Efficient and Interpretable Grammatical Error Correction with Mixture of Experts

TL;DR

A mixture-of-experts model, MoECE, for grammatical error correction that achieves the performance of T5-XL with three times fewer effective parameters and produces interpretable corrections by also identifying the error type during inference.

Abstract

Error type information has been widely used to improve the performance of grammatical error correction (GEC) models, whether for generating corrections, re-ranking them, or combining GEC models. Combining GEC models that have complementary strengths in correcting different error types is very effective in producing better corrections. However, system combination incurs a high computational cost due to the need to run inference on the base systems before running the combination method itself. Therefore, it would be more efficient to have a single model with multiple sub-networks that specialize in correcting different error types. In this paper, we propose a mixture-of-experts model, MoECE, for grammatical error correction. Our model successfully achieves the performance of T5-XL with three times fewer effective parameters. Additionally, our model produces interpretable corrections by also identifying the error type during inference.

Figures (7)

• Figure 1: $F_{0.5}$ scores of a T5-v1.1-Base model on the six most frequent error types and all error types (ALL) in the BEA-2019 development set at different numbers of training steps.
• Figure 2: Illustration of a transformer block with an MoE layer with $M=4$ and $K=2$. To simplify the notation, ${\bm{x}}$ represents the input of the MoE layer and the feed-forward layer instead of the transformer block.
• Figure 3: $F_{0.5}$ scores of comparable models that produce the error type of corrections according to the effective parameter count (in millions). Legends: [1] lai-etal-2022-type [2] tarnavskyi-etal-2022-ensembling_short, [3] rothe-etal-2021-simple_short, [4] omelianchuk-etal-2020-gector_short, [5] stahlberg-kumar-2020-seq2edits_short. We do not compare against sun-wang-2022-adjusting and bout-etal-2023-efficient which only produce the correction tokens.
• Figure 4: Average routing score of MoECE-GS-Base for each token based on the error type.
• Figure 5: Experts' correction accuracy for each token based on the error type.