120 likes | 205 Views
Systematization of Crowdsoucing for Data Annotation. Aobo , Feb. 2010. Outline. Overview Related Work Analysis and Classification Recommendation Future work Conclusions Reference. Overview. Contribution Provide a faceted analysis of existing crowdsourcing annotation applications.
E N D
Systematization of Crowdsoucing for Data Annotation Aobo, Feb. 2010
Outline • Overview • Related Work • Analysis and Classification • Recommendation • Future work • Conclusions • Reference
Overview • Contribution • Provide a faceted analysis of existing crowdsourcing annotation applications. • Discuss recommendations on how practioners can take advantage of crowdsourcing. • Discuss the potential opportunities in this area. • Defination • Crowdsoucing • GWAP • Distributed Human-based Computation • AMT • HIT
Retaled Works • “A Taxonomy of Distributed Human Computation” • Author: J. Quinn and B. Bederson • Year: 2009 Contribution • Divide the DHC applications into seven genres. • Proposed six dimensions to help characterize the different approaches . • Propose some recommandation and future directions
Related work • “A Survey of Human Computation Systems” • Author: Yuen, Chen and King • Year: 2009 Contribution • General survey of various human computation systems separately. • Compare the GWAPs based on the game structure, verification method, and game mechanism • Present the performance aspect issues of GWAPs.
Analysis and Classification • Dimensions
Analysis and Classification • GWAP • High score in : GUI desine, Implementation cost, Annotation speed, • Low score in : Anonotation cost, Difficulty, Participation time, Domain Coverage, Popularization • Medium score in: Annotation accuracy, Data size • NLP tasks: -Word Sense Disambiguation - Coreference Annotation
Analysis and Classification • AMT • High score in : Annotation cost • Low score in : GUI design, Implementation Cost, Number of Participants, Data size • Medium score in: Popularization, Difficulty, Domain coverage, Paticipation time, Popularization Annotation accuracy, • NLP tasks: - Parsing - Part-of-Speech Tagging
Analysis and Classification • Wisdom of Volunteers • High score in : Number of Participants, Data size , Difficulty, Paticipation time • Low score in : GUI design, Fun • Medium score in: Implementation Cost, Annotation accuracy, • NLP tasks: - Paraphrasing - Machine Translation task - Summarization
Recommendation • GWAP • Submit the GWAP games to a popular game website which provides and recommend new games for players • Uniform game developing platform • AMT • Make the task fun • Rank the employers by their contribuation • Award employers who provide original data to be annotated • Donate the whole or part of the benefit to an charity • Wisdom of Volunteers • Rank the users by their contribuation • Push the tasks to the public users
Conclusions • Propose different dimentions of existing crowdsourcing annotation applications. • Discuss recommendations on each crowdsourcing approach • Discuss the potential opportunities in this area
Reference • Benjamin B. Bederson Alexander J. Quinn. 2009. A taxonomy of distributed human computation. • AniketKittur, Ed H. Chi, and BongwonSuh. 2008. Crowdsourcing user studies with mechanical turk. • Rion Snow, Brendan O’Connor, Daniel Jurafsky, and Andrew Ng. 2008. Cheap and fast – but is it good? evaluating non-expert annotations for natural language tasks. • A. Sorokin and D. Forsyth. 2008. Utility data annotation with amazon mechanical turk. • Luis von Ahn and Laura Dabbish. 2008a. Designing games with a purpose. Commun. ACM, 51(8):58– 67, August. • Luis von Ahn and Laura Dabbish. 2008b. General techniques for designing games with a purpose. Commun. ACM, 51(8):58–67. • Man-Ching Yuen, Ling-Jyh Chen, and Irwin King. 2009. A survey of human computation systems. Computational Science and Engineering, IEEE International Conference on, 4:723–728.