Input: labeled document pool L, unlabeled document pool U, batch size b |
---|
// Initialization |
E R 0 = the set of events/relations annotated on L |
Learn a TEES model M 0 from E R 0 |
i = 0 // the index of the current round |
// Active Learning Loop |
while U is not empty: |
i += 1 |
for each document D i j in U: |
Document informativity score I(D i j )=0 |
for each sentence S k in D i j : |
Apply M i−1 to S k and collect the resultant events/relations set \(ER_{S_{k}}\) |
for each event/relation er s.t. er \(\notin ER_{s_k}\): |
I(D i j ) += informativity score I(S k ,e r) |
I(D i j ) = I(D i j ) / sizeOf(D i j ) |
Rank D i j in U based on I(D i j ) and select the top b documents, |
designated as B |
Remove B from U, add B to L, and add the annotations on B to E R i−1, |
designated as E R i |
Learn a new model M i from E R i |