It lets the decoder use context from both sides of a word, matching BERT’s pretraining and improving fluency.