[JOSHUA-129] Constrained decoding heuristics aren't quite right

• Type: Bug
• Status: Reopened
• Resolution: Unresolved
• Fix Version/s: 6.2
• Assignee: Unassigned
• Reporter: Matt Post
• Created: 26/Jul/13 02:02
• Updated: 20/Jun/16 18:29

Description

To implement constrained decoding, we make sure that the LM state of each proposed node matches the input sentence. This is necessary but insufficient. See the test case in test/decoder/constrained, where some constrained sentences have multiple outputs not all equal to the constraints.