File size: 223,070 Bytes
6fa4bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
{
    "paper_id": "2022",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T01:11:35.471170Z"
    },
    "title": "UNIREX: A Unified Learning Framework for Language Model Rationale Extraction",
    "authors": [
        {
            "first": "Aaron",
            "middle": [],
            "last": "Chan",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Southern California",
                "location": {}
            },
            "email": "chanaaro@usc.edu"
        },
        {
            "first": "Maziar",
            "middle": [],
            "last": "Sanjabi",
            "suffix": "",
            "affiliation": {},
            "email": "maziars@fb.com"
        },
        {
            "first": "Lambert",
            "middle": [],
            "last": "Mathias",
            "suffix": "",
            "affiliation": {},
            "email": "mathiasl@fb.com"
        },
        {
            "first": "Liang",
            "middle": [],
            "last": "Tan",
            "suffix": "",
            "affiliation": {},
            "email": "liangtan@fb.com"
        },
        {
            "first": "Shaoliang",
            "middle": [],
            "last": "Nie",
            "suffix": "",
            "affiliation": {},
            "email": "snie@fb.com"
        },
        {
            "first": "Xiaochang",
            "middle": [],
            "last": "Peng",
            "suffix": "",
            "affiliation": {},
            "email": "xiaochang@fb.com"
        },
        {
            "first": "Xiang",
            "middle": [],
            "last": "Ren",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Southern California",
                "location": {}
            },
            "email": "xiangren@usc.edu"
        },
        {
            "first": "Hamed",
            "middle": [],
            "last": "Firooz",
            "suffix": "",
            "affiliation": {},
            "email": "mhfirooz@fb.com"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "An extractive rationale explains a language model's (LM's) prediction on a given task instance by highlighting the text inputs that most influenced the prediction. Ideally, rationale extraction should be faithful (reflective of LM's actual behavior) and plausible (convincing to humans), without compromising the LM's (i.e., task model's) task performance. Although attribution algorithms and select-predict pipelines are commonly used in rationale extraction, they both rely on certain heuristics that hinder them from satisfying all three desiderata. In light of this, we propose UNIREX, a flexible learning framework which generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly train the task model and rationale extractor on the task using selected objectives. UNIREX enables replacing prior works' heuristic design choices with a generic learned rationale extractor in (1) and optimizing it for all three desiderata in (2)-(3). To facilitate comparison between methods w.r.t. multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric. Across five English text classification datasets, our best UNIREX configuration outperforms the strongest baselines by an average of 32.9% NRG. Plus, we find that UNIREXtrained rationale extractors' faithfulness can even generalize to unseen datasets and tasks.",
    "pdf_parse": {
        "paper_id": "2022",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "An extractive rationale explains a language model's (LM's) prediction on a given task instance by highlighting the text inputs that most influenced the prediction. Ideally, rationale extraction should be faithful (reflective of LM's actual behavior) and plausible (convincing to humans), without compromising the LM's (i.e., task model's) task performance. Although attribution algorithms and select-predict pipelines are commonly used in rationale extraction, they both rely on certain heuristics that hinder them from satisfying all three desiderata. In light of this, we propose UNIREX, a flexible learning framework which generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly train the task model and rationale extractor on the task using selected objectives. UNIREX enables replacing prior works' heuristic design choices with a generic learned rationale extractor in (1) and optimizing it for all three desiderata in (2)-(3). To facilitate comparison between methods w.r.t. multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric. Across five English text classification datasets, our best UNIREX configuration outperforms the strongest baselines by an average of 32.9% NRG. Plus, we find that UNIREXtrained rationale extractors' faithfulness can even generalize to unseen datasets and tasks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Large neural language models (LMs) have yielded state-of-the-art performance on various natural language processing (NLP) tasks (Devlin et al., 2018; Liu et al., 2019) . However, LMs' complex reasoning processes are notoriously opaque (Rudin, 2019) , posing concerns about the societal implications of using LMs for high-stakes decision-making * Work done while AC was a research intern at Meta AI. (Bender et al., 2021) . Thus, explaining LMs' behavior is crucial for promoting trust, ethics, and safety in NLP systems (Doshi-Velez and Kim, 2017; Lipton, 2018) . Given a LM's (i.e., task model's) predicted label on a text classification instance, an extractive rationale is a type of explanation that highlights the tokens that most influenced the model to predict that label (Luo et al., 2021) . Ideally, rationale extraction should be faithful (Ismail et al., 2021; Jain et al., 2020) and plausible (DeYoung et al., 2019) , without hurting the LM's task performance (DeYoung et al., 2019) ( Fig. 1) .",
                "cite_spans": [
                    {
                        "start": 128,
                        "end": 149,
                        "text": "(Devlin et al., 2018;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 150,
                        "end": 167,
                        "text": "Liu et al., 2019)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 235,
                        "end": 248,
                        "text": "(Rudin, 2019)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 399,
                        "end": 420,
                        "text": "(Bender et al., 2021)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 520,
                        "end": 547,
                        "text": "(Doshi-Velez and Kim, 2017;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 548,
                        "end": 561,
                        "text": "Lipton, 2018)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 778,
                        "end": 796,
                        "text": "(Luo et al., 2021)",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 848,
                        "end": 869,
                        "text": "(Ismail et al., 2021;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 870,
                        "end": 888,
                        "text": "Jain et al., 2020)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 903,
                        "end": 925,
                        "text": "(DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 995,
                        "end": 1002,
                        "text": "Fig. 1)",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Configuring the rationale extractor and its training can greatly impact these desiderata, yet prior works have commonly adopted two suboptimal heuristics. First, many works rely in some way on attribution algorithms (AAs), which extract rationales via handcrafted functions (Sundararajan et al., 2017; Ismail et al., 2021; Situ et al., 2021) . AAs cannot be directly trained and tend to be computeintensive (Bastings and Filippova, 2020) . Also, AAs can be a bottleneck for plausibility, as producing human-like rationales is a complex objec-tive requiring high capacity rationale extractors (Narang et al., 2020; DeYoung et al., 2019) . Second, many works use a specialized select-predict pipeline (SPP), where a predictor module is trained to solve the task using only tokens chosen by a selector module (Jain et al., 2020; Yu et al., 2021; . Instead of faithfulness optimization, SPPs heuristically aim for \"faithfulness by construction\" by treating the selected tokens as a rationale for the predictor's output (which depends only on those tokens). Still, SPPs typically have worse task performance than vanilla LMs since SPPs hide the full input from the predictor.",
                "cite_spans": [
                    {
                        "start": 274,
                        "end": 301,
                        "text": "(Sundararajan et al., 2017;",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 302,
                        "end": 322,
                        "text": "Ismail et al., 2021;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 323,
                        "end": 341,
                        "text": "Situ et al., 2021)",
                        "ref_id": "BIBREF42"
                    },
                    {
                        "start": 407,
                        "end": 437,
                        "text": "(Bastings and Filippova, 2020)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 592,
                        "end": 613,
                        "text": "(Narang et al., 2020;",
                        "ref_id": null
                    },
                    {
                        "start": 614,
                        "end": 635,
                        "text": "DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 806,
                        "end": 825,
                        "text": "(Jain et al., 2020;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 826,
                        "end": 842,
                        "text": "Yu et al., 2021;",
                        "ref_id": "BIBREF52"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "To tackle this challenge, we propose the UNIfied Learning Framework for Rationale EXtraction (UNIREX), which generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly train the task model and rationale extractor on the task using selected objectives (Sec. 3). UNIREX enables replacing prior works' heuristic design choices in (1) with a generic learned rationale extractor and optimizing it for all three desiderata in (2)-(3).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "UNIREX provides significant flexibility in performing (1)-(3). For (1), any model architecture is applicable, but we study Transformer LM based rationale extractors in this work (Zaheer et al., 2020; DeYoung et al., 2019) . We focus on two architectures: (A) Dual LM, where task model and rationale extractor are separate and (B) Shared LM, where task model and rationale extractor share parameters. For (2), any faithfulness and plausibility criteria can be used. Following DeYoung et al. (2019), we focus on comprehensiveness and sufficiency as faithfulness criteria, while using similarity to gold rationales as plausibility criteria. For (3), trade-offs between the three desiderata can be easily managed during rationale extractor optimization by setting arbitrary loss weights for the faithfulness and plausibility objectives. Plus, though computing the faithfulness criteria involves discrete (nondifferentiable) token selection, using Shared LM can approximate end-to-end training and enable both task model and rationale extractor to be optimized w.r.t. all three desiderata (Sec. 3.3).",
                "cite_spans": [
                    {
                        "start": 178,
                        "end": 199,
                        "text": "(Zaheer et al., 2020;",
                        "ref_id": "BIBREF53"
                    },
                    {
                        "start": 200,
                        "end": 221,
                        "text": "DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "To evaluate all three desiderata in aggregate, we introduce the Normalized Relative Gain (NRG) metric. Across five English text classification datasets -SST, Movies, CoS-E, MultiRC, and e-SNLI (Carton et al., 2020; DeYoung et al., 2019 )our best UNIREX configuration outperforms the strongest baselines by an average of 32.9% NRG (Sec. 4.2) , showing that UNIREX can optimize rationale extractors for all three desiderata. In addition, we verify our UNIREX design choices via extensive ablation studies (Sec. 4.3) . Furthermore, UNIREX-trained extractors have high generalization power, yielding high plausiblity with minimal gold rationale supervision (Sec. 4.4) and high faithfulness on unseen datasets and tasks (Sec. 4.5).",
                "cite_spans": [
                    {
                        "start": 193,
                        "end": 214,
                        "text": "(Carton et al., 2020;",
                        "ref_id": null
                    },
                    {
                        "start": 215,
                        "end": 235,
                        "text": "DeYoung et al., 2019",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 330,
                        "end": 340,
                        "text": "(Sec. 4.2)",
                        "ref_id": null
                    },
                    {
                        "start": 503,
                        "end": 513,
                        "text": "(Sec. 4.3)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Finally, our user study shows that humans judge UNIREX rationales as more plausible than rationales extracted using other methods (Sec. 4.6).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Rationale Extraction Let F task = f task (f enc (\u2022)) be a task model for M -class text classification (Sec. A.1), where f enc is the text encoder and f task is the task output head. Typically, F task has a BERT-style architecture (Devlin et al., 2018) , in which f enc is a Transformer (Vaswani et al., 2017) while f task is a linear layer with softmax classifier. Let",
                "cite_spans": [
                    {
                        "start": 230,
                        "end": 251,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 286,
                        "end": 308,
                        "text": "(Vaswani et al., 2017)",
                        "ref_id": "BIBREF48"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "x i = [x t i ]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "n t=1 be the n-token input sequence (e.g., a sentence) for task instance i, and F task (x i ) \u2208 R M be the logit vector for the output of the task model. Let\u0177 i = arg max j F task (x i ) j be the class predicted by F task . Given F task , x i , and\u0177 i , the goal of rationale extraction is to output vector",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "s i = [s t i ] n t=1 \u2208 R n , such that each s t i \u2208 R",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "is an importance score indicating how much token x t i influenced F task to predict class\u0177 i . Let F ext be a rationale extractor, such that",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "s i = F ext (F task , x i ,\u0177 i ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "F ext can be a learned or heuristic function. In practice, the final rationale is often obtained by binarizing s i as r i \u2208 {0, 1} n , via the top-k% strategy: r t i = 1 if s t i is one of the top-k% scores in s i ; otherwise, r t i = 0 (DeYoung et al., 2019; Jain et al., 2020; Pruthi et al., 2020; Chan et al., 2021) . For top-k%, let r (k) i be the \"important\" (i.e., ones) tokens in r i , when using 0 \u2264 k \u2264 100.",
                "cite_spans": [
                    {
                        "start": 237,
                        "end": 259,
                        "text": "(DeYoung et al., 2019;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 260,
                        "end": 278,
                        "text": "Jain et al., 2020;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 279,
                        "end": 299,
                        "text": "Pruthi et al., 2020;",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 300,
                        "end": 318,
                        "text": "Chan et al., 2021)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "Faithfulness means how well a rationale reflects F task 's true reasoning process for predicting\u0177 i (Jacovi and Goldberg, 2020) . Hence, faithfulness metrics measure how much the r (k) i tokens impact p\u0177 i (x i ), which denotes F task 's confidence probability for\u0177 i when using x i as input (DeYoung et al., 2019; Shrikumar et al., 2017; Hooker et al., 2018; Pruthi et al., 2020) . Recently, comprehensiveness and sufficiency have emerged as popular faithfulness metrics (DeYoung et al., 2019) . Comprehensiveness (comp) measures the change in p\u0177 i when r (k) i is removed from the input:",
                "cite_spans": [
                    {
                        "start": 100,
                        "end": 127,
                        "text": "(Jacovi and Goldberg, 2020)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 292,
                        "end": 314,
                        "text": "(DeYoung et al., 2019;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 315,
                        "end": 338,
                        "text": "Shrikumar et al., 2017;",
                        "ref_id": "BIBREF40"
                    },
                    {
                        "start": 339,
                        "end": 359,
                        "text": "Hooker et al., 2018;",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 360,
                        "end": 380,
                        "text": "Pruthi et al., 2020)",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 472,
                        "end": 494,
                        "text": "(DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "comp = p\u0177 i (x i ) \u2212 p\u0177 i (x i \\r (k) i ). Sufficiency (suff) measures the change in p\u0177 i when only r (k) i is kept in the input: suff = p\u0177 i (x i ) \u2212 p\u0177 i (r (k) i ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "High faithfulness is signaled by high comp and low suff.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "Plausibility means how convincing a rationale is to humans (Jacovi and Goldberg, 2020) . This can be measured by automatically computing the similarity between F ext 's rationales (either s i or r i ) and human-annotated gold rationales (DeYoung et al., 2019) , or by asking human annotators to rate whether F ext 's rationales make sense for predicting\u0177 i (Strout et al., 2019; Doshi-Velez and Kim, 2017) . Typically, a gold rationale is a binary vector r * i \u2208 {0, 1} n , where ones/zeros indicate important/unimportant tokens (Lei et al., 2016) .",
                "cite_spans": [
                    {
                        "start": 59,
                        "end": 86,
                        "text": "(Jacovi and Goldberg, 2020)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 237,
                        "end": 259,
                        "text": "(DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 357,
                        "end": 378,
                        "text": "(Strout et al., 2019;",
                        "ref_id": "BIBREF44"
                    },
                    {
                        "start": 379,
                        "end": 405,
                        "text": "Doshi-Velez and Kim, 2017)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 529,
                        "end": 547,
                        "text": "(Lei et al., 2016)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "Task Performance, w.r.t. rationale extraction, concerns how much F task 's task performance (on test set) drops when F task is trained with explainability objectives (i.e., faithfulness, plausibility) for F ext . As long as F task is trained with non-task losses, F task 's task performance can be affected.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "Given task model F task , UNIREX generalizes rationale extractor optimization as follows: (1) choose architecture for a learned rationale extractor F ext ; (2) select explainability objectives (i.e., faithfulness loss L faith and plausibility loss L plaus ); and (3) jointly train F task and F ext using L task (task loss), L faith , and L plaus . UNIREX training consists of two backpropagation paths (Fig. 2) . The first path is used to update F task w.r.t. L task and L faith . Whereas L task is computed w.r.t. the task target y i , L faith is computed only using the task input x i and the top-k% important tokens r (k) i (obtained via F ext ), based on some combination of comp and suff (Sec. 2). The second path is used to update F ext w.r.t. L plaus , which encourages importance scores s i to approximate gold rationale r * i . Thus, UNIREX frames rationale extraction as the following optimization problem:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 402,
                        "end": 410,
                        "text": "(Fig. 2)",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "UNIREX",
                "sec_num": "3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "min F task , Fext L task (x i , y i ; F task ) + \u03b1 f L faith (x i , r (k) i ; F task ) + \u03b1 p L plaus (x i , r * i ; F ext ),",
                        "eq_num": "(1)"
                    }
                ],
                "section": "UNIREX",
                "sec_num": "3"
            },
            {
                "text": "where \u03b1 f and \u03b1 p are loss weights. If F task and F ext share parameters, then the shared parameters will be optimized w.r.t. all losses. During inference, for task input x i , we first use F task to predict y i , then use F ext to output a rationale r i for F task 's prediction\u0177 i . Below, we discuss options for the rationale extractor and explainability objectives.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "UNIREX",
                "sec_num": "3"
            },
            {
                "text": "In UNIREX, F ext is a learned function by default. Learned F ext can be any model that transforms x t i into s t i . Given their success in NLP explainability (DeYoung et al., 2019), we focus on pre-trained Transformer LMs and highlight two architectures: Dual LM (DLM) and Shared LM (SLM) (Fig. 3) . For DLM, F task and F ext are two separate Transformer LMs. DLM provides more dedicated capacity for F ext , which can help F ext output plausible rationales. For SLM, F task and F ext are two Transformer LMs sharing encoder f enc , while F ext has its own output head f ext . SLM leverages multitask learning between F task and F ext , which can improve faithfulness since F ext gets more information about F task 's reasoning process. Unlike heuristic F ext (Sec. A.2), learned F ext can be optimized for faithfulness/plausibility, but cannot be used out of the box without training. Learned F ext is preferred if: (A) optimizing for both faithfulness and plausibility, and (B) gold rationales are available for plausibility optimization (Sec. A.3).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 290,
                        "end": 298,
                        "text": "(Fig. 3)",
                        "ref_id": "FIGREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Rationale Extractor",
                "sec_num": "3.1"
            },
            {
                "text": "After selecting F ext , we specify the explainability objectives, which can be any combination of faithfulness and plausibility criteria. In prior approaches (e.g., AA, SPPs), the rationale extractor is not optimized for both faithfulness and plausibility, but UNIREX makes this possible. For any choice of learned F ext , UNIREX lets us easily \"plug and play\" different criteria and loss weights, based on our needs and domain knowledge, to find those that best balance the rationale extraction desiderata.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "Faithfulness Evaluating rationale faithfulness is still an open problem with many existing metrics, and UNIREX is not tailored for any specific metric. Still, given the prevalence of comp/suff (Sec. 2), we focus on comp/suff based objectives.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "Recall that comp measures the importance of tokens in r (k) i as how p\u0177 i (x i ), F task 's predicted probability for class\u0177 i , changes when those tokens are removed from x i . Intuitively, we want p\u0177 i (x i ) to be higher than p\u0177 i (x i \\r (k) i ), so higher comp is better. Since comp is defined for a single class' probability rather than the label distribution, we can define the comp loss L comp via cross-entropy loss L CE , as in the following difference criterion for L comp :",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Lcomp-diff = LCE(Ftask(xi), yi) \u2212 LCE(Ftask(xi\\r (k) i ), yi)) (2) L CE (F task (xi), yi) = \u2212yi log(F task (xi))",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "For training stability, we compute comp loss for target class y i here instead of F task 's predicted class\u0177 i , since\u0177 i is a moving target during training. Using L comp-diff , it is possible for",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "L CE (F task (x i \\r (k) i ), y i )) to become much larger than L CE (F task (x i ), y i )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": ", leading to arbitrarily negative losses. To avoid this, we can add margin m c to the loss function, giving the margin criterion:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "Lcomp-margin = max(\u2212mc, LCE(Ftask(xi), yi) \u2212 LCE(Ftask(xi\\r (k) i ), yi)) + mc (4)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "Recall that suff measures the importance of tokens in r (k) i as how p\u0177 i (x i ), F task 's predicted probability for class\u0177 i , changes when they are the only tokens kept in x i . Based on suff's definition, we want p\u0177 i (r (k) i ) to be higher than p\u0177 i (x i ), so lower suff is better. For suff loss L suff , we define the difference and margin criteria analogously with margin m s but the opposite sign (since lower suff is better):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Lsuff-diff = LCE(Ftask(r (k) i ), yi) \u2212 LCE(Ftask(xi), yi) (5) Lsuff-margin = max(\u2212ms, LCE(Ftask(r (k) i ), yi) \u2212 LCE(Ftask(xi), yi)) + ms",
                        "eq_num": "(6)"
                    }
                ],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "In our experiments, we find that the marginbased comp/suff criteria are effective (Sec. 4.3), though others (e.g., KL Div, MAE) can be used too (Sec. A.4.1). Note that r (k) i is computed via top-k% thresholding (Sec. 2), so we also need to specify a set K of threshold values. We separately compute the comp/suff losses for each k \u2208 K, then obtain the final comp/suff losses by averaging over all k values via area-over-precision-curve (AOPC) (DeYoung et al., 2019). To reflect this, we denote the comp and suff losses as L comp,K and L suff,K , respectively. Let",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "\u03b1 f L faith = \u03b1 c L comp,K + \u03b1 s L suff,K ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "where \u03b1 c and \u03b1 s are loss weights.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "Plausibility Plausibility is defined as how convincing a rationale is to humans (Jacovi and Goldberg, 2020), i.e., whether humans would agree the rationale supports the model's prediction. While optimizing for plausibility should ideally involve human-in-the-loop feedback, this is prohibitive. Instead, many works consider gold rationales as a cheaper form of plausibility annotation (DeYoung et al., 2019; Narang et al., 2020; Jain et al., 2020) . Thus, if gold rationale supervision is available, then we can optimize for plausibility. With gold rationale r * i for input x i , plausibility optimization entails training F ext to predict binary importance label r * ,t i for each token x t i . This is essentially token classification, so one natural choice for L plaus is the token-level binary cross-entropy (BCE) criterion:",
                "cite_spans": [
                    {
                        "start": 385,
                        "end": 407,
                        "text": "(DeYoung et al., 2019;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 408,
                        "end": 428,
                        "text": "Narang et al., 2020;",
                        "ref_id": null
                    },
                    {
                        "start": 429,
                        "end": 447,
                        "text": "Jain et al., 2020)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Lplaus-BCE = \u2212 t r * ,t i log(Fext(x t i ))",
                        "eq_num": "(7)"
                    }
                ],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "Besides BCE loss, we can also consider other criteria like sequence-level KL divergence and L1 loss. See Sec. A.4.2 for discussion of these and other plausibility criteria.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Explainability Objectives",
                "sec_num": "3.2"
            },
            {
                "text": "After setting F ext , L faith , and L plaus , we can move on to training F task and F ext . Since top-k% rationale binarization (Sec. 3.2) is not differentiable, by default, we cannot backpropagate L faith through all of F ext 's parameters. Thus, F task is trained via L task and L faith , while F ext is only trained via L plaus . This means F ext 's rationales r i are indirectly optimized for faithfulness by regularizing F task such that its behavior aligns with r i . The exception is if we are using the SLM variant, where encoder f enc is shared by F task and F ext . In this case, f enc is optimized w.r.t. all losses, f task is optimized w.r.t. L task and L faith , and f ext is optimized w.r.t. L plaus . SLM is a simple way to approximate end-to-end training of F task and F ext . In contrast, past SPPs have used more complex methods like reinforcement learning (Lei et al., 2016) and the reparameterization trick (Bastings et al., 2019) , whose training instability can hurt task performance (Jain et al., 2020) . Now, we summarize the full learning objective. Given that cross-entropy loss L task = L CE (F task (x i ), y i ) is used to train F task to predict y i , the full learning objective is:",
                "cite_spans": [
                    {
                        "start": 875,
                        "end": 893,
                        "text": "(Lei et al., 2016)",
                        "ref_id": null
                    },
                    {
                        "start": 927,
                        "end": 950,
                        "text": "(Bastings et al., 2019)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 1006,
                        "end": 1025,
                        "text": "(Jain et al., 2020)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training and Inference",
                "sec_num": "3.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "L = Ltask + \u03b1 f Lfaith + \u03b1pLplaus = Ltask + \u03b1cLcomp,K + \u03b1sLsuff,K + \u03b1pLplaus.",
                        "eq_num": "(8)"
                    }
                ],
                "section": "Training and Inference",
                "sec_num": "3.3"
            },
            {
                "text": "During inference, we use F task to predict y i , then use F ext to output r i for F task 's predicted label\u0177 i .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training and Inference",
                "sec_num": "3.3"
            },
            {
                "text": "We present empirical results demonstrating UNIREX's effectiveness in managing trade-offs between faithfulness, plausibility, and task performance during rationale extractor optimization. First, our main experiments compare methods w.r.t. faithfulness, plausibility, and task performance (Sec. 4.2). Second, we perform various ablation studies to verify our design choices for UNIREX (Sec. 4.3). Third, we present experiments highlighting UNIREX's generalization ability, both in terms of limited gold rationale supervision (Sec. 4.4) and zero-shot transfer (Sec. 4.5). Fourth, we conduct a user study to further evaluate UNIREX rationales' plausibility, relative to those generated by other methods (Sec. 4.6). See Sec. A.5 for implementation details (LM architecture, AA settings, training).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "4"
            },
            {
                "text": "Datasets We primarily use SST (Socher et al., 2013; Carton et al., 2020) , Movies (Zaidan and Eisner, 2008) , CoS-E (Rajani et al., 2019), Mul-tiRC (Khashabi et al., 2018) , and e-SNLI (Camburu et al., 2018) , all of which have gold rationale annotations. The latter four datasets were taken from the ERASER benchmark (DeYoung et al., 2019).",
                "cite_spans": [
                    {
                        "start": 30,
                        "end": 51,
                        "text": "(Socher et al., 2013;",
                        "ref_id": "BIBREF43"
                    },
                    {
                        "start": 52,
                        "end": 72,
                        "text": "Carton et al., 2020)",
                        "ref_id": null
                    },
                    {
                        "start": 82,
                        "end": 107,
                        "text": "(Zaidan and Eisner, 2008)",
                        "ref_id": "BIBREF54"
                    },
                    {
                        "start": 148,
                        "end": 171,
                        "text": "(Khashabi et al., 2018)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 185,
                        "end": 207,
                        "text": "(Camburu et al., 2018)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiment Setup",
                "sec_num": "4.1"
            },
            {
                "text": "We use the metrics from the ERASER explainability benchmark (DeYoung et al., 2019) . For faithfulness, we use comprehensiveness (Comp) and sufficiency (Suff), for k = [1, 5, 10, 20, 50] (DeYoung et al., 2019). For plausibility, we use area under precision-recall curve (AUPRC) and token F1 (TF1) to measure similarity to gold rationales (DeYoung et al., 2019; Narang et al., 2020) . For task performance, we follow (DeYoung et al., 2019) and (Carton et al., 2020) in using accuracy (SST, CoS-E) and macro F1 (Movies, MultiRC, e-SNLI).",
                "cite_spans": [
                    {
                        "start": 60,
                        "end": 82,
                        "text": "(DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 337,
                        "end": 359,
                        "text": "(DeYoung et al., 2019;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 360,
                        "end": 380,
                        "text": "Narang et al., 2020)",
                        "ref_id": null
                    },
                    {
                        "start": 442,
                        "end": 463,
                        "text": "(Carton et al., 2020)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": "To aggregately evaluate multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric, which is based on the ARG metric from Ye et al. (2021) . NRG normalizes raw metrics (e.g., F1, sufficiency) to scores between 0 and 1 (higher is better). Given a set of raw metric scores Z = {z 1 , z 2 , ...} (each from a different method), NRG(z i ) captures z i 's value relative to min(Z) and max(Z). If higher values are better for the given metric (e.g., F1), then we have:",
                "cite_spans": [
                    {
                        "start": 139,
                        "end": 155,
                        "text": "Ye et al. (2021)",
                        "ref_id": "BIBREF50"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": "NRG(z i ) = z i \u2212min(Z) max(Z)\u2212min(Z)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": ". If lower values are better (e.g., sufficiency), then we have:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": "NRG(z i ) = max(Z)\u2212z i max(Z)\u2212min(Z)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": ". After computing NRG for multiple raw metrics, we can aggregate them w.r.t. desiderata via averaging. Let FNRG, PNRG, and TNRG be the NRG values for faithfulness, plausibility, and task performance, respectively. Finally, we compute the composite NRG as:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": "CNRG = FNRG+PNRG+TNRG 3",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": ". Results Reporting For all results, we report average over three seeds and the five k values. We Figure 4 : Composite NRG Comparison (w/o Plausibility Optimization). Composite NRG (CNRG) is the mean of the three desiderata NRG scores. For each dataset, we use CNRG to compare methods that do not optimize for plausibility. Figure 5 : Composite NRG Comparison (w/ Plausibility Optimization). Composite NRG (CNRG) is the mean of the three desiderata NRG scores. For each dataset, we use CNRG to compare methods that do optimize for plausibility. denote each UNIREX configuration with \"([rationale extractor]-[explainability objectives])\". F, P, and FP denote faithfulness, plausibility, and faith-fulness+plausibility, respectively.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 98,
                        "end": 106,
                        "text": "Figure 4",
                        "ref_id": null
                    },
                    {
                        "start": 324,
                        "end": 332,
                        "text": "Figure 5",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": "Baselines The first category is AAs, which are not trained: AA (Grad) (Simonyan et al., 2013) , AA (Input*Grad) (Denil et al., 2014) , AA (DeepLIFT) (Lundberg and Lee, 2017), AA (IG) (Sundararajan et al., 2017) . We also experiment with IG for L2E (Situ et al., 2021) , which distills knowledge from an AA to an LM. The second category is SPPs: FRESH (Jain et al., 2020) and A2R (Yu et al., 2021) . For FRESH, we use a strong variant where IG rationales are directly given to the predictor, rather than output by a trained selector. A2R aims to improve SPP task performance by regularizing the predictor with an attention-based predictor that uses the full input. In addition, we introduce FRESH+P and A2R+P, which augment FRESH and A2R, respectively, with plausibility optimization. The third category is AA-based regularization: SGT (Ismail et al., 2021) , which uses a sufficiency-based criterion to optimize for faithfulness. We also consider SGT+P, which augments SGT with plausibility optimization. Fig. 4 -6 display the main results. In Fig. 4 /5, we compare the CNRG for all methods and datasets, without/with gold rationales. In both plots, we see that UNIREX variants achieve the best CNRG across all datasets, indicating that they are effective in balancing the three desiderata. In particular, UNIREX (DLM-FP) and UNIREX (SLM- FP) have very high CNRG scores, both yielding more than 30% improvement over the strongest baselines. Fig. 6 compares methods w.r.t. desiderata NRG (i.e., FNRG, PNRG, TNRG). Here, the left/right plots show methods without/with gold rationales. Again, we see that UNIREX variants achieve a good NRG balance of faithfulness, plausibility, and task performance. Meanwhile, many baselines (e.g., AA (IG), A2R, SGT+P) do well on some desiderata but very poorly on others.",
                "cite_spans": [
                    {
                        "start": 70,
                        "end": 93,
                        "text": "(Simonyan et al., 2013)",
                        "ref_id": "BIBREF41"
                    },
                    {
                        "start": 112,
                        "end": 132,
                        "text": "(Denil et al., 2014)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 183,
                        "end": 210,
                        "text": "(Sundararajan et al., 2017)",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 248,
                        "end": 267,
                        "text": "(Situ et al., 2021)",
                        "ref_id": "BIBREF42"
                    },
                    {
                        "start": 351,
                        "end": 370,
                        "text": "(Jain et al., 2020)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 379,
                        "end": 396,
                        "text": "(Yu et al., 2021)",
                        "ref_id": "BIBREF52"
                    },
                    {
                        "start": 835,
                        "end": 856,
                        "text": "(Ismail et al., 2021)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1005,
                        "end": 1011,
                        "text": "Fig. 4",
                        "ref_id": null
                    },
                    {
                        "start": 1044,
                        "end": 1050,
                        "text": "Fig. 4",
                        "ref_id": null
                    },
                    {
                        "start": 1441,
                        "end": 1447,
                        "text": "Fig. 6",
                        "ref_id": "FIGREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Metrics",
                "sec_num": null
            },
            {
                "text": "We present five ablation studies to validate the effectiveness of our UNIREX design choices. The ablation results are displayed in Table 1 . In this table, each of the five sections shows results for a different ablation. Thus, all numbers within the same section and column are comparable.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 131,
                        "end": 138,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Ablation Studies",
                "sec_num": "4.3"
            },
            {
                "text": "Extractor Type In the Ext Type (F) section, we compare four heuristic rationale extractors, using AA-F. Rand uses random importance scores, Gold directly uses the gold rationales, Inv uses the inverse of the gold rationales, and IG uses IG. All heuristics yield similar task performance, but IG dominates on all faithfulness metrics. This makes sense because IG is computed using F task 's inputs/parameters/outputs, while the others do not have this information. For plausibility, Gold is the best, Inv is the worst, and Rand and IG are about the same, as none of the heuristics are optimized for plausibility. In the Ext Type (FP) section, we compare four learned rationale extractors. By default, attribution algorithms' dimension scores are pooled into token scores via sum pooling. AA-FP (Sum) uses IG with sum pooling, while AA-FP (MLP) replaces the sum pooler with a MLP-based pooler to increase capacity for plausibility optimization. Task performance for all four methods is similar, AA-FP (Sum) dominates on faithfulness, and DLM-FP and SLM-FP dominate on plausibility. AA-FP (MLP) does not perform as well on faithfulness but slightly improves on plausibility compared to AA-FP (Sum).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ablation Studies",
                "sec_num": "4.3"
            },
            {
                "text": "Comp/Suff Losses The Comp/Suff Loss section compares different combinations of Comp and Suff losses, using SLM-FP. Note that SLM-FP (Comp+Suff) is equivalent to SLM-FP shown in other tables/sections. As expected, SLM-FP (Comp) does best on Comp, but SLM-FP (Comp+Suff) actually does best on Suff. Meanwhile, SLM-FP, (Suff) does second-best on Suff but is much worse on Comp. This shows that Comp and Suff are complementary for optimization.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ablation Studies",
                "sec_num": "4.3"
            },
            {
                "text": "Suff Criterion The Suff Criterion section compares different Suff criteria, using SLM-FP. SLM-FP (KLDiv) uses the KL divergence criterion, SLM-FP (MAE) uses the MAE criterion, and SLM-FP (Margin) uses the margin criterion. SLM-FP (Margin) is equivalent to SLM-FP in other ta- bles/sections. All criteria yield similar performance and plausibility, while Margin is slightly better on faithfulness.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ablation Studies",
                "sec_num": "4.3"
            },
            {
                "text": "SLM Extractor Head The SLM Ext Head section compares different extractor heads, using SLM-FP. Linear is the default choice and uses a linear layer. MLP-2048-2 uses a MLP with two 2048-dim hidden layers. MLP-4096-3 uses a MLP with three 4096-dim hidden layers. All three output head types yield similar performance, but decreasing head capacity yields better faithfulness, while increasing head capacity heads yields better plausibility. This trades off faithfulness and plausibility, although larger heads will be more compute-intensive.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ablation Studies",
                "sec_num": "4.3"
            },
            {
                "text": "UNIREX supports arbitrary amounts of gold rationale supervision and allows us to account for data efficiency. In Fig. 7 , we compare plausibility (in AUPRC) for \u03b3 = [0.5, 1, 5, 10, 20, 100] (i.e., % of train instances with gold rationales). We compare AA (IG) and four UNIREX variants (AA-F, AA-FP, DLM-FP, SLM-FP). AA (IG) and AA-F do not use gold rationales and thus have the same AUPRC for all \u03b3. Standard deviation is shown by the error bands. UNIREX (DLM-FP) and UNIREX (SLM-FP) dominate across all \u03b3 values, with AUPRC slowly decreasing as \u03b3 decreases. Even at \u03b3 = 0.5, they can still achieve high AUPRC. This suggests that UNIREX's gold rationale batching procedure (Sec. A.3) is effective for learning from minimal gold rationale supervision and demonstrates how UNIREX enables us to manage this trade-off. See Sec. A.6 for similar results on CoS-E. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 113,
                        "end": 119,
                        "text": "Fig. 7",
                        "ref_id": "FIGREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Gold Rationale Data Efficiency",
                "sec_num": "4.4"
            },
            {
                "text": "In Table 2 , we investigate if F ext 's faithfulness, via UNIREX training on some source dataset, can generalize to unseen target datasets/tasks in a zero-shot setting (i.e., no fine-tuning on target datasets). Plausibility is not evaluated here, since these unseen datasets do not have gold rationales. As the source model, we compare various SST-trained models: AA (IG) and UNIREX (AA-F, DLM-FP). First, we evaluate on unseen datasets for a seen task (sentiment analysis (SA)): Yelp (Zhang et al., 2015) and Amazon (McAuley and Leskovec, 2013) . Second, we evaluate on unseen datasets for unseen tasks: Stormfront (hate speech detection (HSD), binary F1) (de Gibert et al., 2018), OffenseEval (offensive speech detection (OSD), macro F1) (Zampieri et al., 2019) , and SemEval2018 (irony detection (ID), binary F1) (Van Hee et al., 2018). We want to show that, even if F task yields poor task performance on unseen datasets, F ext 's rationales can still be faithful. As expected, all methods achieve much lower task performance in the third setting than in the first two settings. However, faithfulness does not appear to be strongly correlated with task performance, as unseen tasks' comp/suff scores are similar to seen tasks'. Across all datasets, DLM-FP has the best faithfulness and is the only method whose comp is always higher than suff. AA-F is not as consistently strong as DLM-FP, but almost always beats AA (IG) on comp and suff. Meanwhile, AA (IG) has the worst comp and suff overall. Ultimately, these results suggest that UNIREX-trained models' faithfulness (i.e., alignment between F task 's and F ext 's outputs) is a dataset/task agnostic property (i.e., can generalize across datasets/tasks), further establishing UNIREX's utility in low-resource settings. ",
                "cite_spans": [
                    {
                        "start": 485,
                        "end": 505,
                        "text": "(Zhang et al., 2015)",
                        "ref_id": "BIBREF56"
                    },
                    {
                        "start": 510,
                        "end": 545,
                        "text": "Amazon (McAuley and Leskovec, 2013)",
                        "ref_id": null
                    },
                    {
                        "start": 740,
                        "end": 763,
                        "text": "(Zampieri et al., 2019)",
                        "ref_id": "BIBREF55"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 3,
                        "end": 10,
                        "text": "Table 2",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Zero-Shot Faithfulness Transfer",
                "sec_num": "4.5"
            },
            {
                "text": "Gold rationale based plausibility evaluation is noisy because gold rationales are for the target label, not a model's predicted label. Thus, we conduct two five-annotator user studies (Table 3) to get a better plausibility measurement. Given 50 random test instances from SST, we get the rationales for SGT+P, A2R+P, UNIREX (AA-FP), and UNIREX (DLM-FP), plus the gold rationales. For each instance, we threshold all rationales to have the same number of positive tokens as the gold rationale. The first user study is forward simulation (Hase and Bansal, 2020; Jain et al., 2020) . Here, the annotator is given an input and a rationale for some model's prediction, then asked what (binary) sentiment label the model most likely predicted. For forward simulation, we also consider a No Rationale baseline, where no tokens are highlighted. For No Rationale and Gold, the target label is the correct choice. Annotators are also asked to rate their confidence (4point Likert scale) in their answer to this question. The second user study involves giving a subjective rating of how plausible the rationale is (Hase and Bansal, 2020) . Here, the annotator is given the input, rationale, and model's predicted label, then asked to rate (5-point Likert scale) how aligned the rationale is with the prediction. In both forward simulation and subjective rating, we find that DLM-FP performs best among all non-oracle methods and even beats Gold on accuracy, further supporting that DLM-FP rationales are plausible. As expected, the fact that Gold does not achieve near-100% accuracy shows the discrepancy between evaluating plausibility based on the target label (i.e., gold rationale similarity) and F task 's predicted label (forward simulation). Meanwhile, SGT+P and AA-FP, which had lower AUPRC/TF1 in our automatic evaluation, also do worse in accuracy/alignment. Also, users found SGT+P and AA-FP rationales harder to understand, as shown by their lower confidence scores. Meanwhile, A2R+P had high AUPRC/TF1, but gets very low accuracy/alignment because A2R+P's predicted label often not the target label, leading to misalignment with its gold-like rationale. A2R+P is a great example of how automatic plausibility evaluation can be misleading. For the accuracy, confidence, and alignment questions, we achieved Fleiss' Kappa (Fleiss, 1971) inter-annotator agreement scores of 0.2456 (fair), 0.1282 (slight), and, 0.1561 (slight), respectively. This lack of agreement shows the difficulty of measuring plausibility.",
                "cite_spans": [
                    {
                        "start": 536,
                        "end": 559,
                        "text": "(Hase and Bansal, 2020;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 560,
                        "end": 578,
                        "text": "Jain et al., 2020)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 1103,
                        "end": 1126,
                        "text": "(Hase and Bansal, 2020)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 2322,
                        "end": 2336,
                        "text": "(Fleiss, 1971)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 184,
                        "end": 193,
                        "text": "(Table 3)",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "User Study on Plausibility",
                "sec_num": "4.6"
            },
            {
                "text": "Faithfulness Many prior works have tried to improve the faithfulness of extractive rationales through the use of AAs (Bastings and Filippova, 2020) . Typically, this involves designing gradientbased (Sundararajan et al., 2017; Denil et al., 2014; Lundberg and Lee, 2017; Li et al., 2015) or perturbation-based (Li et al., 2016; Poerner et al., 2018; K\u00e1d\u00e1r et al., 2017) AAs. However, attribution algorithms cannot be optimized and tend to be compute-intensive (often requiring multiple LM forward/backward passes). Recently, Ismail et al. (2021) addressed the optimization issue by regularizing the task model to yield faithful rationales via the AA, while other works (Situ et al., 2021; Schwarzenberg et al., 2021) addressed the compute cost issue by training an LM (requiring only one forward pass) to mimic an AA's behavior. Another line of work aims to produce faithful rationales by construction, via SPPs (Jain et al., 2020; Yu et al., 2021; Bastings et al., 2019; Yu et al., 2019; Lei et al., 2016) . Still, SPPs' faithfulness can only guarantee sufficiency -not comprehensiveness (DeYoung et al., 2019) . Also, SPPs generally perform worse than vanilla LMs because they hide much of the original text input from the predictor and are hard to train end-to-end.",
                "cite_spans": [
                    {
                        "start": 117,
                        "end": 147,
                        "text": "(Bastings and Filippova, 2020)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 199,
                        "end": 226,
                        "text": "(Sundararajan et al., 2017;",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 227,
                        "end": 246,
                        "text": "Denil et al., 2014;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 247,
                        "end": 270,
                        "text": "Lundberg and Lee, 2017;",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 271,
                        "end": 287,
                        "text": "Li et al., 2015)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 310,
                        "end": 327,
                        "text": "(Li et al., 2016;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 328,
                        "end": 349,
                        "text": "Poerner et al., 2018;",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 350,
                        "end": 369,
                        "text": "K\u00e1d\u00e1r et al., 2017)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 525,
                        "end": 545,
                        "text": "Ismail et al. (2021)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 669,
                        "end": 688,
                        "text": "(Situ et al., 2021;",
                        "ref_id": "BIBREF42"
                    },
                    {
                        "start": 689,
                        "end": 716,
                        "text": "Schwarzenberg et al., 2021)",
                        "ref_id": "BIBREF39"
                    },
                    {
                        "start": 912,
                        "end": 931,
                        "text": "(Jain et al., 2020;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 932,
                        "end": 948,
                        "text": "Yu et al., 2021;",
                        "ref_id": "BIBREF52"
                    },
                    {
                        "start": 949,
                        "end": 971,
                        "text": "Bastings et al., 2019;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 972,
                        "end": 988,
                        "text": "Yu et al., 2019;",
                        "ref_id": "BIBREF51"
                    },
                    {
                        "start": 989,
                        "end": 1006,
                        "text": "Lei et al., 2016)",
                        "ref_id": null
                    },
                    {
                        "start": 1089,
                        "end": 1111,
                        "text": "(DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "5"
            },
            {
                "text": "Plausibility Existing approaches for improving extractive rationale plausibility typically involve supervising LM-based extractors (Bhat et al., 2021) or SPPs (Jain et al., 2020; DeYoung et al., 2019) with gold rationales. However, existing LM-based extractors have not been trained for faithfulness, while SPPs' faithfulness by construction comes at the great cost of task performance. Meanwhile, more existing works focus on improving the plausibility of free-text rationales (Narang et al., 2020; Lakhotia et al., 2020; Camburu et al., 2018) , often with task-specific pipelines Kumar and Talukdar, 2020) .",
                "cite_spans": [
                    {
                        "start": 131,
                        "end": 150,
                        "text": "(Bhat et al., 2021)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 159,
                        "end": 178,
                        "text": "(Jain et al., 2020;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 179,
                        "end": 200,
                        "text": "DeYoung et al., 2019)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 478,
                        "end": 499,
                        "text": "(Narang et al., 2020;",
                        "ref_id": null
                    },
                    {
                        "start": 500,
                        "end": 522,
                        "text": "Lakhotia et al., 2020;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 523,
                        "end": 544,
                        "text": "Camburu et al., 2018)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 582,
                        "end": 607,
                        "text": "Kumar and Talukdar, 2020)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "5"
            },
            {
                "text": "Connection to UNIREX Unlike prior works, UNIREX enables both the task model and rationale extractor to be jointly optimized for faithfulness, plausibility, and task performance. As a result, UNIREX-trained rationale extractors achieve a better balance of faithfulness and plausibility, without compromising the task model's performance. Also, by using a learned rationale extractor, which generally only requires one model forward pass, UNIREX does not have the computational expenses that limit many AAs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "5"
            },
            {
                "text": "Here, we formalize the text classification problem in more detail.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "Let D = {X , Y} N i=1 be a dataset, where X = {x i } N i=1",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "are the text inputs, Y = {y * i } N i=1 are the labels, and N is the number of instances (x i , y * i ) in D. We also assume D can be partitioned into train set D train , dev set D dev , and test set D test . Let F task = f task (f enc (\u2022)) be a task LM, where f enc is the text encoder, and f task is the task output head. Typically, F task has a BERT-style architecture (Devlin et al., 2018) , in which f enc is a Transformer (Vaswani et al., 2017) while f task is a linear layer. Below, we define the sequence classification (SST, Movies, MultiRC, e-SNLI) and multi-choice QA (CoS-E) tasks, which are different types of text classification.",
                "cite_spans": [
                    {
                        "start": 372,
                        "end": 393,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 428,
                        "end": 450,
                        "text": "(Vaswani et al., 2017)",
                        "ref_id": "BIBREF48"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "Sequence Classification In sequence classification, x i is a token sequence (e.g., a single sentence, a pair of sentences), while y * i is the target class for x i . Here, we assume a fixed label space",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "Y = {1, ..., M } of size M , where y * i \u2208 Y for all i. Thus, f task outputs a vector of size M , such that F task (x i ) = f task (f enc (x i )) =\u0177 i \u2208 R M is the logit vector used to classify x i . Given\u0177 i = [\u0177 i,j ] M",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "j=1 , let y i = arg max j\u0177i,j be the class predicted by F task . The goal of sequence classification is to learn F task such that y * i = y i , for all (x i , y * i ) (Minaee et al., 2021) .",
                "cite_spans": [
                    {
                        "start": 167,
                        "end": 188,
                        "text": "(Minaee et al., 2021)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "Multi-Choice QA Instead of a fixed label space, multi-choice QA has a different (but fixed-size) set of answer choices per instance. For instance i, let q i be the question (e.g., \"A friend is greeting me, what would they say?\") and",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "A i = {a i,j } M j=1",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "be the corresponding answer choices (e.g., {\"say hello\", \"greet\", \"associate\", \"socialize\", \"smile\"}), where M is now the number of answer choices. Define x i,j = q i \u2295 a i,j , where \u2295 denotes concatenation. In multi-choice QA, we have",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "x i = {x i,j } M j=1 , while y * i \u2208 A i",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "is the correct answer for x i . Thus, f task outputs a scalar, such that",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "F task (x i,j ) = f task (f enc (x i,j )) =\u0177 i,j \u2208 R is the logit for x i,j . Given\u0177 i = [\u0177 i,j ] M",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "j=1 , let j \u2032 = arg max j\u0177i,j , where y i = a i,j \u2032 is the answer predicted by F task . The goal of multi-choice QA is to learn F task such that y * i = y i , for all (x i , y * i ) (Talmor et al., 2018) .",
                "cite_spans": [
                    {
                        "start": 182,
                        "end": 203,
                        "text": "(Talmor et al., 2018)",
                        "ref_id": "BIBREF46"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.1 Text Classification",
                "sec_num": null
            },
            {
                "text": "A heuristic F task is an AA, which can be any handcrafted function that calculates an importance score s t i for each input token x t i (Bastings and Filippova, 2020) . AAs are typically gradient-based (Sundararajan et al., 2017; Denil et al., 2014; Lundberg and Lee, 2017; Li et al., 2015) or perturbationbased (Li et al., 2016; Poerner et al., 2018; K\u00e1d\u00e1r et al., 2017) methods. Gradient-based methods compute s t i via the gradient of F task 's output\u0177 i w.r.t. x t i , via one or more F task backward passes. Perturbation-based methods measure s t i as\u0177 i 's change when perturbing (e.g., removing) x t i , via multiple F task forward passes.",
                "cite_spans": [
                    {
                        "start": 136,
                        "end": 166,
                        "text": "(Bastings and Filippova, 2020)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 202,
                        "end": 229,
                        "text": "(Sundararajan et al., 2017;",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 230,
                        "end": 249,
                        "text": "Denil et al., 2014;",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 250,
                        "end": 273,
                        "text": "Lundberg and Lee, 2017;",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 274,
                        "end": 290,
                        "text": "Li et al., 2015)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 312,
                        "end": 329,
                        "text": "(Li et al., 2016;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 330,
                        "end": 351,
                        "text": "Poerner et al., 2018;",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 352,
                        "end": 371,
                        "text": "K\u00e1d\u00e1r et al., 2017)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.2 Heuristic Rationale Extractors",
                "sec_num": null
            },
            {
                "text": "AAs can be used out of the box without training and are designed to satisfy certain faithfulnessrelated axiomatic properties (Sundararajan et al., 2017; Lundberg and Lee, 2017) . However, AAs' lack of learnable parameters means they cannot be optimized for faithfulness/plausibility. Thus, if F task is trained for explainability using AA-based rationales, then only F task is optimized. Also, faithful AAs tend to be compute-intensive, requiring many F task backward/forward passes per instance (Sundararajan et al., 2017; Lundberg and Lee, 2017; Li et al., 2016) .",
                "cite_spans": [
                    {
                        "start": 125,
                        "end": 152,
                        "text": "(Sundararajan et al., 2017;",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 153,
                        "end": 176,
                        "text": "Lundberg and Lee, 2017)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 496,
                        "end": 523,
                        "text": "(Sundararajan et al., 2017;",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 524,
                        "end": 547,
                        "text": "Lundberg and Lee, 2017;",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 548,
                        "end": 564,
                        "text": "Li et al., 2016)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.2 Heuristic Rationale Extractors",
                "sec_num": null
            },
            {
                "text": "If a learned rationale extractor is chosen, UNIREX enables users to specify how much gold rationale supervision to use. Ideally, each train instance would be annotated with a gold rationale. In this case, we could directly minimize the plausibility loss for each train instance. However, since gold rationales can be expensive to annotate, UNIREX provides a special batching procedure for training with limited gold rationale supervision.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.3 Gold Rationale Supervision",
                "sec_num": null
            },
            {
                "text": "Given N train = |D train | train instances, let 0 < \u03b3 < 100 be the percentage of train instances with gold rationales, N gold = \u2308 \u03b3 100 N train \u2309 \u2265 1 be the number of train instances with gold rationales, b be the desired train batch size, and \u03b2 > 1 be a scaling factor. Define D gold \u2286 D train as the set of train instances with gold rationales, where |D gold | = N gold . Note that, if all train instances have gold rationales, then D gold = D train and \u03b3 = 100.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.3 Gold Rationale Supervision",
                "sec_num": null
            },
            {
                "text": "Each batch is constructed as follows: (1) randomly sample b gold = max(1, b \u03b2 ) instances from D gold without replacement, then (2) randomly sample b \u2212 b gold instances from D train \\D gold without replacement. This results in a batch with b total train instances, b gold with gold rationales and the rest without. Since N gold is generally small, we only sample from D gold without replacement for a given batch, but not a given epoch. Thus, instances from D gold may appear more than once in the same epoch. However, we do sample from D train \\D gold without replacement for each batch and epoch, so every instance in D train \\D gold appears exactly once per epoch.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.3 Gold Rationale Supervision",
                "sec_num": null
            },
            {
                "text": "After constructing the batch, we compute the plausibility loss for the batch as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.3 Gold Rationale Supervision",
                "sec_num": null
            },
            {
                "text": "b i=1 1 (x i ,y * i )\u2208D gold L plaus (F ext (x i ), r * i )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.3 Gold Rationale Supervision",
                "sec_num": null
            },
            {
                "text": ", where L plaus is the plausibility loss for train instance (x i , y * i ). This function zeroes out the plausibility loss for instances without gold rationales, so that plausibility is only being optimized with respect to instances with gold rationales. However, in Sec. ??, we show that it is possible to achieve high plausibility via rationale extractors trained on minimal gold rationale supervision.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.3 Gold Rationale Supervision",
                "sec_num": null
            },
            {
                "text": "Sufficiency In addition, to the criteria presented in Sec. 3.2, we consider two other sufficiency loss functions. The first is the KL divergence criterion used in (Ismail et al., 2021) , which considers the entire label distribution and is defined as",
                "cite_spans": [
                    {
                        "start": 163,
                        "end": 184,
                        "text": "(Ismail et al., 2021)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.1 Faithfulness",
                "sec_num": null
            },
            {
                "text": "L suff-KL = KL(F task (r (k) i )) || F task (x i )).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.1 Faithfulness",
                "sec_num": null
            },
            {
                "text": "The second is the mean absolute error (MAE) criterion, which is defined as",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.1 Faithfulness",
                "sec_num": null
            },
            {
                "text": "L suff-MAE = |L CE (F task (r (k) i )), y * i ) \u2212 L CE (F task (x i ), y * i )|.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.1 Faithfulness",
                "sec_num": null
            },
            {
                "text": "Unlike the difference criterion L suff-diff and margin criterion L suff-margin (Sec. 3.2), the MAE criterion assumes that using r (k) i as input should not yield better task performance than using x i as input. In our experiments, we find that L suff-margin is effective, though others (e.g., KL divergence, MAE) can be used too.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.1 Faithfulness",
                "sec_num": null
            },
            {
                "text": "Similar to faithfulness, UNIREX places no restrictions on the choice of plausibility objective. As described in Sec. 3.2, given gold rationale r * i for input x i , plausibility optimization entails training F ext to predict binary importance label r * ,t i for each token x t i . This is essentially binary token classification, so one natural choice for L plaus is the token-level binary cross-entropy (BCE) criterion: L plaus-BCE = \u2212 t r * ,t i log(F ext (x t i )) (Sec. 3.2). Another option is the sequence-level KL divergence criterion, which is defined as:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "L plaus-KL = KL(F ext (x i ) || r * i )",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": ". Additionally, we can directly penalize F ext (x i ) in the logit space via a linear loss, defined as:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "L plaus-linear = \u03a6(r * i ) F ext (x i ),",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "where \u03a6(u) = \u22122u + 1 maps positive and negative tokens to \u22121 and +1, respectively. The linear loss directly pushes the logits corresponding to positive/negative tokens to be higher/lower and increase the margin between them. To prevent linear loss values from becoming arbitrarily negative, we can also lower bound the loss with a margin m p , yielding:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "L plaus-linear-margin = max(\u2212m p , L plaus-linear ) + m p .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "A.5 Implementation Details LM Architecture While many prior works use BERT (Devlin et al., 2018) Transformer LMs, BERT is limited to having sequences with up to 512 tokens, which is problematic since many datasets (e.g., Movies) contain much longer sequences. Meanwhile, BigBird (Zaheer et al., 2020) is a state-of-the-art Transformer LM designed to handle long input sequences with up to 4096 tokens. Thus, we use BigBird-Base, which is initialized with RoBERTa-Base (Liu et al., 2019) , in all of our experiments (i.e., both baselines and UNIREX). We obtain the pre-trained BigBird-Base model from the Hugging Face Transformers library (Wolf et al., 2019) . Note that UNIREX is agnostic to the choice of LM architecture, so RNNs, CNNs, and other Transformer LMs are also supported by UNIREX. However, we leave exploration of other LM architectures for future work.",
                "cite_spans": [
                    {
                        "start": 75,
                        "end": 96,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 279,
                        "end": 300,
                        "text": "(Zaheer et al., 2020)",
                        "ref_id": "BIBREF53"
                    },
                    {
                        "start": 468,
                        "end": 486,
                        "text": "(Liu et al., 2019)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 638,
                        "end": 657,
                        "text": "(Wolf et al., 2019)",
                        "ref_id": "BIBREF49"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "Training Building upon Sec. ??, we discuss additional training details here. We find that \u03b1 c = 0.5 and \u03b1 s = 0.5 are usually best. For the batching factor \u03b2 (Sec. A.3), we use 2. For model selection, we choose the model with the best dev performance averaged over three seeds. We can also perform model selection based on dev explainability metrics, but we leave this extended tuning for future work. All experiments are implemented using PyTorch-Lightning (Paszke et al., 2019 ; Falcon and The PyTorch Lightning team, 2019). Fig. ?? shows the gold rationale data efficiency results for CoS-E, using the same setup as Sec. ??. Overall, we see that the CoS-E results are quite similar to the SST results. Again, UNIREX (DLM-FP) and UNIREX (SLM-FP) dominate across all \u03b3 values, with AUPRC slowly decreasing as \u03b3 de-creases. Interestingly, UNIREX (AA-FP) yields a noticeable dip in AUPRC for lower \u03b3 values. Since AA-FP has limited capacity (via the task model) for plausibility optimization, it is possible that this fluctuation is due to random noise. We leave further analysis of this for future work. ",
                "cite_spans": [
                    {
                        "start": 458,
                        "end": 478,
                        "text": "(Paszke et al., 2019",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 527,
                        "end": 534,
                        "text": "Fig. ??",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "A.4.2 Plausibility",
                "sec_num": null
            },
            {
                "text": "In this subsection, we present additional results from our experiments. Besides the aggregated results shown in Sec. 4 of the main text, Tables 4-10 contain more detailed results, using both raw and NRG metrics. Specifically , Tables 4-8 show all raw/NRG results for each dataset, Table 9 shows the ablation results for all raw metrics, and Table 10 includes the zero-shot explainability transfer results for UNIREX (SLM-FP). Generally, the computation of NRG should involve globally aggregating the raw metrics for all available methods, as done in the main results. However, for a number of more focused experiments (Tables 9-10), only a subset of the available methods are considered. Thus, to make the faithfulness results in Tables 9-10 easier to digest, we introduce a metric called Comp-Suff Difference (CSD), which locally aggregates comp and suff as: CSD = comp \u2212 suff. Therefore, since higher/lower comp/suff signal higher faithfulness, then higher CSD signals higher faithfulness. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 225,
                        "end": 237,
                        "text": ", Tables 4-8",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 281,
                        "end": 288,
                        "text": "Table 9",
                        "ref_id": null
                    },
                    {
                        "start": 341,
                        "end": 349,
                        "text": "Table 10",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "A.7 Additional Empirical Results",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Interpretable neural predictions with differentiable binary variables",
                "authors": [
                    {
                        "first": "Jasmijn",
                        "middle": [],
                        "last": "Bastings",
                        "suffix": ""
                    },
                    {
                        "first": "Wilker",
                        "middle": [],
                        "last": "Aziz",
                        "suffix": ""
                    },
                    {
                        "first": "Ivan",
                        "middle": [],
                        "last": "Titov",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1905.08160"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jasmijn Bastings, Wilker Aziz, and Ivan Titov. 2019. Interpretable neural predictions with differentiable binary variables. arXiv preprint arXiv:1905.08160.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "The elephant in the interpretability room: Why use attention as explanation when we have saliency methods? arXiv preprint",
                "authors": [
                    {
                        "first": "Jasmijn",
                        "middle": [],
                        "last": "Bastings",
                        "suffix": ""
                    },
                    {
                        "first": "Katja",
                        "middle": [],
                        "last": "Filippova",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2010.05607"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jasmijn Bastings and Katja Filippova. 2020. The ele- phant in the interpretability room: Why use atten- tion as explanation when we have saliency methods? arXiv preprint arXiv:2010.05607.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "On the dangers of stochastic parrots: Can language models be too big",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Emily",
                        "suffix": ""
                    },
                    {
                        "first": "Timnit",
                        "middle": [],
                        "last": "Bender",
                        "suffix": ""
                    },
                    {
                        "first": "Angelina",
                        "middle": [],
                        "last": "Gebru",
                        "suffix": ""
                    },
                    {
                        "first": "Shmargaret",
                        "middle": [],
                        "last": "Mcmillan-Major",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Shmitchell",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency",
                "volume": "",
                "issue": "",
                "pages": "610--623",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emily M Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big?. In Proceedings of the 2021 ACM Confer- ence on Fairness, Accountability, and Transparency, pages 610-623.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Self-training with few-shot rationalization: Teacher explanations aid student in few-shot nlu",
                "authors": [
                    {
                        "first": "Alessandro",
                        "middle": [],
                        "last": "Meghana Moorthy Bhat",
                        "suffix": ""
                    },
                    {
                        "first": "Subhabrata",
                        "middle": [],
                        "last": "Sordoni",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Mukherjee",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2109.08259"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Meghana Moorthy Bhat, Alessandro Sordoni, and Sub- habrata Mukherjee. 2021. Self-training with few-shot rationalization: Teacher explanations aid student in few-shot nlu. arXiv preprint arXiv:2109.08259.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Natural language inference with natural language explanations",
                "authors": [
                    {
                        "first": "Oana-Maria",
                        "middle": [],
                        "last": "Camburu",
                        "suffix": ""
                    },
                    {
                        "first": "Tim",
                        "middle": [],
                        "last": "Rockt\u00e4schel",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Lukasiewicz",
                        "suffix": ""
                    },
                    {
                        "first": "Phil",
                        "middle": [],
                        "last": "Blunsom",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1812.01193"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Oana-Maria Camburu, Tim Rockt\u00e4schel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natu- ral language inference with natural language expla- nations. arXiv preprint arXiv:1812.01193.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Anirudh Rathore, and Chenhao Tan. 2020. Evaluating and characterizing human rationales",
                "authors": [
                    {
                        "first": "Samuel",
                        "middle": [],
                        "last": "Carton",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2010.04736"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Samuel Carton, Anirudh Rathore, and Chenhao Tan. 2020. Evaluating and characterizing human ratio- nales. arXiv preprint arXiv:2010.04736.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Salkg: Learning from knowledge graph explanations for commonsense reasoning",
                "authors": [
                    {
                        "first": "Aaron",
                        "middle": [],
                        "last": "Chan",
                        "suffix": ""
                    },
                    {
                        "first": "Jiashu",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Boyuan",
                        "middle": [],
                        "last": "Long",
                        "suffix": ""
                    },
                    {
                        "first": "Soumya",
                        "middle": [],
                        "last": "Sanyal",
                        "suffix": ""
                    },
                    {
                        "first": "Tanishq",
                        "middle": [],
                        "last": "Gupta",
                        "suffix": ""
                    },
                    {
                        "first": "Xiang",
                        "middle": [],
                        "last": "Ren",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Aaron Chan, Jiashu Xu, Boyuan Long, Soumya Sanyal, Tanishq Gupta, and Xiang Ren. 2021. Salkg: Learn- ing from knowledge graph explanations for common- sense reasoning. Advances in Neural Information Processing Systems, 34.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Aitor Garc\u00eda-Pablos, and Montse Cuadros",
                "authors": [
                    {
                        "first": "Ona",
                        "middle": [],
                        "last": "De Gibert",
                        "suffix": ""
                    },
                    {
                        "first": "Naiara",
                        "middle": [],
                        "last": "Perez",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Hate speech dataset from a white supremacy forum",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1809.04444"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ona de Gibert, Naiara Perez, Aitor Garc\u00eda-Pablos, and Montse Cuadros. 2018. Hate speech dataset from a white supremacy forum. arXiv preprint arXiv:1809.04444.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Extraction of salient sentences from labelled documents",
                "authors": [
                    {
                        "first": "Misha",
                        "middle": [],
                        "last": "Denil",
                        "suffix": ""
                    },
                    {
                        "first": "Alban",
                        "middle": [],
                        "last": "Demiraj",
                        "suffix": ""
                    },
                    {
                        "first": "Nando De",
                        "middle": [],
                        "last": "Freitas",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1412.6815"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Misha Denil, Alban Demiraj, and Nando De Freitas. 2014. Extraction of salient sentences from labelled documents. arXiv preprint arXiv:1412.6815.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "Ming-Wei",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1810.04805"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Eraser: A benchmark to evaluate rationalized nlp models",
                "authors": [
                    {
                        "first": "Jay",
                        "middle": [],
                        "last": "Deyoung",
                        "suffix": ""
                    },
                    {
                        "first": "Sarthak",
                        "middle": [],
                        "last": "Jain",
                        "suffix": ""
                    },
                    {
                        "first": "Nazneen",
                        "middle": [],
                        "last": "Fatema Rajani",
                        "suffix": ""
                    },
                    {
                        "first": "Eric",
                        "middle": [],
                        "last": "Lehman",
                        "suffix": ""
                    },
                    {
                        "first": "Caiming",
                        "middle": [],
                        "last": "Xiong",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "Byron C",
                        "middle": [],
                        "last": "Wallace",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1911.03429"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. 2019. Eraser: A benchmark to evaluate rationalized nlp models. arXiv preprint arXiv:1911.03429.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Towards a rigorous science of interpretable machine learning",
                "authors": [
                    {
                        "first": "Finale",
                        "middle": [],
                        "last": "Doshi",
                        "suffix": ""
                    },
                    {
                        "first": "-",
                        "middle": [],
                        "last": "Velez",
                        "suffix": ""
                    },
                    {
                        "first": "Been",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1702.08608"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Finale Doshi-Velez and Been Kim. 2017. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "William Falcon and The PyTorch Lightning team",
                "authors": [],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.5281/zenodo.3828935"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "William Falcon and The PyTorch Lightning team. 2019. PyTorch Lightning.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Measuring nominal scale agreement among many raters",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Joseph",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Fleiss",
                        "suffix": ""
                    }
                ],
                "year": 1971,
                "venue": "Psychological bulletin",
                "volume": "76",
                "issue": "5",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Joseph L Fleiss. 1971. Measuring nominal scale agree- ment among many raters. Psychological bulletin, 76(5):378.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Evaluating explainable ai: Which algorithmic explanations help users predict model behavior?",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Hase",
                        "suffix": ""
                    },
                    {
                        "first": "Mohit",
                        "middle": [],
                        "last": "Bansal",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2005.01831"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Peter Hase and Mohit Bansal. 2020. Evaluating ex- plainable ai: Which algorithmic explanations help users predict model behavior? arXiv preprint arXiv:2005.01831.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "A benchmark for interpretability methods in deep neural networks",
                "authors": [
                    {
                        "first": "Sara",
                        "middle": [],
                        "last": "Hooker",
                        "suffix": ""
                    },
                    {
                        "first": "Dumitru",
                        "middle": [],
                        "last": "Erhan",
                        "suffix": ""
                    },
                    {
                        "first": "Pieter-Jan",
                        "middle": [],
                        "last": "Kindermans",
                        "suffix": ""
                    },
                    {
                        "first": "Been",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1806.10758"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. 2018. A benchmark for interpretabil- ity methods in deep neural networks. arXiv preprint arXiv:1806.10758.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Improving deep learning interpretability by saliency guided training",
                "authors": [
                    {
                        "first": "Aya Abdelsalam",
                        "middle": [],
                        "last": "Ismail",
                        "suffix": ""
                    },
                    {
                        "first": "Hector",
                        "middle": [
                            "Corrada"
                        ],
                        "last": "Bravo",
                        "suffix": ""
                    },
                    {
                        "first": "Soheil",
                        "middle": [],
                        "last": "Feizi",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. 2021. Improving deep learning inter- pretability by saliency guided training. Advances in Neural Information Processing Systems, 34.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness?",
                "authors": [
                    {
                        "first": "Alon",
                        "middle": [],
                        "last": "Jacovi",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2004.03685"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Alon Jacovi and Yoav Goldberg. 2020. Towards faith- fully interpretable nlp systems: How should we define and evaluate faithfulness? arXiv preprint arXiv:2004.03685.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Learning to faithfully rationalize by construction",
                "authors": [
                    {
                        "first": "Sarthak",
                        "middle": [],
                        "last": "Jain",
                        "suffix": ""
                    },
                    {
                        "first": "Sarah",
                        "middle": [],
                        "last": "Wiegreffe",
                        "suffix": ""
                    },
                    {
                        "first": "Yuval",
                        "middle": [],
                        "last": "Pinter",
                        "suffix": ""
                    },
                    {
                        "first": "Byron C",
                        "middle": [],
                        "last": "Wallace",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2005.00115"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sarthak Jain, Sarah Wiegreffe, Yuval Pinter, and Byron C Wallace. 2020. Learning to faith- fully rationalize by construction. arXiv preprint arXiv:2005.00115.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Representation of linguistic form and function in recurrent neural networks",
                "authors": [
                    {
                        "first": "Akos",
                        "middle": [],
                        "last": "K\u00e1d\u00e1r",
                        "suffix": ""
                    },
                    {
                        "first": "Grzegorz",
                        "middle": [],
                        "last": "Chrupa\u0142a",
                        "suffix": ""
                    },
                    {
                        "first": "Afra",
                        "middle": [],
                        "last": "Alishahi",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Computational Linguistics",
                "volume": "43",
                "issue": "4",
                "pages": "761--780",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Akos K\u00e1d\u00e1r, Grzegorz Chrupa\u0142a, and Afra Alishahi. 2017. Representation of linguistic form and func- tion in recurrent neural networks. Computational Linguistics, 43(4):761-780.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Looking beyond the surface: A challenge set for reading comprehension over multiple sentences",
                "authors": [
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Khashabi",
                        "suffix": ""
                    },
                    {
                        "first": "Snigdha",
                        "middle": [],
                        "last": "Chaturvedi",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    },
                    {
                        "first": "Shyam",
                        "middle": [],
                        "last": "Upadhyay",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "252--262",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. 2018. Looking beyond the surface: A challenge set for reading com- prehension over multiple sentences. In Proceedings of the 2018 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Pa- pers), pages 252-262.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Nile: Natural language inference with faithful natural language explanations",
                "authors": [
                    {
                        "first": "Sawan",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    },
                    {
                        "first": "Partha",
                        "middle": [],
                        "last": "Talukdar",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2005.12116"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sawan Kumar and Partha Talukdar. 2020. Nile: Natu- ral language inference with faithful natural language explanations. arXiv preprint arXiv:2005.12116.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Fid-ex: Improving sequence-to-sequence models for extractive rationale generation",
                "authors": [
                    {
                        "first": "Kushal",
                        "middle": [],
                        "last": "Lakhotia",
                        "suffix": ""
                    },
                    {
                        "first": "Bhargavi",
                        "middle": [],
                        "last": "Paranjape",
                        "suffix": ""
                    },
                    {
                        "first": "Asish",
                        "middle": [],
                        "last": "Ghoshal",
                        "suffix": ""
                    },
                    {
                        "first": "Wen-Tau",
                        "middle": [],
                        "last": "Yih",
                        "suffix": ""
                    },
                    {
                        "first": "Yashar",
                        "middle": [],
                        "last": "Mehdad",
                        "suffix": ""
                    },
                    {
                        "first": "Srinivasan",
                        "middle": [],
                        "last": "Iyer",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2012.15482"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kushal Lakhotia, Bhargavi Paranjape, Asish Ghoshal, Wen-tau Yih, Yashar Mehdad, and Srinivasan Iyer. 2020. Fid-ex: Improving sequence-to-sequence mod- els for extractive rationale generation. arXiv preprint arXiv:2012.15482.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Visualizing and understanding neural models in nlp",
                "authors": [
                    {
                        "first": "Jiwei",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Xinlei",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Eduard",
                        "middle": [],
                        "last": "Hovy",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1506.01066"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jiwei Li, Xinlei Chen, Eduard Hovy, and Dan Jurafsky. 2015. Visualizing and understanding neural models in nlp. arXiv preprint arXiv:1506.01066.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Understanding neural networks through representation erasure",
                "authors": [
                    {
                        "first": "Jiwei",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Will",
                        "middle": [],
                        "last": "Monroe",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1612.08220"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jiwei Li, Will Monroe, and Dan Jurafsky. 2016. Un- derstanding neural networks through representation erasure. arXiv preprint arXiv:1612.08220.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Zachary C Lipton",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Queue",
                "volume": "16",
                "issue": "3",
                "pages": "31--57",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zachary C Lipton. 2018. The mythos of model inter- pretability: In machine learning, the concept of in- terpretability is both important and slippery. Queue, 16(3):31-57.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Roberta: A robustly optimized bert pretraining approach",
                "authors": [
                    {
                        "first": "Yinhan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Myle",
                        "middle": [],
                        "last": "Ott",
                        "suffix": ""
                    },
                    {
                        "first": "Naman",
                        "middle": [],
                        "last": "Goyal",
                        "suffix": ""
                    },
                    {
                        "first": "Jingfei",
                        "middle": [],
                        "last": "Du",
                        "suffix": ""
                    },
                    {
                        "first": "Mandar",
                        "middle": [],
                        "last": "Joshi",
                        "suffix": ""
                    },
                    {
                        "first": "Danqi",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    },
                    {
                        "first": "Veselin",
                        "middle": [],
                        "last": "Stoyanov",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1907.11692"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining ap- proach. arXiv preprint arXiv:1907.11692.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "A unified approach to interpreting model predictions",
                "authors": [
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Scott",
                        "suffix": ""
                    },
                    {
                        "first": "Su-In",
                        "middle": [],
                        "last": "Lundberg",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 31st international conference on neural information processing systems",
                "volume": "",
                "issue": "",
                "pages": "4768--4777",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Scott M Lundberg and Su-In Lee. 2017. A unified ap- proach to interpreting model predictions. In Proceed- ings of the 31st international conference on neural information processing systems, pages 4768-4777.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Local interpretations for explainable natural language processing: A survey",
                "authors": [
                    {
                        "first": "Siwen",
                        "middle": [],
                        "last": "Luo",
                        "suffix": ""
                    },
                    {
                        "first": "Hamish",
                        "middle": [],
                        "last": "Ivison",
                        "suffix": ""
                    },
                    {
                        "first": "Caren",
                        "middle": [],
                        "last": "Han",
                        "suffix": ""
                    },
                    {
                        "first": "Josiah",
                        "middle": [],
                        "last": "Poon",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2103.11072"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Siwen Luo, Hamish Ivison, Caren Han, and Josiah Poon. 2021. Local interpretations for explainable natu- ral language processing: A survey. arXiv preprint arXiv:2103.11072.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Hidden factors and hidden topics: understanding rating dimensions with review text",
                "authors": [
                    {
                        "first": "Julian",
                        "middle": [],
                        "last": "Mcauley",
                        "suffix": ""
                    },
                    {
                        "first": "Jure",
                        "middle": [],
                        "last": "Leskovec",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 7th ACM conference on Recommender systems",
                "volume": "",
                "issue": "",
                "pages": "165--172",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Julian McAuley and Jure Leskovec. 2013. Hidden fac- tors and hidden topics: understanding rating dimen- sions with review text. In Proceedings of the 7th ACM conference on Recommender systems, pages 165-172.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Deep learning-based text classification: A comprehensive review",
                "authors": [
                    {
                        "first": "Shervin",
                        "middle": [],
                        "last": "Minaee",
                        "suffix": ""
                    },
                    {
                        "first": "Nal",
                        "middle": [],
                        "last": "Kalchbrenner",
                        "suffix": ""
                    },
                    {
                        "first": "Erik",
                        "middle": [],
                        "last": "Cambria",
                        "suffix": ""
                    },
                    {
                        "first": "Narjes",
                        "middle": [],
                        "last": "Nikzad",
                        "suffix": ""
                    },
                    {
                        "first": "Meysam",
                        "middle": [],
                        "last": "Chenaghlu",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "ACM Computing Surveys (CSUR)",
                "volume": "54",
                "issue": "3",
                "pages": "1--40",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Nar- jes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. 2021. Deep learning-based text classification: A comprehensive review. ACM Computing Surveys (CSUR), 54(3):1-40.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Noah Fiedel, and Karishma Malkan. 2020. Wt5?! training text-to-text models to explain their predictions",
                "authors": [
                    {
                        "first": "Sharan",
                        "middle": [],
                        "last": "Narang",
                        "suffix": ""
                    },
                    {
                        "first": "Colin",
                        "middle": [],
                        "last": "Raffel",
                        "suffix": ""
                    },
                    {
                        "first": "Katherine",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Roberts",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2004.14546"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. 2020. Wt5?! training text-to-text models to explain their predictions. arXiv preprint arXiv:2004.14546.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "An information bottleneck approach for controlling conciseness in rationale extraction",
                "authors": [
                    {
                        "first": "Mandar",
                        "middle": [],
                        "last": "Bhargavi Paranjape",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Joshi",
                        "suffix": ""
                    },
                    {
                        "first": "Hannaneh",
                        "middle": [],
                        "last": "Thickstun",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Hajishirzi",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2005.00652"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Bhargavi Paranjape, Mandar Joshi, John Thickstun, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2020. An information bottleneck approach for controlling conciseness in rationale extraction. arXiv preprint arXiv:2005.00652.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Pytorch: An imperative style, high-performance deep learning library",
                "authors": [
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Paszke",
                        "suffix": ""
                    },
                    {
                        "first": "Sam",
                        "middle": [],
                        "last": "Gross",
                        "suffix": ""
                    },
                    {
                        "first": "Francisco",
                        "middle": [],
                        "last": "Massa",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Lerer",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Bradbury",
                        "suffix": ""
                    },
                    {
                        "first": "Gregory",
                        "middle": [],
                        "last": "Chanan",
                        "suffix": ""
                    },
                    {
                        "first": "Trevor",
                        "middle": [],
                        "last": "Killeen",
                        "suffix": ""
                    },
                    {
                        "first": "Zeming",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "Natalia",
                        "middle": [],
                        "last": "Gimelshein",
                        "suffix": ""
                    },
                    {
                        "first": "Luca",
                        "middle": [],
                        "last": "Antiga",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Advances in neural information processing systems",
                "volume": "32",
                "issue": "",
                "pages": "8026--8037",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026- 8037.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Evaluating neural network explanation methods using hybrid documents and morphological agreement",
                "authors": [
                    {
                        "first": "Nina",
                        "middle": [],
                        "last": "Poerner",
                        "suffix": ""
                    },
                    {
                        "first": "Benjamin",
                        "middle": [],
                        "last": "Roth",
                        "suffix": ""
                    },
                    {
                        "first": "Hinrich",
                        "middle": [],
                        "last": "Sch\u00fctze",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1801.06422"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nina Poerner, Benjamin Roth, and Hinrich Sch\u00fctze. 2018. Evaluating neural network explanation meth- ods using hybrid documents and morphological agreement. arXiv preprint arXiv:1801.06422.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Evaluating explanations: How much do explanations from the teacher aid students?",
                "authors": [
                    {
                        "first": "Danish",
                        "middle": [],
                        "last": "Pruthi",
                        "suffix": ""
                    },
                    {
                        "first": "Bhuwan",
                        "middle": [],
                        "last": "Dhingra",
                        "suffix": ""
                    },
                    {
                        "first": "Livio Baldini",
                        "middle": [],
                        "last": "Soares",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Collins",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [],
                        "last": "Zachary",
                        "suffix": ""
                    },
                    {
                        "first": "Graham",
                        "middle": [],
                        "last": "Lipton",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "W"
                        ],
                        "last": "Neubig",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Cohen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2012.00893"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Danish Pruthi, Bhuwan Dhingra, Livio Baldini Soares, Michael Collins, Zachary C Lipton, Graham Neubig, and William W Cohen. 2020. Evaluating explana- tions: How much do explanations from the teacher aid students? arXiv preprint arXiv:2012.00893.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Explain yourself! leveraging language models for commonsense reasoning",
                "authors": [
                    {
                        "first": "Bryan",
                        "middle": [],
                        "last": "Nazneen Fatema Rajani",
                        "suffix": ""
                    },
                    {
                        "first": "Caiming",
                        "middle": [],
                        "last": "Mccann",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Xiong",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1906.02361"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain your- self! leveraging language models for commonsense reasoning. arXiv preprint arXiv:1906.02361.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead",
                "authors": [
                    {
                        "first": "Cynthia",
                        "middle": [],
                        "last": "Rudin",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Nature Machine Intelligence",
                "volume": "1",
                "issue": "5",
                "pages": "206--215",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Cynthia Rudin. 2019. Stop explaining black box ma- chine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "Efficient explanations from empirical explainers",
                "authors": [
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Schwarzenberg",
                        "suffix": ""
                    },
                    {
                        "first": "Nils",
                        "middle": [],
                        "last": "Feldhus",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "M\u00f6ller",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2103.15429"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Robert Schwarzenberg, Nils Feldhus, and Sebastian M\u00f6ller. 2021. Efficient explanations from empirical explainers. arXiv preprint arXiv:2103.15429.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "Learning important features through propagating activation differences",
                "authors": [
                    {
                        "first": "Avanti",
                        "middle": [],
                        "last": "Shrikumar",
                        "suffix": ""
                    },
                    {
                        "first": "Peyton",
                        "middle": [],
                        "last": "Greenside",
                        "suffix": ""
                    },
                    {
                        "first": "Anshul",
                        "middle": [],
                        "last": "Kundaje",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "International Conference on Machine Learning",
                "volume": "",
                "issue": "",
                "pages": "3145--3153",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Avanti Shrikumar, Peyton Greenside, and Anshul Kun- daje. 2017. Learning important features through propagating activation differences. In International Conference on Machine Learning, pages 3145-3153. PMLR.",
                "links": null
            },
            "BIBREF41": {
                "ref_id": "b41",
                "title": "Deep inside convolutional networks: Visualising image classification models and saliency maps",
                "authors": [
                    {
                        "first": "Karen",
                        "middle": [],
                        "last": "Simonyan",
                        "suffix": ""
                    },
                    {
                        "first": "Andrea",
                        "middle": [],
                        "last": "Vedaldi",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Zisserman",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1312.6034"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Karen Simonyan, Andrea Vedaldi, and Andrew Zis- serman. 2013. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034.",
                "links": null
            },
            "BIBREF42": {
                "ref_id": "b42",
                "title": "Learning to explain: Generating stable explanations fast",
                "authors": [
                    {
                        "first": "Xuelin",
                        "middle": [],
                        "last": "Situ",
                        "suffix": ""
                    },
                    {
                        "first": "Ingrid",
                        "middle": [],
                        "last": "Zukerman",
                        "suffix": ""
                    },
                    {
                        "first": "Cecile",
                        "middle": [],
                        "last": "Paris",
                        "suffix": ""
                    },
                    {
                        "first": "Sameen",
                        "middle": [],
                        "last": "Maruf",
                        "suffix": ""
                    },
                    {
                        "first": "Gholamreza",
                        "middle": [],
                        "last": "Haffari",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing",
                "volume": "1",
                "issue": "",
                "pages": "5340--5355",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xuelin Situ, Ingrid Zukerman, Cecile Paris, Sameen Maruf, and Gholamreza Haffari. 2021. Learning to explain: Generating stable explanations fast. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5340- 5355.",
                "links": null
            },
            "BIBREF43": {
                "ref_id": "b43",
                "title": "Recursive deep models for semantic compositionality over a sentiment treebank",
                "authors": [
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "Alex",
                        "middle": [],
                        "last": "Perelygin",
                        "suffix": ""
                    },
                    {
                        "first": "Jean",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Chuang",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Christopher",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    },
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Andrew",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Potts",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 2013 conference on empirical methods in natural language processing",
                "volume": "",
                "issue": "",
                "pages": "1631--1642",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empiri- cal methods in natural language processing, pages 1631-1642.",
                "links": null
            },
            "BIBREF44": {
                "ref_id": "b44",
                "title": "Do human rationales improve machine explanations? arXiv preprint",
                "authors": [
                    {
                        "first": "Julia",
                        "middle": [],
                        "last": "Strout",
                        "suffix": ""
                    },
                    {
                        "first": "Ye",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Raymond J",
                        "middle": [],
                        "last": "Mooney",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1905.13714"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Julia Strout, Ye Zhang, and Raymond J Mooney. 2019. Do human rationales improve machine explanations? arXiv preprint arXiv:1905.13714.",
                "links": null
            },
            "BIBREF45": {
                "ref_id": "b45",
                "title": "Axiomatic attribution for deep networks",
                "authors": [
                    {
                        "first": "Mukund",
                        "middle": [],
                        "last": "Sundararajan",
                        "suffix": ""
                    },
                    {
                        "first": "Ankur",
                        "middle": [],
                        "last": "Taly",
                        "suffix": ""
                    },
                    {
                        "first": "Qiqi",
                        "middle": [],
                        "last": "Yan",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "International Conference on Machine Learning",
                "volume": "",
                "issue": "",
                "pages": "3319--3328",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In Inter- national Conference on Machine Learning, pages 3319-3328. PMLR.",
                "links": null
            },
            "BIBREF46": {
                "ref_id": "b46",
                "title": "Commonsenseqa: A question answering challenge targeting commonsense knowledge",
                "authors": [
                    {
                        "first": "Alon",
                        "middle": [],
                        "last": "Talmor",
                        "suffix": ""
                    },
                    {
                        "first": "Jonathan",
                        "middle": [],
                        "last": "Herzig",
                        "suffix": ""
                    },
                    {
                        "first": "Nicholas",
                        "middle": [],
                        "last": "Lourie",
                        "suffix": ""
                    },
                    {
                        "first": "Jonathan",
                        "middle": [],
                        "last": "Berant",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1811.00937"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Commonsenseqa: A question answering challenge targeting commonsense knowl- edge. arXiv preprint arXiv:1811.00937.",
                "links": null
            },
            "BIBREF47": {
                "ref_id": "b47",
                "title": "Semeval-2018 task 3: Irony detection in english tweets",
                "authors": [
                    {
                        "first": "Cynthia",
                        "middle": [],
                        "last": "Van Hee",
                        "suffix": ""
                    },
                    {
                        "first": "Els",
                        "middle": [],
                        "last": "Lefever",
                        "suffix": ""
                    },
                    {
                        "first": "V\u00e9ronique",
                        "middle": [],
                        "last": "Hoste",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of The 12th International Workshop on Semantic Evaluation",
                "volume": "",
                "issue": "",
                "pages": "39--50",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Cynthia Van Hee, Els Lefever, and V\u00e9ronique Hoste. 2018. Semeval-2018 task 3: Irony detection in en- glish tweets. In Proceedings of The 12th Interna- tional Workshop on Semantic Evaluation, pages 39- 50.",
                "links": null
            },
            "BIBREF48": {
                "ref_id": "b48",
                "title": "Attention is all you need",
                "authors": [
                    {
                        "first": "Ashish",
                        "middle": [],
                        "last": "Vaswani",
                        "suffix": ""
                    },
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Shazeer",
                        "suffix": ""
                    },
                    {
                        "first": "Niki",
                        "middle": [],
                        "last": "Parmar",
                        "suffix": ""
                    },
                    {
                        "first": "Jakob",
                        "middle": [],
                        "last": "Uszkoreit",
                        "suffix": ""
                    },
                    {
                        "first": "Llion",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    },
                    {
                        "first": "Aidan",
                        "middle": [
                            "N"
                        ],
                        "last": "Gomez",
                        "suffix": ""
                    },
                    {
                        "first": "\u0141ukasz",
                        "middle": [],
                        "last": "Kaiser",
                        "suffix": ""
                    },
                    {
                        "first": "Illia",
                        "middle": [],
                        "last": "Polosukhin",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Advances in neural information processing systems",
                "volume": "",
                "issue": "",
                "pages": "5998--6008",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information pro- cessing systems, pages 5998-6008.",
                "links": null
            },
            "BIBREF49": {
                "ref_id": "b49",
                "title": "Huggingface's transformers: State-ofthe-art natural language processing",
                "authors": [
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Wolf",
                        "suffix": ""
                    },
                    {
                        "first": "Lysandre",
                        "middle": [],
                        "last": "Debut",
                        "suffix": ""
                    },
                    {
                        "first": "Victor",
                        "middle": [],
                        "last": "Sanh",
                        "suffix": ""
                    },
                    {
                        "first": "Julien",
                        "middle": [],
                        "last": "Chaumond",
                        "suffix": ""
                    },
                    {
                        "first": "Clement",
                        "middle": [],
                        "last": "Delangue",
                        "suffix": ""
                    },
                    {
                        "first": "Anthony",
                        "middle": [],
                        "last": "Moi",
                        "suffix": ""
                    },
                    {
                        "first": "Pierric",
                        "middle": [],
                        "last": "Cistac",
                        "suffix": ""
                    },
                    {
                        "first": "Tim",
                        "middle": [],
                        "last": "Rault",
                        "suffix": ""
                    },
                    {
                        "first": "R\u00e9mi",
                        "middle": [],
                        "last": "Louf",
                        "suffix": ""
                    },
                    {
                        "first": "Morgan",
                        "middle": [],
                        "last": "Funtowicz",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1910.03771"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, R\u00e9mi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of- the-art natural language processing. arXiv preprint arXiv:1910.03771.",
                "links": null
            },
            "BIBREF50": {
                "ref_id": "b50",
                "title": "Crossfit: A few-shot learning challenge for cross-task generalization in nlp",
                "authors": [
                    {
                        "first": "Qinyuan",
                        "middle": [],
                        "last": "Ye",
                        "suffix": ""
                    },
                    {
                        "first": "Xiang",
                        "middle": [],
                        "last": "Bill Yuchen Lin",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ren",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2104.08835"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Qinyuan Ye, Bill Yuchen Lin, and Xiang Ren. 2021. Crossfit: A few-shot learning challenge for cross-task generalization in nlp. arXiv preprint arXiv:2104.08835.",
                "links": null
            },
            "BIBREF51": {
                "ref_id": "b51",
                "title": "Rethinking cooperative rationalization: Introspective extraction and complement control",
                "authors": [
                    {
                        "first": "Mo",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    },
                    {
                        "first": "Shiyu",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Tommi",
                        "middle": [
                            "S"
                        ],
                        "last": "Jaakkola",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1910.13294"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Mo Yu, Shiyu Chang, Yang Zhang, and Tommi S Jaakkola. 2019. Rethinking cooperative rationaliza- tion: Introspective extraction and complement con- trol. arXiv preprint arXiv:1910.13294.",
                "links": null
            },
            "BIBREF52": {
                "ref_id": "b52",
                "title": "Understanding interlocking dynamics of cooperative rationalization",
                "authors": [
                    {
                        "first": "Mo",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Shiyu",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Tommi",
                        "middle": [],
                        "last": "Jaakkola",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mo Yu, Yang Zhang, Shiyu Chang, and Tommi Jaakkola. 2021. Understanding interlocking dynamics of coop- erative rationalization. Advances in Neural Informa- tion Processing Systems, 34.",
                "links": null
            },
            "BIBREF53": {
                "ref_id": "b53",
                "title": "Big bird: Transformers for longer sequences",
                "authors": [
                    {
                        "first": "Manzil",
                        "middle": [],
                        "last": "Zaheer",
                        "suffix": ""
                    },
                    {
                        "first": "Guru",
                        "middle": [],
                        "last": "Guruganesh",
                        "suffix": ""
                    },
                    {
                        "first": "Joshua",
                        "middle": [],
                        "last": "Kumar Avinava Dubey",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Ainslie",
                        "suffix": ""
                    },
                    {
                        "first": "Santiago",
                        "middle": [],
                        "last": "Alberti",
                        "suffix": ""
                    },
                    {
                        "first": "Philip",
                        "middle": [],
                        "last": "Ontanon",
                        "suffix": ""
                    },
                    {
                        "first": "Anirudh",
                        "middle": [],
                        "last": "Pham",
                        "suffix": ""
                    },
                    {
                        "first": "Qifan",
                        "middle": [],
                        "last": "Ravula",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "NeurIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago On- tanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. In NeurIPS.",
                "links": null
            },
            "BIBREF54": {
                "ref_id": "b54",
                "title": "Modeling annotators: A generative approach to learning from annotator rationales",
                "authors": [
                    {
                        "first": "Omar",
                        "middle": [],
                        "last": "Zaidan",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Eisner",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of the 2008 conference on Empirical methods in natural language processing",
                "volume": "",
                "issue": "",
                "pages": "31--40",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Omar Zaidan and Jason Eisner. 2008. Modeling an- notators: A generative approach to learning from annotator rationales. In Proceedings of the 2008 con- ference on Empirical methods in natural language processing, pages 31-40.",
                "links": null
            },
            "BIBREF55": {
                "ref_id": "b55",
                "title": "Semeval-2019 task 6: Identifying and categorizing offensive language in social media (offenseval)",
                "authors": [
                    {
                        "first": "Marcos",
                        "middle": [],
                        "last": "Zampieri",
                        "suffix": ""
                    },
                    {
                        "first": "Shervin",
                        "middle": [],
                        "last": "Malmasi",
                        "suffix": ""
                    },
                    {
                        "first": "Preslav",
                        "middle": [],
                        "last": "Nakov",
                        "suffix": ""
                    },
                    {
                        "first": "Sara",
                        "middle": [],
                        "last": "Rosenthal",
                        "suffix": ""
                    },
                    {
                        "first": "Noura",
                        "middle": [],
                        "last": "Farra",
                        "suffix": ""
                    },
                    {
                        "first": "Ritesh",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1903.08983"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Semeval-2019 task 6: Identifying and catego- rizing offensive language in social media (offenseval). arXiv preprint arXiv:1903.08983.",
                "links": null
            },
            "BIBREF56": {
                "ref_id": "b56",
                "title": "Character-level Convolutional Networks for Text Classification",
                "authors": [
                    {
                        "first": "Xiang",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Junbo",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "Yann",
                        "middle": [],
                        "last": "Lecun",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1509.01626"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level Convolutional Networks for Text Classification. arXiv:1509.01626 [cs].",
                "links": null
            },
            "BIBREF61": {
                "ref_id": "b61",
                "title": "Comp/Suff Loss UNIREX (SLM-FP, Comp) 93.59",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Comp/Suff Loss UNIREX (SLM-FP, Comp) 93.59 (\u00b10.11) 0.040 (\u00b10.096) 0.350 (\u00b10.048) 0.310 (\u00b10.049) 82.79 (\u00b10.62) 70.74 (\u00b10.81)",
                "links": null
            },
            "BIBREF63": {
                "ref_id": "b63",
                "title": "Suff Criterion UNIREX",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Slm-Fp",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Div",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Suff Criterion UNIREX (SLM-FP, KL Div) 93.06 (\u00b10.25) 0.174 (\u00b10.100) 0.306 (\u00b10.098) 0.131 (\u00b10.005) 82.62 (\u00b10.88) 70.43 (\u00b10.65)",
                "links": null
            },
            "BIBREF66": {
                "ref_id": "b66",
                "title": "SLM Ext Head UNIREX (SLM-FP, Linear) 93",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "68",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "SLM Ext Head UNIREX (SLM-FP, Linear) 93.68 (\u00b10.67) 0.189 (\u00b10.030) 0.302 (\u00b10.039) 0.113 (\u00b10.013) 82.55 (\u00b10.84) 70.65 (\u00b10.44)",
                "links": null
            },
            "BIBREF69": {
                "ref_id": "b69",
                "title": "Hate Speech Detection Stormfront Vanilla 10",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "48",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hate Speech Detection Stormfront Vanilla 10.48 (\u00b11.66) -0.066 (\u00b10.072) 0.153 (\u00b10.002) 0.219 (\u00b10.071)",
                "links": null
            },
            "BIBREF70": {
                "ref_id": "b70",
                "title": "Offensive Speech Detection OffenseEval Vanilla",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Offensive Speech Detection OffenseEval Vanilla 33.51 (\u00b10.99) -0.125 (\u00b10.068) 0.104 (\u00b10.007) 0.229 (\u00b10.064)",
                "links": null
            },
            "BIBREF71": {
                "ref_id": "b71",
                "title": ") Irony Detection SemEval2018-Irony Vanilla 29",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "63",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "UNIREX (DLM-FP) 35.52 (\u00b11.26) 0.053 (\u00b10.012) 0.140 (\u00b10.049) 0.087 (\u00b10.045) UNIREX (SLM-FP) 38.17 (\u00b10.96) 0.039 (\u00b10.031) 0.087 (\u00b10.016) 0.048 (\u00b10.024) Irony Detection SemEval2018-Irony Vanilla 29.63 (\u00b14.72) -0.058 (\u00b10.075) 0.154 (\u00b10.001) 0.212 (\u00b10.074)",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "uris": null,
                "text": "Desiderata of Rationale Extraction. Unlike prior works, UNIREX enables optimizing for all three desiderata.",
                "num": null
            },
            "FIGREF1": {
                "type_str": "figure",
                "uris": null,
                "text": "UNIREX Framework. UNIREX enables jointly optimizing the task model (Ftask) and rationale extractor (Fext), w.r.t. faithfulness (Lfaith), plausibility (Lplaus), and task performance (Ltask).",
                "num": null
            },
            "FIGREF2": {
                "type_str": "figure",
                "uris": null,
                "text": "Rationale Extractor Types.",
                "num": null
            },
            "FIGREF3": {
                "type_str": "figure",
                "uris": null,
                "text": "NRG Comparison by Desiderata. We show FNRG, PNRG, and TNRG for all methods, averaged over all datasets.",
                "num": null
            },
            "FIGREF4": {
                "type_str": "figure",
                "uris": null,
                "text": "Gold Rationale Data Efficiency on SST.",
                "num": null
            },
            "FIGREF5": {
                "type_str": "figure",
                "uris": null,
                "text": "Gold Rationale Data Efficiency on CoS-E.",
                "num": null
            },
            "TABREF0": {
                "type_str": "table",
                "html": null,
                "text": "Comp+Suff) 0.302 (\u00b10.039) 0.113 (\u00b10.013) 82.55 (\u00b10.84) 93.68 (\u00b10.67) FP (MLP-2048-2) 0.323 (\u00b10.071) 0.144 (\u00b10.012) 83.82 (\u00b10.77) 93.67 (\u00b10.18) SLM-FP (MLP-4096-3) 0.295 (\u00b10.057) 0.154 (\u00b10.027) 84.53 (\u00b10.61) 93.19 (\u00b10.79)",
                "num": null,
                "content": "<table><tr><td>Ablation</td><td>UNIREX Config</td><td>Faithfulness Comp (\u2191) Suff (\u2193)</td><td>Plausibility AUPRC (\u2191)</td><td>Performance Acc (\u2191)</td></tr><tr><td>Ext Type (F)</td><td>AA-F (Rand) AA-F (Gold) AA-F (Inv)</td><td colspan=\"3\">0.171 (\u00b10.040) 0.327 (\u00b10.050) 44.92 (\u00b10.00) 94.05 (\u00b10.35) 0.232 (\u00b10.088) 0.249 (\u00b10.021) 100.00 (\u00b10.00) 93.81 (\u00b10.54) 0.242 (\u00b10.010) 0.357 (\u00b10.019) 20.49 (\u00b10.00) 93.47 (\u00b11.81)</td></tr><tr><td/><td>AA-F (IG)</td><td colspan=\"3\">0.292 (\u00b10.051) 0.171 (\u00b10.038) 48.13 (\u00b11.14) 92.97 (\u00b10.44)</td></tr><tr><td/><td>AA-FP (Sum)</td><td colspan=\"3\">0.296 (\u00b10.067) 0.185 (\u00b10.048) 47.60 (\u00b12.44) 93.25 (\u00b10.45)</td></tr><tr><td>Ext Type (FP)</td><td>AA-FP (MLP) DLM-FP SLM-FP</td><td colspan=\"3\">0.285 (\u00b10.051) 0.197 (\u00b10.100) 54.82 (\u00b11.97) 93.23 (\u00b10.92) 0.319 (\u00b10.090) 0.167 (\u00b10.036) 85.80 (\u00b10.74) 93.81 (\u00b10.18) 0.302 (\u00b10.039) 0.113 (\u00b10.013) 82.55 (\u00b10.84) 93.68 (\u00b10.67)</td></tr><tr><td>Comp/Suff Loss</td><td/><td/><td/><td/></tr><tr><td/><td>SLM-FP (KL Div)</td><td/><td/><td/></tr><tr><td>Suff Criterion</td><td/><td/><td/><td/></tr></table>"
            },
            "TABREF1": {
                "type_str": "table",
                "html": null,
                "text": "UNIREX Ablation Studies on SST.",
                "num": null,
                "content": "<table/>"
            },
            "TABREF2": {
                "type_str": "table",
                "html": null,
                "text": "\u00b10.009) 0.258 (\u00b10.031) 93.81 (\u00b10.55) UNIREX (AA-F) 0.292 (\u00b10.051) 0.171 (\u00b10.038) 92.97 (\u00b10.44) UNIREX (DLM-FP) 0.319 (\u00b10.090) 0.167 (\u00b10.036) 93.81 (\u00b10.54) Yelp AA (IG) 0.069 (\u00b10.004) 0.219 (\u00b10.028) 92.50 (\u00b12.07) UNIREX (AA-F) 0.138 (\u00b10.078) 0.126 (\u00b10.059) 83.93 (\u00b113.20) UNIREX (DLM-FP) 0.265 (\u00b10.094) 0.097 (\u00b10.033) 92.37 (\u00b10.46)",
                "num": null,
                "content": "<table><tr><td>Task</td><td>Dataset</td><td>Method</td><td>Faithfulness Comp (\u2191) Suff (\u2193)</td><td>Task Performance Perf (\u2191)</td></tr><tr><td>SA</td><td colspan=\"3\">SST 0.119 (Amazon AA (IG) AA (IG) 0.076 (\u00b10.010) 0.224 (\u00b10.037) UNIREX (AA-F) 0.130 (\u00b10.077) 0.073 (\u00b10.039) UNIREX (DLM-FP) 0.232 (\u00b10.072) 0.098 (\u00b10.033)</td><td>91.13 (\u00b10.28) 77.90 (\u00b113.12) 89.35 (\u00b12.22)</td></tr><tr><td>HSD</td><td>Stormfront</td><td colspan=\"2\">AA (IG) UNIREX (AA-F) UNIREX (DLM-FP) 0.167 (\u00b10.084) 0.115 (\u00b10.059) 0.135 (\u00b10.010) 0.245 (\u00b10.059) 0.219 (\u00b10.009) 0.092 (\u00b10.025)</td><td>10.48 (\u00b11.66) 10.36 (\u00b11.94) 10.37 (\u00b12.66)</td></tr><tr><td/><td/><td>AA (IG)</td><td>0.097 (\u00b10.009) 0.244 (\u00b10.052)</td><td>33.51 (\u00b10.99)</td></tr><tr><td colspan=\"2\">OSD OffenseEval</td><td>UNIREX (AA-F)</td><td>0.074 (\u00b10.040) 0.102 (\u00b10.024)</td><td>32.62 (\u00b14.85)</td></tr><tr><td/><td/><td colspan=\"2\">UNIREX (DLM-FP) 0.140 (\u00b10.049) 0.087 (\u00b10.045)</td><td>35.52 (\u00b11.26)</td></tr><tr><td/><td/><td>AA (IG)</td><td>0.128 (\u00b10.014) 0.248 (\u00b10.064)</td><td>29.63 (\u00b14.72)</td></tr><tr><td>ID</td><td>SemEval2018</td><td colspan=\"2\">UNIREX (AA-F) UNIREX (DLM-FP) 0.149 (\u00b10.052) 0.102 (\u00b10.053) 0.069 (\u00b10.041) 0.096 (\u00b10.011)</td><td>49.95 (\u00b18.31) 31.97 (\u00b12.80)</td></tr></table>"
            },
            "TABREF3": {
                "type_str": "table",
                "html": null,
                "text": "Zero-Shot Faithfulness Transfer from SST.",
                "num": null,
                "content": "<table/>"
            },
            "TABREF5": {
                "type_str": "table",
                "html": null,
                "text": "Plausibility User Study on SST.",
                "num": null,
                "content": "<table/>"
            },
            "TABREF7": {
                "type_str": "table",
                "html": null,
                "text": "Main Results on SST.",
                "num": null,
                "content": "<table><tr><td>Method</td><td>Composite NRG (\u2191)</td><td>NRG (\u2191)</td><td>Faithfulness Comp (\u2191)</td><td>Suff (\u2193)</td><td>NRG (\u2191)</td><td>Plausibility AUPRC (\u2191)</td><td>TF1 (\u2191)</td><td colspan=\"2\">Performance NRG (\u2191) F1 (\u2191)</td></tr><tr><td>AA (Grad)</td><td>0.481</td><td>0.457</td><td colspan=\"2\">0.184 (\u00b10.023) 0.107 (\u00b10.017)</td><td>0.028</td><td>13.31 (\u00b10.91)</td><td>5.02 (\u00b10.00)</td><td>0.957</td><td>95.33 (\u00b10.65)</td></tr><tr><td>AA (Input*Grad)</td><td>0.503</td><td>0.359</td><td colspan=\"2\">0.148 (\u00b10.031) 0.137 (\u00b10.019)</td><td>0.194</td><td>8.68 (\u00b10.37)</td><td>37.58 (\u00b10.55)</td><td>0.957</td><td>95.33 (\u00b10.65)</td></tr><tr><td>AA (DeepLIFT)</td><td>0.468</td><td>0.259</td><td colspan=\"2\">0.122 (\u00b10.029) 0.172 (\u00b10.022)</td><td>0.187</td><td>9.00 (\u00b10.16)</td><td>36.15 (\u00b11.45)</td><td>0.957</td><td>95.33 (\u00b10.65)</td></tr><tr><td>AA (IG)</td><td>0.439</td><td>0.173</td><td colspan=\"2\">0.134 (\u00b10.016) 0.219 (\u00b10.044)</td><td>0.188</td><td>8.88 (\u00b10.21)</td><td>36.39 (\u00b11.29)</td><td>0.957</td><td>95.33 (\u00b10.65)</td></tr><tr><td>L2E</td><td>0.550</td><td>0.445</td><td colspan=\"2\">0.000 (\u00b10.007) 0.026 (\u00b10.015)</td><td>0.248</td><td colspan=\"2\">16.68 (\u00b110.20) 38.92 (\u00b14.07)</td><td>0.957</td><td>95.33 (\u00b10.65)</td></tr><tr><td>SGT</td><td>0.553</td><td>0.474</td><td colspan=\"2\">0.124 (\u00b10.053) 0.071 (\u00b10.064)</td><td>0.184</td><td>10.05 (\u00b11.23)</td><td>34.64 (\u00b11.67)</td><td>1.000</td><td>96.33 (\u00b10.76)</td></tr><tr><td>FRESH</td><td>0.645</td><td>0.732</td><td colspan=\"2\">0.234 (\u00b10.034) 0.000 (\u00b10.000)</td><td>0.305</td><td>17.02 (\u00b16.22)</td><td>48.26 (\u00b15.87)</td><td>0.899</td><td>94.00 (\u00b11.44)</td></tr><tr><td>A2R</td><td>0.431</td><td>0.764</td><td colspan=\"2\">0.267 (\u00b10.050) 0.000 (\u00b10.000)</td><td>0.244</td><td colspan=\"2\">35.44 (\u00b121.69) 19.78 (\u00b125.56)</td><td>0.284</td><td>79.78 (\u00b17.14)</td></tr><tr><td>UNIREX (AA-F)</td><td>0.601</td><td>0.744</td><td colspan=\"2\">0.505 (\u00b10.134) 0.122 (\u00b10.100)</td><td>0.189</td><td>9.14 (\u00b12.51)</td><td>36.28 (\u00b11.84)</td><td>0.870</td><td>93.33 (\u00b11.61)</td></tr><tr><td>SGT+P</td><td>0.586</td><td>0.604</td><td colspan=\"2\">0.152 (\u00b10.013) 0.022 (\u00b10.004)</td><td>0.183</td><td>9.16 (\u00b11.59)</td><td>35.33 (\u00b10.41)</td><td>0.971</td><td>95.66 (\u00b11.16)</td></tr><tr><td>FRESH+P</td><td>0.491</td><td>0.691</td><td colspan=\"2\">0.193 (\u00b10.062) 0.000 (\u00b10.000)</td><td>0.710</td><td colspan=\"2\">65.78 (\u00b111.16) 68.70 (\u00b115.78)</td><td>0.070</td><td>74.84 (\u00b112.22)</td></tr><tr><td>A2R+P</td><td>0.585</td><td>0.764</td><td colspan=\"2\">0.267 (\u00b10.076) 0.000 (\u00b10.000)</td><td>0.991</td><td>93.53 (\u00b10.93)</td><td>88.77 (\u00b11.22)</td><td>0.000</td><td>73.22 (\u00b10.75)</td></tr><tr><td>UNIREX (DLM-P)</td><td>0.667</td><td>0.024</td><td colspan=\"2\">0.024 (\u00b10.003) 0.238 (\u00b10.004)</td><td>1.000</td><td>94.32 (\u00b10.12)</td><td>89.53 (\u00b11.63)</td><td>0.978</td><td>95.83 (\u00b10.29)</td></tr><tr><td>UNIREX (AA-FP)</td><td>0.543</td><td>0.514</td><td colspan=\"2\">0.428 (\u00b10.174) 0.195 (\u00b10.105)</td><td>0.193</td><td>8.53 (\u00b10.46)</td><td>37.71 (\u00b13.12)</td><td>0.921</td><td>94.50 (\u00b11.00)</td></tr><tr><td>UNIREX (DLM-FP)</td><td>0.744</td><td>0.326</td><td colspan=\"2\">0.283 (\u00b10.217) 0.216 (\u00b10.005)</td><td>0.991</td><td>93.65 (\u00b10.36)</td><td>88.68 (\u00b12.29)</td><td>0.913</td><td>94.33 (\u00b11.61)</td></tr><tr><td>UNIREX (SLM-FP)</td><td>0.754</td><td>0.362</td><td colspan=\"2\">0.313 (\u00b10.059) 0.213 (\u00b10.014)</td><td>0.965</td><td>91.70 (\u00b11.84)</td><td>86.17 (\u00b11.20)</td><td>0.935</td><td>94.83 (\u00b10.76)</td></tr></table>"
            },
            "TABREF8": {
                "type_str": "table",
                "html": null,
                "text": "Main Results on Movies.",
                "num": null,
                "content": "<table><tr><td>Method</td><td>Composite NRG (\u2191)</td><td>NRG (\u2191)</td><td>Faithfulness Comp (\u2191)</td><td>Suff (\u2193)</td><td>NRG (\u2191)</td><td>Plausibility AUPRC (\u2191)</td><td>TF1 (\u2191)</td><td colspan=\"2\">Performance NRG (\u2191) Acc (\u2191)</td></tr><tr><td>AA (Grad)</td><td>0.537</td><td>0.504</td><td colspan=\"2\">0.331 (\u00b10.012) 0.352 (\u00b10.007)</td><td>0.130</td><td colspan=\"2\">37.33 (\u00b10.62) 22.65 (\u00b10.00)</td><td>0.977</td><td>63.56 (\u00b11.27)</td></tr><tr><td>AA (Input*Grad)</td><td>0.573</td><td>0.361</td><td colspan=\"2\">0.249 (\u00b10.018) 0.385 (\u00b10.008)</td><td>0.383</td><td colspan=\"2\">39.56 (\u00b10.54) 44.43 (\u00b10.40)</td><td>0.977</td><td>63.56 (\u00b11.27)</td></tr><tr><td>AA (DeepLIFT)</td><td>0.605</td><td>0.346</td><td colspan=\"2\">0.254 (\u00b10.035) 0.403 (\u00b10.042)</td><td>0.491</td><td colspan=\"2\">42.82 (\u00b11.83) 51.72 (\u00b11.26)</td><td>0.977</td><td>63.56 (\u00b11.27)</td></tr><tr><td>AA (IG)</td><td>0.578</td><td>0.327</td><td colspan=\"2\">0.216 (\u00b10.007) 0.378 (\u00b10.010)</td><td>0.429</td><td colspan=\"2\">40.07 (\u00b15.47) 48.34 (\u00b13.16)</td><td>0.977</td><td>63.56 (\u00b11.27)</td></tr><tr><td>L2E</td><td>0.544</td><td>0.493</td><td colspan=\"2\">0.005 (\u00b10.003) 0.010 (\u00b10.008)</td><td>0.161</td><td colspan=\"2\">23.56 (\u00b11.09) 37.80 (\u00b11.10)</td><td>0.977</td><td>63.56 (\u00b11.27)</td></tr><tr><td>SGT</td><td>0.618</td><td>0.367</td><td colspan=\"2\">0.197 (\u00b10.040) 0.324 (\u00b10.015)</td><td>0.491</td><td colspan=\"2\">43.68 (\u00b14.68) 51.00 (\u00b13.05)</td><td>0.995</td><td>64.35 (\u00b10.46)</td></tr><tr><td>FRESH</td><td>0.302</td><td>0.546</td><td colspan=\"2\">0.037 (\u00b10.036) 0.000 (\u00b10.000)</td><td>0.261</td><td colspan=\"2\">32.35 (\u00b17.66) 39.37 (\u00b10.70)</td><td>0.101</td><td>24.81 (\u00b13.46)</td></tr><tr><td>A2R</td><td>0.277</td><td>0.516</td><td colspan=\"2\">0.014 (\u00b10.021) 0.000 (\u00b10.000)</td><td>0.282</td><td colspan=\"2\">41.61 (\u00b13.85) 33.12 (\u00b19.06)</td><td>0.032</td><td>21.77 (\u00b11.31)</td></tr><tr><td>UNIREX (AA-F)</td><td>0.690</td><td>0.538</td><td colspan=\"2\">0.297 (\u00b10.141) 0.286 (\u00b10.084)</td><td>0.554</td><td colspan=\"2\">46.97 (\u00b13.41) 53.99 (\u00b11.66)</td><td>0.978</td><td>63.58 (\u00b10.61)</td></tr><tr><td>SGT+P</td><td>0.601</td><td>0.367</td><td colspan=\"2\">0.201 (\u00b10.032) 0.328 (\u00b10.022)</td><td>0.436</td><td colspan=\"2\">41.30 (\u00b16.70) 47.95 (\u00b11.65)</td><td>1.000</td><td>64.57 (\u00b10.33)</td></tr><tr><td>FRESH+P</td><td>0.374</td><td>0.515</td><td colspan=\"2\">0.013 (\u00b10.021) 0.013 (\u00b10.021)</td><td>0.606</td><td colspan=\"2\">53.40 (\u00b112.87) 53.17 (\u00b17.83)</td><td>0.000</td><td>20.36 (\u00b10.66)</td></tr><tr><td>A2R+P</td><td>0.488</td><td>0.500</td><td colspan=\"2\">0.001 (\u00b10.001) 0.000 (\u00b10.000)</td><td>0.951</td><td colspan=\"2\">73.59 (\u00b10.81) 67.63 (\u00b11.54)</td><td>0.012</td><td>20.91 (\u00b10.48)</td></tr><tr><td>UNIREX (DLM-P)</td><td>0.751</td><td>0.267</td><td colspan=\"2\">0.180 (\u00b10.016) 0.390 (\u00b10.035)</td><td>0.997</td><td colspan=\"2\">76.07 (\u00b11.63) 69.76 (\u00b10.27)</td><td>0.990</td><td>64.13 (\u00b10.46)</td></tr><tr><td>UNIREX (AA-FP)</td><td>0.685</td><td>0.551</td><td colspan=\"2\">0.395 (\u00b10.109) 0.381 (\u00b10.101)</td><td>0.537</td><td colspan=\"2\">45.21 (\u00b14.46) 53.91 (\u00b13.23)</td><td>0.968</td><td>63.14 (\u00b10.33)</td></tr><tr><td>UNIREX (DLM-FP)</td><td>0.814</td><td>0.492</td><td colspan=\"2\">0.293 (\u00b10.043) 0.321 (\u00b10.070)</td><td>0.997</td><td colspan=\"2\">76.38 (\u00b10.57) 69.52 (\u00b10.24)</td><td>0.953</td><td>62.50 (\u00b11.34)</td></tr><tr><td>UNIREX (SLM-FP)</td><td>0.807</td><td>0.494</td><td colspan=\"2\">0.390 (\u00b10.087) 0.424 (\u00b10.110)</td><td>0.983</td><td colspan=\"2\">75.12 (\u00b10.41) 69.25 (\u00b10.41)</td><td>0.944</td><td>62.09 (\u00b12.12)</td></tr></table>"
            },
            "TABREF9": {
                "type_str": "table",
                "html": null,
                "text": "Main Results on CoS-E.",
                "num": null,
                "content": "<table><tr><td>Method</td><td>Composite NRG (\u2191)</td><td>NRG (\u2191)</td><td>Faithfulness Comp (\u2191)</td><td>Suff (\u2193)</td><td>NRG (\u2191)</td><td>Plausibility AUPRC (\u2191)</td><td>TF1 (\u2191)</td><td colspan=\"2\">Performance NRG (\u2191) F1 (\u2191)</td></tr><tr><td>AA (Grad)</td><td>0.498</td><td>0.462</td><td colspan=\"2\">0.222 (\u00b10.028) 0.120 (\u00b10.018)</td><td>0.035</td><td colspan=\"2\">22.27 (\u00b10.17) 13.81 (\u00b10.00)</td><td>0.997</td><td>69.80 (\u00b10.60)</td></tr><tr><td>AA (Input*Grad)</td><td>0.506</td><td>0.289</td><td colspan=\"2\">0.225 (\u00b10.048) 0.260 (\u00b10.059)</td><td>0.231</td><td colspan=\"2\">18.51 (\u00b10.23) 43.45 (\u00b10.05)</td><td>0.997</td><td>69.80 (\u00b10.60)</td></tr><tr><td>AA (DeepLIFT)</td><td>0.493</td><td>0.249</td><td colspan=\"2\">0.225 (\u00b10.012) 0.292 (\u00b10.014)</td><td>0.234</td><td colspan=\"2\">18.80 (\u00b10.19) 43.51 (\u00b10.04)</td><td>0.997</td><td>69.80 (\u00b10.60)</td></tr><tr><td>AA (IG)</td><td>0.499</td><td>0.280</td><td colspan=\"2\">0.162 (\u00b10.086) 0.222 (\u00b10.086)</td><td>0.220</td><td colspan=\"2\">18.71 (\u00b10.40) 41.79 (\u00b11.33)</td><td>0.997</td><td>69.80 (\u00b10.60)</td></tr><tr><td>L2E</td><td>0.522</td><td>0.366</td><td colspan=\"2\">0.007 (\u00b10.006) 0.042 (\u00b10.024)</td><td>0.205</td><td colspan=\"2\">24.48 (\u00b12.71) 32.63 (\u00b16.12)</td><td>0.997</td><td>69.80 (\u00b10.60)</td></tr><tr><td>SGT</td><td>0.594</td><td>0.564</td><td colspan=\"2\">0.214 (\u00b10.105) 0.033 (\u00b10.077)</td><td>0.224</td><td colspan=\"2\">18.60 (\u00b10.42) 42.42 (\u00b10.51)</td><td>0.995</td><td>69.73 (\u00b10.13)</td></tr><tr><td>FRESH</td><td>0.675</td><td>0.571</td><td colspan=\"2\">0.176 (\u00b10.029) 0.000 (\u00b10.000)</td><td>0.617</td><td colspan=\"2\">24.68 (\u00b17.98) 48.02 (\u00b13.04)</td><td>0.838</td><td>64.47 (\u00b13.41)</td></tr><tr><td>A2R</td><td>0.217</td><td>0.404</td><td colspan=\"2\">-0.010 (\u00b10.029) 0.000 (\u00b10.000)</td><td>0.249</td><td colspan=\"2\">18.72 (\u00b10.67) 45.45 (\u00b10.02)</td><td>0.000</td><td>36.39 (\u00b10.00)</td></tr><tr><td>UNIREX (AA-F)</td><td>0.711</td><td>0.956</td><td colspan=\"2\">0.505 (\u00b10.050) -0.071 (\u00b10.020)</td><td>0.236</td><td colspan=\"2\">18.82 (\u00b10.40) 43.68 (\u00b10.38)</td><td>0.939</td><td>66.17 (\u00b14.58)</td></tr><tr><td>SGT+P</td><td>0.630</td><td>0.665</td><td colspan=\"2\">0.280 (\u00b10.029) 0.283 (\u00b10.039)</td><td>0.226</td><td colspan=\"2\">18.63 (\u00b10.52) 42.71 (\u00b10.39)</td><td>1.000</td><td>69.91 (\u00b10.81)</td></tr><tr><td>FRESH+P</td><td>0.404</td><td>0.413</td><td colspan=\"2\">0.000 (\u00b10.013) 0.000 (\u00b10.000)</td><td>0.739</td><td colspan=\"2\">55.87 (\u00b110.13) 63.70 (\u00b19.58)</td><td>0.060</td><td>38.41 (\u00b15.34)</td></tr><tr><td>A2R+P</td><td>0.516</td><td>0.422</td><td colspan=\"2\">0.011 (\u00b10.024) 0.000 (\u00b10.000)</td><td>0.977</td><td colspan=\"2\">70.86 (\u00b11.30) 76.21 (\u00b11.68)</td><td>0.150</td><td>41.42 (\u00b18.73)</td></tr><tr><td>UNIREX (DLM-P)</td><td>0.708</td><td>0.123</td><td colspan=\"2\">0.127 (\u00b10.010) 0.322 (\u00b10.017)</td><td>0.999</td><td colspan=\"2\">71.80 (\u00b10.27) 77.94 (\u00b10.57)</td><td>1.000</td><td>69.91 (\u00b10.76)</td></tr><tr><td>UNIREX (AA-FP)</td><td>0.706</td><td>1.000</td><td colspan=\"2\">0.545 (\u00b10.045) -0.077 (\u00b10.099)</td><td>0.231</td><td colspan=\"2\">19.13 (\u00b10.71) 42.66 (\u00b11.18)</td><td>0.888</td><td>66.17 (\u00b14.58)</td></tr><tr><td>UNIREX (DLM-FP)</td><td>0.751</td><td>0.327</td><td colspan=\"2\">0.135 (\u00b10.072) 0.165 (\u00b10.029)</td><td>0.998</td><td colspan=\"2\">71.89 (\u00b10.41) 77.63 (\u00b10.62)</td><td>0.929</td><td>67.53 (\u00b11.06)</td></tr><tr><td>UNIREX (SLM-FP)</td><td>0.784</td><td>0.377</td><td colspan=\"2\">0.198 (\u00b10.038) 0.171 (\u00b10.027)</td><td>0.997</td><td colspan=\"2\">71.69 (\u00b10.21) 77.79 (\u00b10.09)</td><td>0.979</td><td>69.20 (\u00b11.58)</td></tr></table>"
            },
            "TABREF10": {
                "type_str": "table",
                "html": null,
                "text": "Main Results on MultiRC.",
                "num": null,
                "content": "<table><tr><td>Method</td><td>Composite NRG (\u2191)</td><td>NRG (\u2191)</td><td>Faithfulness Comp (\u2191)</td><td>Suff (\u2193)</td><td>NRG (\u2191)</td><td>Plausibility AUPRC (\u2191)</td><td>TF1 (\u2191)</td><td colspan=\"2\">Performance NRG (\u2191) F1 (\u2191)</td></tr><tr><td>AA (Grad)</td><td>0.587</td><td>0.518</td><td colspan=\"2\">0.313 (\u00b10.009) 0.380 (\u00b10.025)</td><td>0.244</td><td>59.80 (\u00b11.32)</td><td>15.27 (\u00b10.00)</td><td>0.999</td><td>90.78 (\u00b10.27)</td></tr><tr><td>AA (Input*Grad)</td><td>0.503</td><td>0.287</td><td colspan=\"2\">0.205 (\u00b10.005) 0.446 (\u00b10.020)</td><td>0.223</td><td>32.98 (\u00b11.37)</td><td>43.13 (\u00b10.86)</td><td>0.999</td><td>90.78 (\u00b10.27)</td></tr><tr><td>AA (DeepLIFT)</td><td>0.508</td><td>0.270</td><td colspan=\"2\">0.195 (\u00b10.012) 0.448 (\u00b10.014)</td><td>0.254</td><td>33.47 (\u00b11.31)</td><td>46.44 (\u00b10.04)</td><td>0.999</td><td>90.78 (\u00b10.27)</td></tr><tr><td>AA (IG)</td><td>0.596</td><td>0.473</td><td colspan=\"2\">0.308 (\u00b10.011) 0.414 (\u00b10.020)</td><td>0.317</td><td>47.83 (\u00b11.04)</td><td>37.87 (\u00b11.39)</td><td>0.999</td><td>90.78 (\u00b10.27)</td></tr><tr><td>L2E</td><td>0.606</td><td>0.460</td><td colspan=\"2\">0.009 (\u00b10.015) 0.036 (\u00b10.022)</td><td>0.358</td><td>58.11 (\u00b10.97)</td><td>31.35 (\u00b10.27)</td><td>0.999</td><td>90.78 (\u00b10.27)</td></tr><tr><td>SGT</td><td>0.595</td><td>0.503</td><td colspan=\"2\">0.288 (\u00b10.025) 0.361 (\u00b10.038)</td><td>0.298</td><td>42.46 (\u00b13.03)</td><td>41.70 (\u00b11.78)</td><td>0.985</td><td>90.23 (\u00b10.16)</td></tr><tr><td>FRESH</td><td>0.518</td><td>0.661</td><td colspan=\"2\">0.120 (\u00b10.075) 0.000 (\u00b10.000)</td><td>0.361</td><td>38.77 (\u00b16.82)</td><td>53.71 (\u00b13.30)</td><td>0.530</td><td>72.92 (\u00b18.71)</td></tr><tr><td>A2R</td><td>0.273</td><td>0.564</td><td colspan=\"2\">0.053 (\u00b10.048) 0.000 (\u00b10.000)</td><td>0.256</td><td colspan=\"2\">48.48 (\u00b111.14) 29.54 (\u00b124.72)</td><td>0.000</td><td>52.72 (\u00b114.08)</td></tr><tr><td>UNIREX (AA-F)</td><td>0.622</td><td>0.539</td><td colspan=\"2\">0.330 (\u00b10.018) 0.383 (\u00b10.055)</td><td>0.340</td><td>45.29 (\u00b13.02)</td><td>43.69 (\u00b11.98)</td><td>0.987</td><td>90.31 (\u00b10.19)</td></tr><tr><td>SGT+P</td><td>0.608</td><td>0.524</td><td colspan=\"2\">0.286 (\u00b10.034) 0.339 (\u00b10.032)</td><td>0.311</td><td>43.03 (\u00b11.69)</td><td>42.59 (\u00b11.63)</td><td>0.988</td><td>90.36 (\u00b10.08)</td></tr><tr><td>FRESH+P</td><td>0.614</td><td>0.695</td><td colspan=\"2\">0.143 (\u00b10.072) 0.000 (\u00b10.000)</td><td>0.603</td><td colspan=\"2\">56.21 (\u00b110.47) 64.09 (\u00b15.59)</td><td>0.544</td><td>73.44 (\u00b112.88)</td></tr><tr><td>A2R+P</td><td>0.800</td><td>0.751</td><td colspan=\"2\">0.182 (\u00b10.097) 0.000 (\u00b10.000)</td><td>0.992</td><td>87.30 (\u00b10.44)</td><td>77.31 (\u00b10.72)</td><td>0.656</td><td>77.31 (\u00b10.72)</td></tr><tr><td>UNIREX (DLM-P)</td><td>0.842</td><td>0.525</td><td colspan=\"2\">0.311 (\u00b10.011) 0.371 (\u00b10.032)</td><td>1.000</td><td>87.85 (\u00b10.13)</td><td>77.63 (\u00b10.35)</td><td>1.000</td><td>90.80 (\u00b10.33)</td></tr><tr><td>UNIREX (AA-FP)</td><td>0.626</td><td>0.529</td><td colspan=\"2\">0.341 (\u00b10.008) 0.406 (\u00b10.046)</td><td>0.363</td><td>44.79 (\u00b10.81)</td><td>47.18 (\u00b10.83)</td><td>0.985</td><td>90.21 (\u00b10.08)</td></tr><tr><td>UNIREX (DLM-FP)</td><td>0.857</td><td>0.588</td><td colspan=\"2\">0.335 (\u00b10.018) 0.346 (\u00b10.023)</td><td>0.991</td><td>86.99 (\u00b10.40)</td><td>77.53 (\u00b10.15)</td><td>0.992</td><td>90.51 (\u00b10.12)</td></tr><tr><td>UNIREX (SLM-FP)</td><td>0.864</td><td>0.603</td><td colspan=\"2\">0.353 (\u00b10.017) 0.356 (\u00b10.015)</td><td>0.994</td><td>87.58 (\u00b10.14)</td><td>77.22 (\u00b10.28)</td><td>0.994</td><td>90.59 (\u00b10.09)</td></tr></table>"
            },
            "TABREF11": {
                "type_str": "table",
                "html": null,
                "text": "Main Results on e-SNLI.",
                "num": null,
                "content": "<table><tr><td>Ablation</td><td>Method</td><td>Performance Acc (\u2191)</td><td>CSD (\u2191)</td><td>Faithfulness Comp (\u2191)</td><td>Suff (\u2193)</td><td>Plausibility AUPRC (\u2191) TF1 (\u2191)</td></tr><tr><td/><td>UNIREX (AA-F, Rand)</td><td/><td/><td/><td/><td/></tr><tr><td>Ext Type (F)</td><td/><td/><td/><td/><td/><td/></tr></table>"
            }
        }
    }
}