic” weights analyzed by the LDA model Adaptive Crowdsourcing, SIGMOD 2015
Effect of Parameter • all connected microtasks having the same estimated accuracies, which failed to capture accuracy diversity • is a large estimated accuracies is very close to the observed ones, which means the estimation could not benefit from the graph-based inference Adaptive Crowdsourcing, SIGMOD 2015
Effect of Adaptive Assignment The proposed adaptive assignment achieved the best accuracy Adaptive Crowdsourcing, SIGMOD 2015
Comparison with Baselines iCrowd achieves 10% - 20% improvement on accuracy Adaptive Crowdsourcing, SIGMOD 2015
Efficiency of Microtask Assignment • Evaluating efficiency using simulation • iCrowd scales well in terms of number of tasks Adaptive Crowdsourcing, SIGMOD 2015
Conclusion • An adaptive crowdsourcing framework iCrowd • A graph-based estimation model to capture diverse accuracies of workers across different tasks • Efficient algorithms to enable instant task assignments • A system deployed on Amazon Mechanical Turk • Higher crowdsourcing quality on accuracy Adaptive Crowdsourcing, SIGMOD 2015
Thanks SQL Suggestion, ICDE 2011
Problem Formulation • Microtask: • A crowdsourcing requester publishes a set of microtasks • Worker: • A set of workers will work on the microtasks in • Worker’s Diverse Accuracy • Probability • Microtask Assignment • Each microtask is assigned to workers • Objective: maximizing overall accuracy • Limited knowledge of accuracy • Dynamic worker set Adaptive Crowdsourcing, SIGMOD 2015
Implementation of iCrowd on AMT Adaptive Crowdsourcing, SIGMOD 2015
Optimal Microtask Assignment • Discussions • Dynamical data structure for maintaining top workers • Periodical microtask assignment updates • For a worker not assigned with microtasks, actively testing by using • Qualification microtasks with ground truth • Globally completed microtask with consensus answers Adaptive Crowdsourcing, SIGMOD 2015
Effect of Qualification Assignment Influence-based qualification test assignment performs better Adaptive Crowdsourcing, SIGMOD 2015
Accuracy Diversity in Complicated Tasks • More complicated crowdsourcing tasks may require domain knowledge, e.g., • Comparing two items based on a specified criteria Which food has more calories? • Food A: Chocolate • Food B: Honey Comparison Domain Worker Adaptive Crowdsourcing, SIGMOD 2015
Diverse Accuracy Estimation • Observation:accuracies of a worker tend to be comparable on similarmicrotasks in same domains • Idea: Estimating worker accuracies by task similarity • Model: • graph based estimation model • Algorithm: • Estimating accuracies in a fast way Adaptive Crowdsourcing, SIGMOD 2015
Graph-Based Accuracy Estimation • Observed Accuracy Estimation • If is a qualification microtask: if ’s answer is equal to the ground truth; otherwise • If ’s ground truth is unknown, use voting-based method: compare ’s answer with the answers from other workers on this microtask • Discussion on Microtask Similarity • For textual microtasks, employing existing similarity metrics, e.g., Jaccard, Edit-Distance, Topic-based similarity, etc. Adaptive Crowdsourcing, SIGMOD 2015
Graph-Based Accuracy Estimation • Observed Accuracy Estimation • If is a qualification microtask: if ’s answer is equal to the ground truth; otherwise • If ’s ground truth is unknown, use voting-based method: compare ’s answer with the answers from other workers on this microtask • Discussion on Microtask Similarity • For textual microtasks, employing existing similarity metrics, e.g., Jaccard, Edit-Distance, Topic-based similarity, etc. Adaptive Crowdsourcing, SIGMOD 2015
Optimal Microtask Assignment • Formal Optimization Problem • Finding an assignment scheme s.t. • Complexity • Reduced from a NP-hard problem, k-set packing • A Greedy-Based Assignment Algorithm Adaptive Crowdsourcing, SIGMOD 2015