Enabling Robust Grammatical Error Correction in New Domains: Data Sets, Metrics, and Analyses

@article{doi:10.1162/tacl_a_00282,
author = {Napoles, Courtney and Nădejde, Maria and Tetreault, Joel},
title = {Enabling Robust Grammatical Error Correction in New Domains: Data Sets, Metrics, and Analyses},
journal = {Transactions of the Association for Computational Linguistics},
volume = {7},
number = {},
pages = {551-566},
year = {2019},
doi = {10.1162/tacl_a_00282},
URL = {https://doi.org/10.1162/tacl_a_00282},
eprint = {https://doi.org/10.1162/tacl_a_00282},
abstract = { Until now, grammatical error correction (GEC) has been primarily evaluated on text written by non-native English speakers, with a focus on student essays. This paper enables GEC development on text written by native speakers by providing a new data set and metric. We present a multiple-reference test corpus for GEC that includes 4,000 sentences in two new domains (formal and informal writing by native English speakers) and 2,000 sentences from a diverse set of non-native student writing. We also collect human judgments of several GEC systems on this new test set and perform a meta-evaluation, assessing how reliable automatic metrics are across these domains. We find that commonly used GEC metrics have inconsistent performance across domains, and therefore we propose a new ensemble metric that is robust on all three domains of text.}
}

Repository Contents

Data

data/ contains the dev and test splits, with a subdirectory for each domain containing

• the original sentences (source)
• system outputs (amu, lstm, lstm-r, marian, nus, transformer)
• human corrections (ref[0-3])
• negative control used for collecting human ratings (source+error)

Domains are fce, wiki, and yahoo.

DOMAIN-corpus-scores.csv has the mean human rating for each system for that domain. DOMAIN-segment-scores.csv has the mean human rating by sentence for each system.

Data from the yahoo domain was sampled from the Yahoo Answers corpus, created from L6 - Yahoo! Answers Comprehensive Questions and Answers version 1.0. This Yahoo Answers corpus can be requested free of charge for research purposes. Access to data from the yahoo domain will require you to first gain access to this Yahoo Answers corpus.

Once you have gained access to the L6 corpus, please forward the acknowledgment to Grammarly (peng.wang@grammarly.com), along with your affiliation and a short description of how you will be using the data, and we will provide access to data from the yahoo domain.