applications 2 of 2 applications 2 of 2
play

Applications (2 of 2): Applications (2 of 2): Recognition, - PowerPoint PPT Presentation

Applications (2 of 2): Applications (2 of 2): Recognition, Transduction, Discrimination, Segmentation Alignment etc Segmentation, Alignment, etc. Kenneth Church Kenneth.Church@jhu.edu Kenneth.Church@jhu.edu Dec 9, 2009 1 Solitaire


  1. Applications (2 of 2): Applications (2 of 2): Recognition, Transduction, Discrimination, Segmentation Alignment etc Segmentation, Alignment, etc. Kenneth Church Kenneth.Church@jhu.edu Kenneth.Church@jhu.edu Dec 9, 2009 1

  2. Solitaire � Multiplayer Games: Auctions (Ads) http://www.scienceoftheweb.org/15 ‐ 396/lectures/lecture09.pdf Right Rail Right Rail: Avoid distortions from Mainline Ad commercial interests Dec 9, 2009 2

  3. A Single Auction � A Stream of Continuous Auctions g • Standard Example of Second Price Auction Sta da d a p e o Seco d ce uct o – Single Auction for a Single Apple • Theoretical Result – Second Price Auction � Truth Telling – http://en.wikipedia.org/wiki/Vickrey_auction – Optimal Strategy: • Bid what the apple is worth to you • Don’t worry about what it is worth to others • Don t worry about what it is worth to others – First Price Auction � Truth Telling • Does theory generalize to a continuous stream? Does theory generalize to a continuous stream? Dec 9, 2009 3

  4. Pricing: Cost Per Click (CPC) Pricing: Cost Per Click (CPC) • Equilibrium • B i = your bid i y – Advertisers • B i+1 = next bid • Awareness • CTR i = your click through rate • Sales Sa es • CTR i+1 = next click through rate • New Customers • ROI • CPC i = your price – Users Users – (if we show your ad and user clicks) (if h d d li k ) • Minimize pain • Improvement: CTR � Q (Prior) • Obtain Value • Single Auction: • Single Auction: – Market Maker Market Maker • Maximize – CPC i = B i+1 Revenue • Continuous Stream: • Continuous Stream: • Truth Telling? – CPC i = B i+1 CTR i+1 / CTR i Dec 9, 2009 4

  5. Multi ‐ Player Games � Many Technical Opportunities Many Technical Opportunities • Economics – http://www.wired.com/culture/culturereviews/magazine/17 ‐ 06/nep_googlenomics?currentPage=all • Machine Learning – Learning to Rank – Estimate CTR (Q/Priors) Estimate CTR (Q/Priors) – Sparse Data: • What is the CTR for a new ad? – Errors can be expensive – Errors can be expensive • If CTR is too low for new ad � Penalize Growth • If too high � Reward Bad Guys to do Bad Things • Truth Telling for Continuous Auctions? • Truth Telling for Continuous Auctions? – Probably not, especially if participants can estimate Q better than market maker • Machine Learning: Solitaire � Multi ‐ Player Games – Can I estimate Q better than you can? Man ‐ eating tiger Dec 9, 2009 5

  6. Applications Applications • Recognition: Shannon’s Noisy Channel Model – Speech Optical Character Recognition (OCR) Spelling Speech, Optical Character Recognition (OCR), Spelling • Transduction – Part of Speech (POS) Tagging – Machine Translation (MT) • • Parsing ??? Parsing: ??? Ranking • – Information Retrieval (IR) – Lexicography • Di Discrimination: i i i – Sentiment, Text Classification, Author Identification, Word Sense Disambiguation (WSD) • Segmentation – Asian Morphology (Word Breaking), Text Tiling • Alignment: Bilingual Corpora, Dotplots • Compression Language Modeling: good for everything • Dec 9, 2009 6

  7. Speech � Language Shannon’s: Noisy Channel Model Shannon s: Noisy Channel Model Channel Language Model • I � Noisy Channel � O Model y • I ΄ ≈ ARGMAX I Pr ( I|O ) = ARGMAX I Pr ( I ) Pr ( O|I ) Application Trigram Language Model Independent Independent Word Rank More likely alternatives The This One Two A Three Channel Model We 9 Please In need need 7 7 are will the would also do are will the would also do Application Application Input Input Output Output to 1 Speech Recognition writer rider resolve 85 have know do… OCR (Optical The This One Two A Three all all 9 9 C Character all a1l Please In Recognition) The This One Two A Three of 2 Please In Spelling Correction government goverment the the 1 1 important 657 document question first… issues 14 thing point to Dec 9, 2009 7

  8. Speech � Language Using (Abusing) Shannon’s Noisy Channel Model: Part of g ( g) y Speech Tagging and Machine Translation • Speech p – Words � Noisy Channel � Acoustics • OCR – Words � Noisy Channel � Optics • Spelling Correction – Words � Noisy Channel � Typos d � N i l � T W Ch • Part of Speech Tagging (POS): – POS � Noisy Channel � Words – POS � Noisy Channel � Words • Machine Translation: “Made in America” – English � Noisy Channel � French g y Didn’t have the guts to use this slide at Eurospeech (Geneva) Dec 9, 2009 8

  9. Dec 9, 2009 9

  10. Spelling Correction Dec 9, 2009 10

  11. Dec 9, 2009 11

  12. Dec 9, 2009 12

  13. Dec 9, 2009 13

  14. Dec 9, 2009 14

  15. Evaluation Evaluation Dec 9, 2009 15

  16. Performance Performance Dec 9, 2009 16

  17. The Task is Hard without Context The Task is Hard without Context Dec 9, 2009 17

  18. Easier with Context Easier with Context • actuall actual actually actuall, actual, actually – … in determining whether the defendant actually will die will die. • constuming, consuming, costuming • conviced, convicted, convinced i d i t d i d • confusin, confusing, confusion • workern, worker, workers Dec 9, 2009 18

  19. Easier with Context Dec 9, 2009 19

  20. Context Model Context Model Dec 9, 2009 20

  21. Dec 9, 2009 21

  22. Dec 9, 2009 22

  23. Dec 9, 2009 23

  24. Dec 9, 2009 24

  25. Future Improvements Future Improvements • Add More Factors Add More Factors – Trigrams – Thesaurus Relations Thesaurus Relations – Morphology – Syntactic Agreement S t ti A t – Parts of Speech • Improve Combination Rules b l – Shrink (Meaty Methodology) Dec 9, 2009 25

  26. Dec 9, 2009 26

  27. Conclusion (Spelling Correction) Conclusion (Spelling Correction) • There has been a lot of interest in smoothing There has been a lot of interest in smoothing – Good ‐ Turing estimation – Knesser ‐ Ney Knesser Ney • Is it worth the trouble? • Ans: Yes (at least for recognition applications) Dec 9, 2009 27

  28. Dec 9, 2009 28

  29. Dec 9, 2009 29

  30. Dec 9, 2009 30

  31. Dec 9, 2009 31

  32. Dec 9, 2009 32

  33. Dec 9, 2009 33

  34. Dec 9, 2009 34

  35. Dec 9, 2009 35

  36. Dec 9, 2009 36

  37. Dec 9, 2009 37

  38. Dec 9, 2009 38

  39. Dec 9, 2009 39

  40. Dec 9, 2009 40

  41. Dec 9, 2009 41

  42. Dec 9, 2009 42

  43. Dec 9, 2009 43

  44. Dec 9, 2009 44

  45. Dec 9, 2009 45

  46. Dec 9, 2009 46

  47. Dec 9, 2009 47

  48. Dec 9, 2009 48

  49. Dec 9, 2009 49

  50. Dec 9, 2009 50

  51. Aligning Words Aligning Words Dec 9, 2009 51

  52. Dec 9, 2009 52

  53. Dec 9, 2009 53

  54. Dec 9, 2009 54

  55. Dec 9, 2009 55

  56. Dec 9, 2009 56

  57. Dec 9, 2009 57

  58. Dec 9, 2009 58

  59. Dec 9, 2009 59

  60. Dec 9, 2009 60

  61. Dec 9, 2009 61

  62. Dec 9, 2009 62

  63. Dec 9, 2009 63

  64. Dec 9, 2009 64

  65. Dec 9, 2009 65

  66. Dec 9, 2009 66

  67. Dec 9, 2009 67

  68. Dec 9, 2009 68

Recommend


More recommend