Research Article
Medical Specialty Classification Based on Semiadversarial Data Augmentation
| | Input: Medical note text , the ground truth , target model , attack step size , synonym sets size , original dataset | | | Output: Semiadversarial examples set | | (1) | ⟵ train on | | (2) | ⟵ Sort all words in by the descending order of their importance scores via equation (1) | | (3) | Filter the stopwords from | | (4) | ⟵ length of | | (5) | Fordo | | | ⟵ | | (6) | in ⟵ the words in where index is to | | (7) | ⟵ { } | | (8) | for in do | | (9) | Initiate the candidates set by extracting the top synonyms for from WordNet using cosine similarity | | (10) | end for | | (11) | fordo | | (12) | ⟵ Randomly sample words from to | | (13) | Add to | | (14) | end for | | (15) | ⟵ | | (16) | for in do | | (17) | ⟵ Replace to of with their corresponding candidate in | | (18) | ifthen | | (19) | Add to AESet | | (20) | ⟵ | | (21) | end if | | (22) | end for | | (23) | if there exits whose prediction result Then | | (24) | return AESet | | (25) | end if | | (26) | end for | | (27) | return AESet |
|
