Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.



Published on

  • Be the first to comment

  • Be the first to like this


  1. 1. BIONLP09 and CRFs Farzaneh Sarafraz 18 February 2009    
  2. 2. BioNLP'09 Event rather than entity  Most entities are given  3 tasks  − Event detection and characterization − Event argument recognition − Negations and speculations    
  3. 3. Example quot;I kappa B/MAD­3 masks the nuclear localization  signal of NF­kappa B p65 and requires the  transactivation domain to inhibit NF­kappa B  p65 DNA binding. quot; Event: negative regulation Trigger: masks Theme1: the first p65 Cause: MAD­3 Site: nuclear localization signal    
  4. 4. Example quot;In contrast, NF­kappa B p50 alone fails to  stimulate kappa B­directed transcription, and  based on prior in vitro studies, is not  directly regulated by I kappa B. quot; Event: regulation Theme1: this p50 Trigger: regulated Negation: true for this event Speculation: none    
  5. 5. HMM and MEMM (X1, X2, ...) Observations (Y1, Y2, ...) labels p(Xi , Yi)   X  ranges over observation sequence  Y ranges over and label sequence Requires independence assumption  i.e. each item is labelled independently    
  6. 6. Conditional Random Field p(Y |X)  Y: label sequence X: observation sequence Maximise p     
  7. 7. MMEM Label Bias Problem Probability given the current state  − Transitions leaving a state compete against each other  not all states  − Per­state normalization − Probability bias towards states with few transitions − Demonstrated experimentall    
  8. 8. Label Bias Example Training data:  − A B C D − A B D D − A B C E − A B D C Model says:  − C > D 50% − C > E 50% Why predict E when D is much more common?    
  9. 9. CRF Solution Model probability of transitions and probability   of states CRFs  − Models probability of transition between states − Probability is conditional on current observation − Not normalised − Considers many quot;featuresquot; of observations    
  10. 10. Features quot;edge featuresquot; as well as quot;vertex featuresquot;  − Word is capitalized − Word ends in quot;­ingquot; − Label is quot;proper nounquot; Features are important!     
  11. 11. End.