edit-based generation

**Edit-Based Generation** is a **family of text generation approaches that produce output by applying a sequence of edit operations to an initial sequence** — rather than generating text from scratch, edit-based models transform an existing sequence (draft, template, or source) through insertions, deletions, replacements, and reorderings. **Edit-Based Methods** - **LaserTagger**: Predicts edit operations (KEEP, DELETE, INSERT) for each input token — efficient for text editing tasks. - **GEC (Grammatical Error Correction)**: Detect and correct specific errors — edit-based approach is natural for correction. - **Seq2Edits**: Convert seq2seq problems into edit prediction problems — more efficient for tasks where output is similar to input. - **Levenshtein Transformer**: General-purpose edit-based generation with learned operations. **Why It Matters** - **Efficiency**: When output is similar to input (editing, correction, paraphrasing), edit-based models avoid redundant generation of unchanged portions. - **Controllability**: Edit operations are interpretable — can constrain the types of changes allowed. - **Speed**: For editing tasks, predicting edits is much faster than regenerating the entire output. **Edit-Based Generation** is **text as revision** — generating output by applying targeted edit operations to an existing sequence rather than writing from scratch.