Skip to main content

Advertisement

Table 3 Meta-knowledge assignment results using 10-fold cross validation on the GENIA-MK training set with various settings

From: Extracting semantically enriched events from biomedical literature

Meta-knowledge Disabled features   Disabled learning settings   
   -Meta- -Trigger -Trigger- -Sentence -Citation -Type-based -Biased   
Dimension Value knowledge features argument features features feature regularisation ALL Majority
   clue features   pair features    normalisation factors   
KT Investigation 71.9 61.2 71.7 70.8 71.6 73.2 70.3 71.8 0.0
Analysis 75.6 68.1 74.6 75.6 75.7 75.7 75.0 75.8 0.0
Observation 74.7 71.1 72.7 71.7 75.1 74.0 74.6 75.3 41.4
Fact 67.2 65.5 59.6 56.2 66.8 59.3 64.5 67.3 0.0
Method 61.9 56.0 52.8 61.6 61.7 55.3 52.9 62.0 0.0
Other 74.1 71.6 73.5 73.9 74.7 74.0 74.5 74.8 0.0
  Macro 70.9 65.6 67.5 68.3 70.9 68.6 68.6 71.2 6.9
  Accuracy 73.6 69.4 71.6 71.6 73.9 72.7 73.3 74.1 26.1
CL L3 97.4 97.6 97.7 97.8 97.8 97.8 98.0 97.8 94.5
L2 67.1 69.7 72.0 72.8 73.0 72.6 70.9 73.0 0.0
L1 76.4 78.7 76.0 78.1 78.7 75.9 78.7 78.6 0.0
  Macro 80.3 82.0 81.9 82.9 83.2 82.1 82.5 83.1 31.5
  Accuracy 95.0 95.4 95.5 95.8 95.8 95.7 96.2 95.8 89.6
Polarity Positive 97.7 97.3 97.1 97.6 97.5 97.4 98.1 97.5 95.1
Negative 65.9 65.0 64.1 67.8 67.3 64.2 63.6 67.3 0.0
  Macro 81.8 81.2 80.6 82.7 82.4 80.8 80.9 82.4 47.5
  Accuracy 95.8 95.0 94.7 95.5 95.4 95.1 96.4 95.4 90.7
Manner High 44.1 41.3 32.0 43.7 42.3 37.0 10.0 42.6 0.0
Low 17.6 13.5 14.7 18.4 17.1 13.3 0.8 17.1 0.0
Neutral 97.3 97.1 95.4 97.1 97.0 96.9 97.6 97.0 95.9
  Macro 53.0 50.6 47.4 53.1 52.1 49.1 36.1 52.2 32.0
  Accuracy 94.7 94.4 91.1 94.3 94.1 93.9 95.4 94.1 92.1
Source Current 99.1 98.6 98.5 99.1 98.9 99.1 99.3 98.9 98.6
Other 43.7 32.2 35.4 41.1 40.6 42.8 18.3 41.9 0.0
  Macro 71.4 65.4 66.9 70.1 69.8 70.9 58.8 70.4 49.3
  Accuracy 98.2 97.3 97.0 98.2 97.9 98.2 98.6 97.9 97.2
  1. “-” means that the named feature type or configuration is disabled. The F-score for each meta-knowledge value, together with the macro-averaged F-score and accuracy (= Micro-averaged F-score) for each dimension, are reported. The macro-averaged F-score is calculated by taking the average of the F-scores for each different meta-knowledge value within the dimension.