Table 1 Pseudo-code of the optimization algorithm

1)

randomly split dataset D into train D _T and development D _D datasets

2)

for each trigger type T _i ϵ T

a) for each feature type F _j ϵ F

  i) activate feature F _j on model configuration MC _i

  ii) call TrainModels with D _T , D _D , MC _i and O

  iii) if no improvement, deactivate feature F _j on model configuration MC _i

b) for each context type C _j ϵ C

  i) activate context C _j on model configuration MC _i

  ii) call TrainModels with D _T , D _D , MC _i and O

c) store best performing context on model configuration MC _i

d) for each feature with n-grams FN _j ϵ FN

  i) for each n-grams N _k ϵ N

    (1) activate n-gram N _k for feature FN _j on model configuration MC _i

    (2) call TrainModels with D _T , D _D , MC _i and O

  ii) store best performing n-gram of feature FN _j on model configuration MC _i

e) for each feature with dependency hops FH _j ϵ FH

  i) for each dependency hop H _k ϵ H

    (1) activate hop H _k for feature FH _j on model configuration MC _i

    (2) call TrainModels with D _T , D _D , MC _i and O

  ii) store best performing n-gram of feature FN _j on model configuration MC _i

f) for each feature with vertex feature type FV _j ϵ FV

  i) for each vertex feature type V _k ϵ V

    (1) activate vertex type V _k for feature FV _j on model configuration MC _i

    (2) call TrainModels with D _T , D _D , MC _i and O

  ii) store best performing vertex type of feature FV _j on model configuration MC _i

3)  Return MC

Train models (D _T , D _D , MC _i , O)

1)  for each O _j ϵ O

a) train model M on dataset D _T using MC _i

b) get performance of model M on dataset D _D

c) store performance and model order if better

2)  return better performance and order