Sequence Modeling with Unconstrained Generation Order
This addresses a fundamental limitation in sequence generation for applications like translation and captioning, though it appears incremental as it builds on existing sequence modeling paradigms.
The paper tackles the problem of rigid generation order in sequence models by proposing a model that can generate output sequences by inserting tokens in any arbitrary order, learning decoding order through training. Experiments show this model outperforms fixed-order models on multiple tasks including Machine Translation, Image-to-LaTeX, and Image Captioning.
The dominant approach to sequence generation is to produce a sequence in some predefined order, e.g. left to right. In contrast, we propose a more general model that can generate the output sequence by inserting tokens in any arbitrary order. Our model learns decoding order as a result of its training procedure. Our experiments show that this model is superior to fixed order models on a number of sequence generation tasks, such as Machine Translation, Image-to-LaTeX and Image Captioning.