To Err Is Human, but Llamas Can Learn It Too

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Luhtaru, Agnes, Purason, Taido, Vainikko, Martin, Del, Maksym, Fishel, Mark
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912057776930816
author Luhtaru, Agnes
Purason, Taido
Vainikko, Martin
Del, Maksym
Fishel, Mark
author_facet Luhtaru, Agnes
Purason, Taido
Vainikko, Martin
Del, Maksym
Fishel, Mark
contents This study explores enhancing grammatical error correction (GEC) through artificial error generation (AEG) using language models (LMs). Specifically, we fine-tune Llama 2-based LMs for error generation and find that this approach yields synthetic errors akin to human errors. Next, we train GEC Llama models with the help of these artificial errors and outperform previous state-of-the-art error correction models, with gains ranging between 0.8 and 6 F0.5 points across all tested languages (German, Ukrainian, and Estonian). Moreover, we demonstrate that generating errors by fine-tuning smaller sequence-to-sequence models and prompting large commercial LMs (GPT-3.5 and GPT-4) also results in synthetic errors beneficially affecting error generation models.
format Preprint
id arxiv_https___arxiv_org_abs_2403_05493
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle To Err Is Human, but Llamas Can Learn It Too
Luhtaru, Agnes
Purason, Taido
Vainikko, Martin
Del, Maksym
Fishel, Mark
Computation and Language
This study explores enhancing grammatical error correction (GEC) through artificial error generation (AEG) using language models (LMs). Specifically, we fine-tune Llama 2-based LMs for error generation and find that this approach yields synthetic errors akin to human errors. Next, we train GEC Llama models with the help of these artificial errors and outperform previous state-of-the-art error correction models, with gains ranging between 0.8 and 6 F0.5 points across all tested languages (German, Ukrainian, and Estonian). Moreover, we demonstrate that generating errors by fine-tuning smaller sequence-to-sequence models and prompting large commercial LMs (GPT-3.5 and GPT-4) also results in synthetic errors beneficially affecting error generation models.
title To Err Is Human, but Llamas Can Learn It Too
topic Computation and Language
url https://arxiv.org/abs/2403.05493