The University Carlos III of Madrid at TREC 2011 Crowdsourcing Track: Notebook Paper

760 views
662 views

Published on

This notebook paper describes our participation in both tasks of the TREC 2011 Crowdsourcing Track. For the first one we submitted three runs that used Amazon Mechanical Turk: one where workers made relevance judgments based on a 3-point scale, and two similar runs where workers provided an explicit ranking of documents. All three runs implemented a quality control mechanism at the task level, which was based on a simple reading comprehension test. For the second task we submitted another three runs: one with a stepwise execution of the GetAnotherLabel algorithm by Ipeirotis et al., and two others with a rule-based and a SVM-based model. We also comment on several topics regarding the Track design and evaluation methods.

Published in: Technology, Business
0 Comments
1 Like
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
760
On SlideShare
0
From Embeds
0
Number of Embeds
6
Actions
Shares
0
Downloads
7
Comments
0
Likes
1
Embeds 0
No embeds

No notes for slide

The University Carlos III of Madrid at TREC 2011 Crowdsourcing Track: Notebook Paper

  1. 1. Practical and Effective Design of a Crowdsourcing Task for Unconventional Relevance Judging Julián Urbano @julian_urbano Mónica Marrero, Diego Martín, Jorge Morato, Karina Robles and Juan Lloréns University Carlos III of Madrid TREC 2011Picture by Michael Dornbierer Gaithersburg, USA · November 18th
  2. 2. Task ICrowdsourcing Individual Relevance Judgments
  3. 3. In a Nutshell• Amazon Mechanical Turk, External HITs• All 5 documents per set in a sigle HIT = 435 HITs• $0.20 per HIT = $0.04 per document ran out of time graded slider hterms Hours to complete 8.5 38 20.5 HITs submitted (overhead) 438 (+1%) 535 (+23%) 448 (+3%)Submitted workers (just preview) 29 (102) 83 (383) 30 (163)Average documents per worker 76 32 75 Total cost (including fees) $95.7 $95.7 $95.7
  4. 4. Document Preprocessing• Ensure smooth loading and safe rendering – Null hyperlinks – Embed all external resources – Remove CSS unrelated to style or layout – Remove unsafe HTML elements – Remove irrelevant HTML attributes 4
  5. 5. Display Mode hterms run• With images• Black & white, no images 5
  6. 6. Display Mode (and II)• Previous experiment – Workers seem to prefer images and colors – But some definitelly go for just text• Allow them both, but images by default• Black and white best with highlighting – 7 (24%) workers in graded – 21 (25%) in slider – 12 (40%) in hterms 6
  7. 7. HIT Design 7
  8. 8. Relevance Question• graded: focus on binary labels• Binary label – Bad = 0, Good = 1 – Fair: different probabilities? Chose 1 too• Ranking – Order by relevance, then by failures in Quality Control and then by time spent
  9. 9. Relevance Question (II)• slider: focus on ranking• Do not show handle at the beginning – Bias – Lazy indistinguishable from undecided• Seemed unclear it was a slider
  10. 10. Relevance Question (III) 100 200 300 400 500 600 700 Frequency 0 0 20 40 60 80 100 slider value 10
  11. 11. Relevance Question (IV)• Binary label – Threshold – Normalized between 0 and 100 – Worker-Normalized threshold = 0.4 – Set-Normalized – Set-Normalized Threshold – Cluster• Ranking label – Implicit
  12. 12. Relevance Question (and V)• hterms: focus on ranking, seriously• Still unclear? 600 600 Frequency Frequency 400 400 200 200 0 0 0 20 40 60 80 100 0 20 40 60 80 100 slider value slider value
  13. 13. Quality Control• Worker Level: demographic filters• Task Level: additional info/questions – Implicit: work time, behavioral patterns – Explicit: additional verifiable questions• Process Level: trap questions, training• Aggregation Level: consensus from redundancy
  14. 14. QC: Worker Level• At least 100 total approved HITs• At least 95% approved HITs – 98% in hterm• Work in 50 HITs at most• Also tried – Country – Master Qualifications
  15. 15. QC: Implicit Task Level• Time spent in each document – Images and Text modes together• Don’t use time reported by Amazon – Preview + Work time• Time failure: less than 4.5 secs 15
  16. 16. QC: Implicit Task Level (and II) Time Spent (secs) graded slider hterms Min 3 3 3 1st Q 10 14 11 Median 15 23 19 16
  17. 17. QC: Explicit Task Level• There is previous work with Wikipedia – Number of images – Headings – References – Paragraphs• With music / video – Aproximate song duration• Impractical with arbitrary Web documents
  18. 18. QC: Explicit Task Level (II)• Ideas – Spot nonsensical but syntactically correct sentences “the car bought a computer about eating the sea” • Not easy to find the right spot to insert it • Too annoying for clearly (non)relevant documents – Report what paragraph made them decide • Kinda useless without redundancy • Might be several answers• Reading comprehension test
  19. 19. QC: Explicit Task Level (III)• Previous experiment – Give us 5-10 keywords to describe the document • 4 AMT runs with different demographics • 4 faculty members – Nearly always gave the top 1-2 most frequent terms • Stemming and removing stop words• Offered two sets of 5 keywords, choose the one better describing the document 19
  20. 20. QC: Explicit Task Level (and IV)• Correct – 3 most frequent + 2 in the next 5• Incorrect – 5 in the 25 least frequent• Shuffle and random picks• Keyword failure: chose the incorrect terms 20
  21. 21. QC: Process Level• Previous NIST judgments as trap questions?• No – Need previous judgments – Not expected to be balanced – Overhead cost – More complex procress – Do not tell anything about non-trap examples 21
  22. 22. Reject Work and Block Workers• Limit the number of failures in QC Action Failure graded slider hterms Keyword 1 0 1 Reject HIT Time 2 1 1 Keyword 1 1 1Block Worker Time 2 1 1 Total HITs rejected 3 (1%) 100 (23%) 13 (3%) Total Workers blocked 0 (0%) 40 (48%) 4 (13%) 22
  23. 23. Workers by Country Preview Accept Reject % P %A %R Preview Accept Reject % P %A %R Preview Accept Reject % P % A %RAustralia 8 100%Bangladesh 15 3 2 75% 60% 40%Belgium 2 100%Canada 2 100% 1 100% 1 100%Croatia 4 1 80% 100% 0%Egypt 1 100%Finland 11 50 1 18% 98% 2% 4 100% 8 43 1 15% 98% 2%France 9 24 2 26% 92% 8%Germany 1 100%Guatemala 1 100%India 236 214 1 52% 100% 0% 543 235 63 65% 79% 21% 235 190 7 54% 96% 4%Indonesia 2 100%Jamaica 8 5 62% 100% 0%Japan 6 100% 2 100%Kenya 2 100%Lebanon 3 100%Lithuania 6 1 86% 100% 0%Macedonia 1 100%Moldova 1 100%Netherlands 1 100%Pakistan 1 100% 4 1 80% 100% 0% 1 100%Philippines 8 100% 3 100%Poland 1 100%Portugal 2 100%Romania 1 100% 1 100% 5 100%Saudi Arabia 3 1 75% 100% 0% 4 4 50% 100% 0%Slovenia 2 1 67% 100% 0% 3 1 75% 100% 0% 8 2 80% 100% 0%Spain 16 100% 15 100% 9 100%Switzerland 1 100% 7 100%United Arab Emirates 27 12 1 68% 92% 8% 35 9 80% 100% 0%United Kingdom 8 3 73% 100% 0% 18 18 2 47% 90% 10% 8 16 33% 100% 0%United States 246 166 1 60% 99% 1% 381 110 28 73% 80% 20% 242 174 5 57% 97% 3%Yugoslavia 17 14 2 52% 88% 13% 1 100% Average 527 435 3 77% 99% 1% 1086 428 100 83% 90% 10% 579 435 13 85% 99% 1%
  24. 24. Results (unofficial) Consensus Truth Acc Rec Prec Spec AP NDCG Median .761 .752 .789 .700 .798 .831 graded .667 .702 .742 .651 .731 .785 slider .659 .678 .710 .632 .778 .819 hterms .725 .725 .781 .726 .818 .846 NIST Truth Acc Rec Prec Spec AP NDCG Median .623 .729 .773 .536 .931 .922 graded .748 .802 .841 .632 .922 .958 slider .690 .720 .821 .607 .889 .935 hterms .731 .737 .857 .728 .894 .932
  25. 25. Task IIAggregating MultipleRelevance Judgments
  26. 26. Good and Bad Workers• Bad ones in politics might still be good in sports• Topic categories to distinguish – Type: Closed, limited, navigational, open-ended, etc. – Subject: politics, people, shopping, etc. – Rareness: topic keywords in Wordnet? – Readability: Flesch test
  27. 27. GetAnotherLabel• Input – Some known labels – Worker responses• Output – Expected label of unknowns – Expected quality for each worker – Confusion matrix for each worker 27
  28. 28. Step-Wise GetAnotherLabel• For each worker wi compute expected quality qi on all topics and quality qij on each topic type tj.• For topics in tj, use only workers with qij>qi• We didn’t use all known labels by good workers to compute their expected quality (and final label), but only labels in the topic category• Rareness seemed to work slightly better
  29. 29. Train Rule and SVM Models• Relevant-to-nonrelevant ratio – Unbiased majority voting• For all workers , average correct-to-incorrect ratio when saying relevant/nonrelevant• For all workers, average posterior probability of relevant/nonrelevant – Based on the confusion matrix from GerAnotherLabel
  30. 30. Results (unofficial) Consensus Truth Acc Rec Prec Spec AP NDCG Median .811 .818 .830 .716 .806 .921 rule .854 .818 .943 .915 .791 .904 svm .847 .798 .953 .931 .855 .958 wordnet .630 .663 .730 .574 .698 .823 NIST Truth Acc Rec Prec Spec AP NDCG Median .640 .754 .625 .560 .111 .359 rule .699 .754 .679 .644 .166 .415 svm .714 .750 .700 .678 .082 .331 wordnet .571 .659 .560 .484 .060 .299
  31. 31. Sum Up
  32. 32. • Really work the task design – “Make it simple, but not simpler” (A. Einstein) – Make sure they understand it before scaling up• Find good QC methods at the explicit task level for arbitrary Web pages – Was our question too obvious?• Pretty decent judgments compared to NIST’s• Look at the whole picture: system rankings• Study long-term reliability of Crowdsourcing – You can’t prove God doesn’t exist – You can’t prove Crowdsourcing works

×