Insertionbased

Insertionbased is a term used in natural language processing to describe a family of sequence generation methods in which the output is constructed by iteratively inserting tokens into a partially built sequence, rather than predicting tokens strictly from left to right. The approach enables non-sequential generation and can leverage parallel insertions, potentially speeding up decoding while maintaining competitive output quality.

The best-known example is the Insertion Transformer, introduced as an alternative to left-to-right autoregressive decoding. Insertion-based models generate text by inserting tokens at chosen positions in the current sequence rather than extending it one token at a time from the end.

How it works: a generation process starts from a seed sequence, often a start token. The model proposes insertions: a set of positions between existing tokens and tokens to be inserted, along with probabilities. In a single decoding step, multiple tokens can be inserted, and after several steps the final sequence emerges. Training typically uses teacher forcing to align predicted insertions with the target sequence, with the model learning an insertion plan or order. Variants explore different planners or hierarchical schemes to organize insertions.

Advantages and limitations: insertion-based methods can offer faster decoding due to parallel insertions and can better capture non-local dependencies or alternative token orders. They also allow flexible output lengths. However, they involve more complex training and inference machinery, require reliable insertion plans, and may underperform baseline left-to-right models without careful design and tuning in some tasks.

Applications and scope: these models have been explored for language generation, machine translation, summarization, and code generation, among other sequence-generation tasks. Related concepts include non-autoregressive generation and traditional left-to-right autoregressive models.