In this paper, Seq2Edits: Sequence Transduction Using Span-level Edit Operations, I found all tasks are trained for 1M i terations on 170M sentences extracted from English Wikipedia revisions and 176M sentences from English Wikipedia round-trip translated via German. Whether to get tag and position information between source and target sentences during pre-training? Another question about start and end position, target position is used in code rather than souce position. Which position information should be selected?
Description
In this paper, Seq2Edits: Sequence Transduction Using Span-level Edit Operations, I found all tasks are trained for 1M i terations on 170M sentences extracted from English Wikipedia revisions and 176M sentences from English Wikipedia round-trip translated via German. Whether to get tag and position information between source and target sentences during pre-training? Another question about start and end position, target position is used in code rather than souce position. Which position information should be selected?
Thanks!