add ACL/010_CLEME20_Towards_Interpretable_Evaluation_by_Disentangling_Ed.txt

ACL/010_CLEME20_Towards_Interpretable_Evaluation_by_Disentangling_Ed.txt

@@ -0,0 +1,96 @@
+Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 204–222 July 27 - August 1, 2025 ©2025 Association for Computational Linguistics CLEME2.0: Towards Interpretable Evaluation by Disentangling Edits for Grammatical Error Correction Jingheng Ye1*, Zishan Xu1*, Yinghui Li1, Linlin Song2, Qingyu Zhou3, Hai-Tao Zheng1,4†, Ying Shen5, Wenhao Jiang6, Hong-Gee Kim7, Ruitong Liu1, Xin Su8, Zifei Shan8 1Tsinghua University, 2Huazhong University of Science and Technology, 3ByteDance Inc., 4Peng Cheng Laboratory, 5Sun-Yat Sen University, 6Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), 7Seoul National University, 8Tencent {yejh22,xzs23}@mails.tsinghua.edu.cn Abstract The paper focuses on the interpretability of Grammatical Error Correction (GEC) evalua- tion metrics, which received little attention in previous studies. To bridge the gap, we intro- duce CLEME2.0, a reference-based metric de- scribing four fundamental aspects of GEC sys- tems: hit-correction, wrong-correction, under- correction, and over-correction. They collec- tively contribute to exposing critical qualities and locating drawbacks of GEC systems. Eval- uating systems by combining these aspects also leads to superior human consistency over other reference-based and reference-less met- rics. Extensive experiments on two human judgment datasets and six reference datasets demonstrate the effectiveness and robustness of our method, achieving a new state-of-the- art result. Our codes are released at https: //github.com/THUKElab/CLEME. 1
+Introduction The task of Grammatical Error Correction (GEC) automatically detects and corrects grammatical er- rors in a given text (Bryant et al., 2023). A core component of GEC is the development of auto- matic metrics that can objectively measure model performance (Kobayashi et al., 2024b; Ye et al., 2023c). However, proposing appropriate evalua- tion of GEC has long been a challenging task (Mad- nani et al., 2011), due to the subjectivity (Bryant and Ng, 2015), complexity (Mita et al., 2019) and subtlety (Choshen and Abend, 2018) of GEC. Recent research efforts have focused on devel- oping GEC metrics that closely align with human judgements (Koyama et al., 2024), whereas the interpretability of these metrics has received less emphasis. We define the interpretability of metrics as their capacity to disclose concerned character- istics of systems, which is crucial for identifying *Equal Contribution. †Corresponding Author: Hai-Tao Zheng. (E-mail: zheng.haitao@sz.tsinghua.edu.cn) Ref. PRF-based Metrics (CLEME) Hyp. 1 CLEME2.0 (Ours) True Negative (TN) False Positive (FP) False Negative (FN) True Positive (TP) Necessary False Positive (FPne) Unnecessary False Positive (FPun) Hyp. 2 Source Nowadays technologies have improved a lot compared to the last century. Nowadays technologies was in the last century. improved a lot compared Nowadays were the last century. for technologies improved a lot compared of Nowadays were the last century. the for technologies improved a lot compared Source Nowadays were the last century. the for technologies improved a lot compared Ref Nowadays technologies have improved a lot compared to the last century. Hyp. 1 Nowadays technologies was in the last century. improved a lot compared Hyp. 2 Nowadays were the last century. for technologies improved a lot compared of Ref. PRF-based Metrics (ERRANT) Hyp. 1 Hyp. 2 Source Nowadays technologies have improved a lot compared to the last century. Nowadays technologies was in the last century. improved a lot compared Nowadays were the last century. for technologies improved a lot compared of Nowadays were the last century. the for technologies improved a lot compared FP + FN Figure 1: An example of CLEME2.0. We highlight TP, FP, FPne, FPun, and FN in different colors. weaknesses in a given GEC system. It is gener- ally recognized that excellent systems should ad- here to the principles of grammaticality and faith- fulness (Bryant et al., 2023). Grammaticality de- mands that all grammatical errors be accurately cor- rected, while faithfulness ensures that corrections retain the original meaning and syntactic struc- ture. Nevertheless, the commonly utilized GEC metrics (Bryant et al., 2017; Dahlmeier and Ng, 2012a) are PRF-based (Precision, Recall, and F scores). We claim that PRF-based metrics fail to effectively capture subtle dimensions of GEC systems, consequently hindering progress. As il- lustrated in Figure 1, the edits [were→was] and [for→in] in Hyp. 1 are regarded as 2 FP + FN edits by ERRANT (Bryant et al., 2017) or 2 FP edits by CLEME (Ye et al., 2023c). Meanwhile, the edit [ϵ→of] in Hyp. 2 is categorized as an FP edit for both ERRANT and CLEME. Despite this, these two categories of FP edits carry different im- plications. The former type is correctly placed but wrongly modified, whereas the latter is incorrectly positioned. The inability to differentiate between these FP edits results in ambiguous interpretations 204
+
+of P/R/F0.5 scores, as they fail to quantify gram- maticality and faithfulness. Thus, we introduce CLEME2.0, an interpretable reference-based metric describing four fundamen- tal aspects of GEC systems: 1) the hit-correction score reflects the degree to which a system accu- rately corrects grammatical errors, 2) the wrong- correction score denotes the degree of incorrect corrections made, 3) the under-correction score re- veals the degree of missing corrections, and 4) the over-correction score measures the degree of exces- sive corrections. An excellent GEC system should gain a higher hit-correction score and lower wrong- correction, under-correction, and over-correction scores. The initial three aspects assess grammati- cality, whereas the over-correction score pertains to faithfulness, given that it often alters the orig- inal meaning, a challenge notably observed with LLMs (Coyne et al., 2023; Li et al., 2023a). To achieve this, CLEME2.0 first distinguishes be- tween necessary and unnecessary false positive (FP) edits. The idea is that necessary FP edits indi- cate the system’s wrong-correction degree, while unnecessary FP edits reveal the system’s over- correction degree. As shown in the bottom block of Figure 1, [were→was] and [for→in] in Hyp. 1 are regarded as FPne edits, while [ϵ→of] in Hyp. 2 is considered as an FPun edit. As a result, CLEME2.0 establishes a one-to-one relationship between four distinct system aspects and four types of edits: hit-correction v.s. TP, wrong-correction v.s. FPne, under-correction v.s. FN, and over-correction v.s. FPun. Unlike conven- tional GEC metrics like ERRANT (Bryant et al., 2017) and MaxMatch (Dahlmeier and Ng, 2012a) that evaluate using P/R/F0.5 scores, it offers a nu- anced view into the detailed aspects necessary for characterizing critical features of GEC systems. These separated scores are then consolidated into an overall score via linear weighted summation, giving varying importance to these distinct scores. This aggregate score provides a holistic measure of system performance. Similar to CLEME, our method adopts the chunk partition technique and supports evaluations based on either correction de- pendence or correction independence assumptions, so we dub the metric as CLEME2.0. Moreover, we propose that edits of varying modi- fication levels should uniquely influence the evalua- tion outcomes. For example, corrections involving punctuation are often less significant than correc- tions of content words. Therefore, we integrate two edit weighting techniques into CLEME2.0, similarity-based weighting (Gong et al., 2022) and LLM-based weighting. In particular, these methods compute a specific weight for each edit through a language model rather than assigning equal weight to all edits, thereby enabling CLEME2.0 to grasp contextual semantics and address the limitations of conventional metrics that depend on surface-level form similarity (Kobayashi et al., 2024a). To verify the effectiveness of CLEME2.0, we conduct extensive experiments on two human judg- ment datasets (GJG15 (Grundkiewicz et al., 2015) and SEEDA (Kobayashi et al., 2024b)), where our method consistently achieves high correlations. We also demonstrate the robustness of CLEME2.0 by computing the evaluation results based on six refer- ence datasets with disparate annotation styles. In summary, our contributions are three folds: (1) We introduce CLEME2.0, an interpretable reference-based metric, which is beneficial to reveal crucial aspects of GEC systems. (2) We enhance CLEME2.0 by incorporating two edit weighting techniques, addressing the lim- itations of conventional reference-based met- rics in capturing semantics. (3) Extensive experiments and analyses are con- ducted to confirm the effectiveness and robust- ness of our proposed method. 2
+Related Work Reference-based metrics. Reference-based met- rics evaluate GEC systems by comparing their outputs to manually written references (Ye et al., 2022, 2023a,b; Huang et al., 2023; Li et al., 2024c, 2022c,b, 2024d; Ma et al., 2022; Zhang et al., 2023, 2025a; Li et al., 2025c). The M2 scorer (Dahlmeier and Ng, 2012b) identifies optimal edit sequences between source sentences and system hypothe- ses, using F0.5 scores. However, this method can inflate scores by manipulating edit bound- aries. To mitigate this problem, ERRANT (Bryant et al., 2017) improves edit extraction through a linguistically-informed alignment algorithm, but it remains language-dependent and biased in multi- reference evaluation. CLEME (Ye et al., 2023c) fur- ther provides unbiased F0.5 scores and introduces an extra correction assumption for multi-reference evaluation. PT-M2 (Gong et al., 2022) combines PT-based and existing GEC metrics for higher cor- relations with human judgments. 205
+
+Src Ref Hyp Nowadays the technologies were improved lot compared for the last century. Nowadays technologies have to the last century. Nowadays technologies were in the last centuries. ���= �� ��+ ����+ ��= 1 3
+�����= ���� ��+ ����+ ��= 1 3
+�����= �� ��+ ����+ ��= 1 3
+����= ���� ��+ ����+ ���� = 1 3
+Similarity-based Weighting 
+LLM-based Weighting 
+the were have were for to in centuries. FPne TP FN FPun �����= �1 ∙���−�2 ∙(1 −�����) −�3 ∙(1 −�����) −�4 ∙(1 −����) = �1 ∙1 3   − �2 ∙(1 −1 3 ) −�3 ∙(1 −1 3 ) −�4 ∙(1 −1 3 ) LM for to in Nowadays the technologies 
+were improved lot compared 
+for the last century. Nowadays technologies have 
+improved lot compared  to  
+the last century. Nowadays the technologies 
+were improved lot compared 
+in the  last century. 0.11 
+ |s1-s2|
+ s1
+ s2 ② Disentangled Scores ① Edit Extraction improved lot compared improved lot compared ③ Comprehensive Scores 
+century. century. LLM
+ 4 
+Nowadays technologies have 
+improved lot compared  to  
+the last century. Nowadays technologies have 
+improved lot compared in the  
+last century. 1~5 for to in Figure 2: Overview of CLEME2.0. Initially, we extract edits and categorize hypothesis edits as TP, FN, FPne, and FPun. Next, we compute four distinct scores. Finally, we integrate these scores into an overall score utilizing one of the edit weighting techniques. Reference-less metrics. To overcome the limi- tations of reference-based metrics, recent studies focus on reference-less scoring. Inspired by qual- ity estimation in NMT (Liu et al., 2022; Dong et al., 2023), Napoles et al. (2016a) propose Grammaticality-Based Metrics (GBMs) using an existing GEC system or a pre-trained ridge regres- sion model. Asano et al. (2017) enhance GBMs by adding criteria like grammaticality, fluency, and meaning preservation. Yoshimura et al. (2020) in- troduce SOME, which uses sub-metrics optimized for manual assessment with regression models. Scribendi Score (Islam and Magnani, 2021) com- bines language perplexity and token/Levenshtein distance ratios. IMPARA (Maeda et al., 2022) in- corporates a Quality Estimator and a Semantic Es- timator based on BERT to evaluate GEC output quality and semantic similarity. While reference- less metrics align well with human judgments, they lack interpretability due to the heavy dependence on trained models, thus posing latent risks. 3
+Method Our CLEME2.0 can be generally divided into three main steps, with the overview shown in Fig- ure 2. Additionally, we incorporate two distinct edit weighting techniques to enhance performance. 3.1 Edit Extraction Given a source sentence X and a target (either hypothesis or reference) sentence Y , we extract the edits describing the modification from X to Y . Here, we utilize the chunk partition technique from CLEME (Ye et al., 2023c) to execute the pro- cess of edit extraction. Unlike the traditional met- rics like ERRANT (Bryant et al., 2017) and Max- Match (Dahlmeier and Ng, 2012a), CLEME con- currently aligns all sentences, including the source, the hypothesis, and all the references. This facili- tates segmentation of them all into chunk sequences with an equal number of chunks, irrespective of the varying token counts in different sentences, as de- lineated in Figure 2. It is worth noting that a chunk is a basic edit unit, which can be unchanged, cor- rected, or dummy (empty) (Ye et al., 2023c). 3.2 Disentangled Scores To compute disentangled scores, we initially dis- entangle edits into four elementary types. 1) TP edits refer to the corrected/dummy hypothesis chunks that share the same tokens as the corre- 206
+
+sponding reference chunks. 2) FPne edits are the corrected/dummy hypothesis chunks that have dif- ferent tokens from those in the corresponding ref- erence chunks wherein the reference chunks are also corrected/dummy ones. 3) FPun edits are the corrected hypothesis chunks but their correspond- ing reference chunks remain unchanged. 4) FN edits indicate the unchanged hypothesis chunks but the corresponding reference chunks are cor- rected/dummy. It is highlighted that traditional met- rics (Dahlmeier and Ng, 2012a; Bryant et al., 2017; Li et al., 2023c) do not distinguish between FPne and FPun, treating both as FP, thereby resulting in confusion between wrong-correction and over- correction. Actually, we have FP = FPne +FPun. Furthermore, we can differentiate between nec- essary and unnecessary edits. TP, FPne, and FN edits are all necessary edits, since their correspond- ing reference chunks are also corrected/dummy, implying the existence of grammatical errors in the related parts of X. On the contrary, FPun edit are unnecessary edits because the systems propose corrections not represented in references. Conse- quently, we can define four disentangled scores. Hit-correction score. This paper defines the hit- correction score as the ratio of TP edits to all neces- sary reference edits. Its purpose is to quantify the accuracy with which systems offer correct correc- tions. The formula is as follows: Hit = TP necessity = TP TP + FPne + FN (1) Wrong-correction score. Conversely, the wrong- correction score is defined as the ratio of FPne edits to all necessary reference edits. This score seeks to evaluate the degree to which systems generate erroneous corrections for grammatical errors. The formula for this score is as follows: Wrong = FPne necessity = FPne TP + FPne + FN (2) Under-correction score. Similarly, the under- correction score is proposed to measure the degree to which systems omit to correct grammatical er- rors, which is computed as follows: Under = FN necessity = FN TP + FPne + FN (3) Over-correction score. The score is introduced in response to frequent observations that LLMs are prone to over-correcting texts. This score is deter- mined by the proportion of FPun edits to all hypoth- esis corrected/dummy edits, aiming to gauge the level to which systems offer excessive corrections: Over = FPun TP + FP = FPun TP + FPne + FPun (4) With the disentangled scores indicating disparate aspects of GEC systems, researchers can identify specific weaknesses and implement targeted im- provements without expensive human labor. 3.3 Comprehensive Score Once the four disentangled scores have been com- puted, they need to be merged into a comprehensive score that encapsulates the global performance of the systems. We employ a weighted summation approach to organize these four scores for inter- pretability and simplification. By definition, sys- tems with higher hit-correction scores are usually preferable, a tendency that inversely applies to the remaining scores. Thus, the comprehensive score can be calculated using the following formula: Score = α1 · Hit + α2 · (1 −Wrong) + α3 · (1 −Under) + α4 · (1 −Over) (5) where αi is the trade-off factor for each disentan- gled score, and we constrain that 0 < αi < 1 and α1 + α2 + α3 + α4 = 1. 3.4 Edit Weighting Existing reference-based metrics, such as ER- RANT and CLEME, depend heavily on superficial literal similarity. This means that, regardless of length or modification, all types of edits have equal weighting in the evaluation scores. This aspect fails to acknowledge that human evaluators might semantically consider the edits’ varying importance levels. Therefore, we introduce two distinct edit weighting techniques to compute the importance weights of edits. These weights are then incorpo- rated into the calculation of the aforementioned disentangled scores as depicted in Equation (1) ∼ (4). Take the hit-correction score as a typical exam- ple, we reformulate the Equation (1) as follows: Hit = wT P wT P + wF Pne + wF N (6) Similarity-based weighting. We use PTScore to assign edit weights (Gong et al., 2022). Since it per- forms based on BERTScore (Zhang et al., 2019), a 207
+
+tool for evaluating text generation through similar- ity scores, we refer to this technique as similarity- based weighting. The rationale is to prioritize edits with a more significant modification of the meaning and quality of the text. By simulating a partially accurate version X′ of the source sentence X, PTScore can associate specific weights to edits within a sentence. The computation process is as follows: X′ = replace(X, ehyp) (7) w = PTScore(X′, R) −PTScore(X, R) (8) where R is the reference sentence, while the func- tion replace() is used to replace a specific chunk of the source X with the corrected/dummy hypoth- esis chunk ehyp. A positive weight w > 0 indicates a beneficial correction, whereas a negative value suggests a wrong correction. The absolute value |w| is utilized as the edit weight following (Gong et al., 2022), and the significance1 of an edit in a sentence grows with a larger |w|. LLM-based weighting. Recent studies have be- gun investigating the effectiveness of LLM-based evaluation, known for their advanced semantic comprehension (Qin et al., 2024b), in assessing various NLP tasks (Pavlovic and Poesio, 2024; Sot- tana et al., 2023). Building on this trend, we prompt Llama-2-7B (Touvron et al., 2023) to assign edit weights from 1 to 5, where a higher value signifies more critical edits. This methodology is rooted in the idea that LLMs, due to their extensive train- ing on diverse data, are adept at grasping intricate language patterns and text structure. Detailed im- plementation instructions and the prompting frame- work are available in Appendix A. 4
+Experiments 4.1 Experimental Settings Human ranking datasets. We conduct compre- hensive experiments across two human judgment datasets with disparate annotation protocols. • GJG15 (Grundkiewicz et al., 2015) is con- structed to manually evaluate classical sys- tems (Junczys-Dowmunt and Grundkiewicz, 2014; Rozovskaya et al., 2014) in the CoNLL- 2014 shared task (Ng et al., 2014). 1For more detailed analysis, refer to our case study in Section 5.1 and PT-M2 (Gong et al., 2022). • SEEDA. Kobayashi et al. (2024b) reveal sev- eral shortcomings in GJS15 and subsequently propose SEEDA, an alternative dataset fea- turing human judgments across two levels of granularity. To align with the contemporary trend in GEC, SEEDA is primarily focused on mainstream neural-based systems. Both of human judgment datasets derive the over- all human rankings for all GEC systems by em- ploying Expected Wins (EW) (Bojar et al., 2013) and TrueSkill (TS) (Sakaguchi et al., 2014) meth- ods. Following the previous approaches (Ye et al., 2023c; Kobayashi et al., 2024b), we compute the Pearson (γ) and Spearman (ρ) correlations between metrics and human judgments, in order to ascertain the effectiveness and robustness of GEC metrics within the context of system-level ranking. Reference datasets. Reference-based metrics rely on a reference set to establish a system rank- ing list, the properties of which may significantly influence the performance of the metrics. To in- vestigate the impact of variable reference sets, we assess human consistency across 6 reference datasets. These datasets encompass a range of an- notation styles, and a number of human annota- tors, including CoNLL-2014 (Grundkiewicz et al., 2015), BN-10GEC (Bryant and Ng, 2015) and SN- 8GEC (Sakaguchi et al., 2016). Notably, SN-8GEC is partitioned into 4 sub-sets, i.e., Expert-Minimal, Expert-Fluency, Non-Expert-Minimal, and Non- Expert-Fluency. A more thorough breakdown of these datasets and the statistics is provided in Ap- pendix B. Corpus and sentence levels. GEC evaluation metrics can compute an overall system-level score for a given system in two settings (Gong et al., 2022). Given the metric M, source sentences S, hypothesis sentences H and reference sentences R, 1) corpus-level metrics compute the system score based on the whole corpus M(S, H, R), and 2) sentence-level metrics use the average of the sentence-level scores PI i M(Si, Hi, Ri)/I. Trade-off factors. We leverage a
+cross- evaluation search method to identify two optimal sets of trade-off factors for both the corpus and sen- tence levels. At the corpus level, the factors are as- signed as α1, α2, α3, α4 = 0.45, 0.35, 0.15, 0.05, while for the sentence level, they are adjusted to α1, α2, α3, α4 = 0.35, 0.25, 0.20, 0.20. The 208
+
+Metric CoNLL-2014 BN-10GEC E-Minimal E-Fluency NE-Minimal NE-Fluency Avg. EW TS EW TS EW TS EW TS EW TS EW TS M2 γ 0.623 0.672 0.547 0.610 0.597 0.650 0.590 0.659 0.575 0.634 0.582 0.649 0.616 ρ 0.687 0.720 0.648 0.692 0.654 0.703 0.654 0.709 0.577 0.648 0.648 0.703 0.670 GLEU γ 0.701 0.750 0.678 0.761 0.533 0.513 0.693 0.771 -0.044 -0.113 0.674 0.767 0.557 ρ 0.467 0.555 0.754 0.806 0.577 0.511 0.710 0.757 -0.005 -0.055 0.725 0.819 0.551 ERRANT γ 0.642 0.688 0.586 0.644 0.578 0.631 0.594 0.663 0.585 0.637 0.597 0.659 0.625 ρ 0.659 0.698 0.637 0.698 0.742 0.786 0.720 0.775 0.747 0.797 0.753 0.797 0.734 PT-M2 γ 0.693 0.737 0.650 0.706 0.626 0.667 0.621 0.681 0.630 0.675 0.620 0.682 0.666 ρ 0.758 0.769 0.690 0.824 0.709 0.736 0.758 0.802 0.736 0.758 0.758 0.802 0.758 CLEME-dep γ 0.648 0.691 0.602 0.656 0.594 0.644 0.589 0.654 0.595 0.643 0.612 0.673 0.633 ρ 0.709 0.742 0.692 0.747 0.797 0.813 0.714 0.775 0.786 0.835 0.720 0.791 0.760 CLEME-ind γ 0.649 0.691 0.609 0.659 0.593 0.643 0.587 0.653 0.601 0.647 0.611 0.672 0.635 ρ 0.709 0.731 0.692 0.747 0.791 0.802 0.731 0.791 0.797 0.841 0.714 0.786 0.761 CLEME2.0-dep (Ours) γ 0.700 0.765 0.675 0.745 0.690 0.768 0.695 0.788 0.702 0.778 0.704 0.800 0.734 ρ 0.665 0.736 0.626 0.692 0.736 0.808 0.742 0.830 0.775 0.846 0.599 0.714 0.730 CLEME2.0-ind (Ours) γ 0.718 0.777 0.731 0.793 0.708 0.784 0.736 0.824 0.757 0.826 0.801 0.848 0.775 ρ 0.665 0.736 0.698 0.758 0.736 0.808 0.742 0.830 0.775 0.846 0.670 0.769 0.753 CLEME2.0-sim-dep (Ours) γ 0.783 0.853 0.721 0.801 0.765 0.834 0.737 0.827 0.761 0.824 0.741 0.834 0.790 ρ 0.819 0.890 0.802 0.863 0.791 0.868 0.758 0.852 0.830 0.896 0.786 0.857 0.834 CLEME2.0-sim-ind (Ours) γ 0.806 0.871 0.772 0.839 0.780 0.841 0.761 0.844 0.782 0.834 0.798 0.877 0.817 ρ 0.846 0.901 0.835 0.885 0.819 0.885 0.758 0.852 0.846 0.896 0.863 0.923 0.859 SentM2 γ 0.871 0.864 0.567 0.646 0.805♣0.836♣0.655 0.732 0.729♣0.785♣0.621 0.699 0.734 ρ 0.731 0.758 0.593 0.648 0.806♣0.845♣0.731 0.764 0.797♣0.846♣0.632 0.687 0.737 SentGLEU γ 0.784 0.828 0.756 0.826 0.742♣0.773♣0.785 0.846 0.723♣0.762♣0.778 0.848 0.788 ρ 0.720 0.775 0.769 0.824 0.764♣0.797♣0.791 0.846 0.764♣0.830♣0.768 0.846 0.791 SentERRANT γ 0.870 0.846 0.885 0.896 0.768♣0.803♣0.806 0.732 0.710♣0.765♣0.793 0.847 0.810 ρ 0.742 0.747 0.786 0.830 0.775♣0.819♣0.813 0.764 0.780♣0.841♣0.830 0.857 0.799 SentPT-M2 γ 0.949 0.938 0.602♣0.682♣0.831♣0.855♣0.689 0.763 0.770♣0.822♣0.648 0.725 0.772 ρ 0.907 0.874 0.626♣0.670♣0.808♣0.819♣0.797 0.841 0.813♣0.857♣0.742 0.786 0.795 SentCLEME-dep γ 0.876 0.844 0.915 0.913 0.806♣0.838♣0.849 0.886 0.742♣0.795♣0.876 0.921 0.855 ρ 0.824 0.808 0.835 0.874 0.775♣0.819♣0.824 0.863 0.797♣0.846♣0.791 0.846 0.825 SentCLEME-ind γ 0.868 0.857 0.855♣0.876♣0.821♣0.856♣0.841 0.877 0.782♣0.831♣0.852 0.896 0.851 ρ 0.725 0.758 0.659♣0.714♣0.775♣0.819♣0.808 0.846 0.819♣0.874♣0.762 0.825 0.782 SentCLEME2.0-dep (Ours) γ 0.870 0.881 0.766 0.830 0.941♣0.954♣0.892 0.938 0.913♣0.918♣0.916 0.949 0.897 ρ 0.714 0.725 0.681 0.747 0.857♣0.885♣0.824 0.901 0.857���0.912♣0.720 0.791 0.801 SentCLEME2.0-ind (Ours) γ 0.866 0.881 0.799 0.853 0.941♣0.956♣0.915 0.952 0.915♣0.917♣0.883 0.904 0.899 ρ 0.709 0.720 0.681 0.747 0.879♣0.912♣0.857 0.923 0.824♣0.885♣0.654 0.720 0.793 SentCLEME2.0-sim-dep (Ours) γ 0.926 0.937 0.797 0.861 0.939♣0.948♣0.908 0.952 0.871♣0.872♣0.918 0.947 0.906 ρ 0.907 0.912 0.808 0.863 0.852♣0.879♣0.885 0.945 0.753♣0.780♣0.896 0.940 0.868 SentCLEME2.0-sim-ind (Ours) γ 0.915 0.936 0.808 0.866 0.945♣0.956♣0.923 0.963 0.885♣0.887♣0.931 0.961 0.915 ρ 0.868 0.879 0.753 0.824 0.863♣0.901♣0.879 0.956 0.775♣0.802♣0.835 0.923 0.855 Table 1: Correlation results on GJG15 Ranking. CLEME2.0-sim is based on similarity-based weighting. We highlight the highest scores in bold and the second-highest scores with underlines. ♣We exclude unchanged references for higher correlations due to low-quality annotations in some reference sets. Results without excluding references are presented in Appendix C.1. details of the chosen values of trade-off factors can be seen in Appendix B.5. Evaluation assumptions. CLEME can evaluate GEC systems based on correction dependence (- dep) or independence (-ind) assumptions. The cor- rection independence assumption offers a more re- laxed edit-matching process, implying that systems might yield better scores when multiple references are available. Inspired by this work, CLEME2.0 also supports both assumptions, and we will study their effects on our method. 4.2 Results of GJG15 Ranking The correlations between the GEC metrics and hu- man judgments on the GJG15 rankings are shown in Table 1, and we have the following insights. 209
+
+CLEME2.0 outperforms other metrics at both corpus and sentence levels. For corpus-level, CLEME2.0-sim-ind achieves the highest average correlations, closely followed by CLEME2.0-sim- dep. CLEME2.0-ind and CLEME2.0-dep can also gain comparable correlations with other metrics, even though they do not utilize any edit weighting techniques. On the other hand, sentence-level met- rics exhibit a similar pattern. SentCLEME2.0-sim- dep and SentCLEME2.0-sim-ind achieve the high- est Pearson and Spearson correlations, respectively. These results significantly demonstrate the effec- tiveness and robustness of our proposed method across different settings. Sentence-level metrics outperform their corpus- level counterparts. This observation is consis- tent with recent studies (Gong et al., 2022; Ye et al., 2023c). This is because system-level rankings treat each sample equally regardless of edit numbers, mirroring how sentence-level metrics are evalu- ated. On the other hand, corpus-level metrics em- phasize samples with more edits, thus causing the gap between automatic metrics and human evalu- ation. SentPT-M2 shows superior performance on CoNLL-2014 but performs worse on BN-10GEC, E-Minimal, and NE-Fluency compared to our ap- proach, revealing a lack of robustness of the metric. Generally, our method aligns more closely with human assessments than existing popular met- rics. Notably, our method with similarity-based weighting surpasses unweighted ones, thanks to the integration of semantic factors. However, on E-Minimal and NE-Minimal, weighted and un- weighted results are comparable. We suspect this is because these datasets have minimal yet cru- cial annotations, reducing the possibility of varying weights and the efficacy of edit weighting. Furthermore, we present comprehensive results of CLEME2.0 on CoNLL-2014 and offer insights into our method for analyzing and identifying weak- nesses in GEC systems in Appendix C.5. 4.3 Results of SEEDA Ranking We carry out an additional experiment on the SEEDA-Sentence and SEEDA-Edit datasets, where we compare our method against various GEC met- rics. As presented in Table 2, our approach con- sistently achieves the best outcomes across both datasets. According to Kobayashi et al. (2024b), the correlations of most metrics tend to decrease when transitioning from classical to neural evalua- Metric SEEDA-S SEEDA-E Avg. γ
+ρ γ
+ρ M2 0.658 0.487 0.791 0.764 0.675 PT-M2 0.845 0.769 0.896 0.909 0.855 ERRANT 0.557 0.406 0.697 0.671 0.583 PT-ERRANT 0.818 0.720 0.888 0.888 0.829 GoToScorer 0.929 0.881 0.901 0.937 0.912 GLEU 0.847 0.886 0.911 0.897 0.885 Scribendi Score 0.631 0.641 0.830 0.848 0.738 SOME 0.892 0.867 0.901 0.951 0.903 IMPARA 0.911 0.874 0.889 0.944 0.903 CLEME-dep 0.633 0.501 0.755 0.757 0.662 CLEME-ind 0.616 0.466 0.736 0.708 0.632 CLEME2.0-dep (Ours) 0.937 0.865 0.945 0.939 0.922 CLEME2.0-ind (Ours) 0.908 0.844 0.961 0.946 0.915 CLEME2.0-sim-dep (Ours) 0.923 0.914 0.948 0.974 0.940 CLEME2.0-sim-ind (Ours) 0.921 0.907 0.953 0.981 0.941 Sent-M2 0.802 0.692 0.887 0.846 0.807 SentERRANT 0.758 0.643 0.860 0.825 0.772 SentCLEME-dep 0.866 0.809 0.944 0.939 0.890 SentCLEME-ind 0.864 0.858 0.935 0.911 0.892 SentCLEME2.0-dep (Ours) 0.905 0.844 0.955 0.946 0.913 SentCLEME2.0-ind (Ours) 0.875 0.837 0.953 0.953 0.905 SentCLEME2.0-sim-dep (Ours) 0.924 0.858 0.923 0.953 0.915 SentCLEME2.0-sim-ind (Ours) 0.921 0.886 0.957 0.960 0.931 Table 2: Results of human correlations on SEEDA Rank- ing based on TrueSkill (TS). tion systems. This implies that conventional met- rics might face difficulties in evaluating the more extensively edited or fluent corrections produced by state-of-the-art neural GEC systems. Nevertheless, our method effectively tackles these challenges, delivering even improved performance for all indi- cators. The results for SEEDA-Edit exceed those for SEEDA-Sentence, due to the greater detail in SEEDA-Edit, aligning more closely with the oper- ation of CLEME2.0. It is crucial to mention that reference-less met- rics such as SOME and IMPARA yield high out- comes, in part, because these are fine-tuned on GEC data. Although fine-tuned metrics generally perform better, they are not without their limi- tations. Firstly, the incorporation of fine-tuning in SOME and IMPARA makes these reference- less metrics more costly. Second, these reference- less metrics may suffer from poor robustness since the assessment process is not guided by human-annotated references. For example, the au- thors of Scribendi Score claim that it can achieve high correlations on the human judgment dataset from Napoles et al. (2016b). However, only moder- ate correlations are observable on SEEDA-Edit. 210
+
+Chunk 1 Chunk 2 Chunk 3 Chunk 4 Chunk 5 Chunk 6 Source Do one who suffered from this disease keep it a secret of infrom their relatives ? Reference Does one who suffers from this disease keep it a secret or inform their relatives ? Hypothesis Do one (0.028) who suffer (0.011) from this disease keep it a secret to inform (0.094) their relatives ? Hit = 0.00, Wrong = 0.79, Under = 0.21, Over = 0.00 Chunk 1 Chunk 2 Chunk 3 Chunk 4 Chunk 5 Chunk 6 Chunk 7 Chunk 8 Chunk 9 Source When we are diagonosed out with certain genetic disease , should we disclose this result to our relatives ? Ref. When we are diagnosed with certain genetic diseases , should we disclose this result to our relatives ? Hyp. When we are diagnosed out (0.056) with certain genetic diseases (0.006) , should we disclose the results (0.019) to their (0.021) relatives ? Hit = 0.10, Wrong = 0.90, Under = 0.0, Over = 0.39 Table 3: Study cases of CLEME2.0 with similarity-based weighting. We highlight TP, FPne, FPun, and FN chunks in different colors. Values in brackets are similarity-based weighting scores. Dataset Corpus-EW Corpus-TS Sentence-EW Sentence-TS γ
+ρ γ
+ρ γ
+ρ γ
+ρ CoNLL-2014 0.697 0.659 0.759 0.720 0.626 0.654 0.696 0.698 BN-10GEC 0.732 0.764 0.796 0.813 0.638 0.637 0.708 0.698 E-Minimal 0.709 0.786 0.779 0.819 0.642 0.692 0.715 0.747 E-Fluency 0.760 0.786 0.831 0.841 0.642 0.665 0.720 0.714 NE-Minimal 0.777 0.823 0.839 0.861 0.654 0.747 0.723 0.791 NE-Fluency 0.823 0.692 0.849 0.709 0.664 0.791 0.742 0.830 Table 4: Correlation results of LLM-based weighting on GJG15 Ranking. 4.4 Results of LLM-based Weighting Table 4 presents the outcomes of LLM-based weighting, noting that its effectiveness is less favor- able than similarity-based weighting. A likely rea- son is the coarse grading method of LLMs, which allocates edit weights from 1 to 5, unlike the finer continuous scale [0, 1]. Although Kobayashi et al. (2024a) argue that LLMs serve as effective evalua- tors for GEC, their research pertains to huge closed- source LLMs (GPT-4 and GPT-3.5) and involves specific prompt engineering. They also identify the importance of the LLM scale since GPT-3.5 may even obtain negative correlations with human judgments. In contrast, we employ a more straight- forward approach with open-source LLama-2-7B. 5
+Analysis 5.1 Case Study Table 3 demonstrates instances of CLEME2.0. In the first set, Chunks 3 and 5 are FPne edits con- tributing to the wrong-correction score, with a higher edit weight of Chunk 5 than Chunk 3 since Chunk 5 introduces an error that entirely alters the sentence’s meaning. In the second set, Chunk 2 ob- tains the highest edit weight of 0.056, underscoring Metric EW TS Avg. γ
+ρ γ
+ρ CLEME2.0-dep-Hit 0.599 0.593 0.673 0.648 0.628 CLEME2.0-dep-Wrong -0.444 -0.533 -0.526 -0.593 -0.524 CLEME2.0-dep-Under 0.496 0.599 0.576 0.659 0.583 CLEME2.0-dep-Over 0.118 0.269 0.073 0.275 0.253 SentCLEME2.0-dep-Hit 0.594 0.593 0.672 0.648 0.627 SentCLEME2.0-dep-Wrong -0.405 -0.429 -0.489 -0.500 -0.456 SentCLEME2.0-dep-Under 0.489 0.511 0.572 0.582 0.539 SentCLEME2.0-dep-Over -0.247 -0.363 -0.346 -0.440 -0.349 Table 5: Correlation results of each disentangled score on GJG15 Ranking. its substantial influence on the evaluation. Despite the correct modification of “diagnosed”, the mis- use of “out” remains, keeping the correction wrong. Chunk 4 illustrates a singular-to-plural correction in the source sentence, with a low weight indicating a minor impact. Chunks 6 and 8 showcase over- corrections. Chunk 6 leaves the original meaning unchanged, whereas Chunk 8 introduces a signifi- cant error by misusing a personal pronoun. The cases highlight the effectiveness of the weighting technique. Otherwise, all edits are given equal weight, failing to distinguish hypothesis ed- its with varying correction levels. We display the cases of LLM-based weighting in Appendix C.2. 5.2 Ablation Study We conduct ablation studies on (Sent)CLEME2.0- dep to analyze the performance of individual dis- entangled scores. A preferable system has re- duced wrong-correction, under-correction, and over-correction scores, so we report corrections between 1-x with human judgments where x is one of the scores. The outcomes are detailed in 211
+
+Metric Time (Seconds) ERRANT 33.4 GLEU 21.5 CLEME-dep 54.1 CLEME-ind 54.1 (Sent)CLEME2.0-dep 54.1 (Sent)CLEME2.0-ind 54.1 (Sent)CLEME2.0-sim-ind 88.4 (Sent)CLEME2.0-sim-ind 87.6 Table 6: Efficiency of metrics. Table 5. Hit-correction and under-correction show moderate correlations. Over-correction scores have small positive correlations at the corpus level, with minimal negative correlations at the sentence level. Notably, wrong-correction scores display negative correlations, but this does not mean they do not im- pact the overall score. In reality, the trade-off factor for wrong-correction scores is relatively substantial. The hypothesis is that focusing evaluations only on wrong-correction scores might prefer systems that make only highly confident edits, potentially lead- ing to assessment bias. Additionally, we utilize the similarity-weighting approach on CLEME to evaluate its efficacy, with the outcomes detailed in Appendix C.3. To exam- ine our method on a broad scale, we also provide the average correlations obtained from a compre- hensive analysis of all potential parameter settings. The results are found in Appendix C.4. 5.3 Efficiency This section provides a comparative analysis of the efficiency of our methods against other prevail- ing metrics. The experiments were executed on a GPU 3090 within the CoNLL-2014 framework, with the evaluation times of the AMU system re- ported. Our observations are as follows: (1) For ERRANT, the primary time expenditure is associ- ated with edit extraction, lasting 33.4 seconds. (2) CLEME and CLEME2.0 primarily incur time costs from edit extraction at 33.4 seconds and chunk par- titioning at 20.7 seconds. (3) For CLEME2.0-sim, the most significant time costs are assignable to edit extraction (33.4 s), chunk partitioning (20.7 s), and edit weighting (34.3 s). PT-M2 exhibits the slow- est runtime when replicating existing mainstream methods, with its evaluation process taking several hours; thus, we did not report a precise runtime due to the time constraints. Some technical solutions can mitigate the runtime when evaluating a sys- tem using these metrics concurrently. For instance, when assessing a system with ERRANT, CLEME, and CLEME2.0, the minimum cumulative duration is calculated as 33.4 seconds for edit extraction, 20.7 seconds for chunk partitioning, and 34.3 sec- onds for edit weighting, totaling 88.4 seconds. 6
+Conclusion This paper introduces CLEME2.0, an interpretable evaluation metric for GEC that effectively high- lights four key aspects of systems. By incorporat- ing edit weighting techniques, we overcome the challenges traditional reference-based metrics face in recognizing semantic subtleties. Extensive ex- periments and analyses confirm the effectiveness and robustness of our method. We anticipate that CLEME2.0 will offer a valuable perspective in the GEC community. Limitation Limitation in languages and datasets. While CLEME2.0 is adaptable to various languages, its efficiency beyond English remains unverified. Ad- ditionally, the reference sets employed in our ex- periments stem from the CoNLL-2014 shared task, which involves a second language dataset. To con- firm the robustness of our methods, it’s necessary to conduct further experiments using evaluation datasets that cover a range of languages and text domains. Finally, we highly encourage the creation of new GEC evaluation datasets to foster progress. Lack of further human evaluation for inter- pretability. The experiments discussed in the paper are primarily concerned with assessing the correlation between automatic metrics and human judgments. However, they fall short of providing a thorough analysis of the method’s interpretabil- ity. Although we showcase the strong correlation performance of CLEME2.0, its interpretability is still unverified. In future work, we will conduct human evaluation experiments to showcase the in- terpretability of our method. Ethics Statement In this paper, we validate the effectiveness and robustness of our proposed approach using the CoNLL-2014, BN-10GEC, and SN-8GEC refer- ence datasets. These datasets are sourced from publicly available resources on legitimate websites 212
+
+and do not contain any sensitive data. Additionally, all the baselines employed in our experiments are publicly accessible GEC metrics, and we have duly cited the respective authors. We confirm that all datasets and baselines utilized in our experiments are consistent with their intended purposes. Acknowledgements This research is supported by National Natu- ral Science Foundation of China (Grant No. 62276154), Research Center for Computer Network (Shenzhen) Ministry of Education, the Natural Science Foundation of Guangdong Province (Grant No. 2023A1515012914 and 440300241033100801770), Ba- sic Research Fund of Shenzhen City (Grant No. JCYJ20210324120012033, JCYJ20240813112009013 and GJHZ20240218113603006), the Major Key Project of PCL (NO. PCL2024A08). References Hiroki Asano, Tomoya Mizumoto, and Kentaro Inui. 2017. Reference-based metrics can be replaced with reference-less metrics in evaluating grammatical er- ror correction systems. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 343– 348. Ondˇrej Bojar, Christian Buck, Chris Callison-Burch, Christian Federmann, Barry Haddow, Philipp Koehn, Christof Monz, Matt Post, Radu Soricut, and Lucia Specia. 2013. Findings of the 2013 Workshop on Statistical Machine Translation. In Proceedings of the Eighth Workshop on Statistical Machine Trans- lation, pages 1–44, Sofia, Bulgaria. Association for Computational Linguistics. Christopher Bryant, Mariano Felice, and Ted Briscoe. 2017. Automatic annotation and evaluation of error types for grammatical error correction. In Proceed- ings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 793–805, Vancouver, Canada. Association for Computational Linguistics. Christopher Bryant and Hwee Tou Ng. 2015. How far are we from fully automatic high quality grammatical error correction? In Proceedings of the 53rd Annual Meeting of the Association for Computational Lin- guistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 697–707, Beijing, China. Association for Computational Linguistics. Christopher Bryant, Zheng Yuan, Muhammad Reza Qorib, Hannan Cao, Hwee Tou Ng, and Ted Briscoe. 2023. Grammatical error correction: A survey of the state of the art. Computational Linguistics, 49(3):643–701. Guiming Hardy Chen, Shunian Chen, Ziche Liu, Feng Jiang, and Benyou Wang. 2024. Humans or llms as the judge? a study on judgement biases. arXiv preprint arXiv:2402.10669. Leshem Choshen and Omri Abend. 2018. Inherent bi- ases in reference-based evaluation for grammatical error correction. In Proceedings of the 56th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 632–642. Zhendong Chu, Shen Wang, Jian Xie, Tinghui Zhu, Yibo Yan, Jingheng Ye, Aoxiao Zhong, Xuming Hu, Jing Liang, Philip S Yu, et al. 2025. Llm agents for education: Advances and applications. arXiv preprint arXiv:2503.11733. Steven Coyne, Keisuke Sakaguchi, Diana Galvan-Sosa, Michael Zock, and Kentaro Inui. 2023. Analyzing the performance of gpt-3.5 and gpt-4 in grammatical error correction. arXiv preprint arXiv:2303.14342. Daniel Dahlmeier and Hwee Tou Ng. 2012a. Better evaluation for grammatical error correction. In Pro- ceedings of the 2012 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 568–572, Montréal, Canada. Association for Compu- tational Linguistics. Daniel Dahlmeier and Hwee Tou Ng. 2012b. Better evaluation for grammatical error correction. In Pro- ceedings of the 2012 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 568–572. Chenhe Dong, Yinghui Li, Haifan Gong, Miaoxin Chen, Junxin Li, Ying Shen, and Min Yang. 2023. A survey of natural language generation. ACM Comput. Surv., 55(8):173:1–173:38. Jiangshu Du, Yibo Wang, Wenting Zhao, Zhongfen Deng, Shuaiqi Liu, Renze Lou, Henry Peng Zou, Pranav Narayanan Venkit, Nan Zhang, Mukund Sri- nath, Haoran Zhang, Vipul Gupta, Yinghui Li, Tao Li, Fei Wang, Qin Liu, Tianlin Liu, Pengzhi Gao, Congy- ing Xia, Chen Xing, Cheng Jiayang, Zhaowei Wang, Ying Su, Raj Sanjay Shah, Ruohao Guo, Jing Gu, Haoran Li, Kangda Wei, Zihao Wang, Lu Cheng, Surangika Ranathunga, Meng Fang, Jie Fu, Fei Liu, Ruihong Huang, Eduardo Blanco, Yixin Cao, Rui Zhang, Philip S. Yu, and Wenpeng Yin. 2024. Llms assist NLP researchers: Critique paper (meta- )reviewing. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, EMNLP 2024, Miami, FL, USA, November 12- 16, 2024, pages 5081–5099. Association for Compu- tational Linguistics. 213
+
+Peiyuan Gong, Xuebo Liu, Heyan Huang, and Min Zhang. 2022. Revisiting grammatical error correc- tion evaluation and beyond. In Proceedings of the 2022 Conference on Empirical Methods in Natu- ral Language Processing, pages 6891–6902, Abu Dhabi, United Arab Emirates. Association for Com- putational Linguistics. Takumi Gotou, Ryo Nagata, Masato Mita, and Kazuaki Hanawa. 2020. Taking the correction difficulty into account in grammatical error correction evaluation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2085–2095. Roman Grundkiewicz, Marcin Junczys-Dowmunt, and Edward Gillian. 2015. Human evaluation of gram- matical error correction systems. In Proceedings of the 2015 Conference on Empirical Methods in Natu- ral Language Processing, pages 461–470. Xinyu Hu, Mingqi Gao, Sen Hu, Yang Zhang, Yicheng Chen, Teng Xu, and Xiaojun Wan. 2024. Are llm-based evaluators confusing nlg quality criteria? arXiv preprint arXiv:2402.12055. Haojing Huang, Jingheng Ye, Qingyu Zhou, Yinghui Li, Yangning Li, Feng Zhou, and Hai-Tao Zheng. 2023. A frustratingly easy plug-and-play detection- and-reasoning module for chinese spelling check. In Findings of the Association for Computational Lin- guistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 11514–11525. Association for Computa- tional Linguistics. Shulin Huang, Shirong Ma, Yinghui Li, Mengzuo Huang, Wuhe Zou, Weidong Zhang, and Haitao Zheng. 2024. Lateval: An interactive llms evalu- ation benchmark with incomplete information from lateral thinking puzzles. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING 2024, 20-25 May, 2024, Torino, Italy, pages 10186–10197. ELRA and ICCL. Md Asadul Islam and Enrico Magnani. 2021. Is this the end of the gold standard? a straightforward reference- less grammatical error correction metric. In Proceed- ings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3009–3015. Marcin Junczys-Dowmunt and Roman Grundkiewicz. 2014. The amu system in the conll-2014 shared task: Grammatical error correction by data-intensive and feature-rich statistical machine translation. In Proceedings of the Eighteenth Conference on Com- putational Natural Language Learning: Shared Task, pages 25–33. Masamune Kobayashi, Masato Mita, and Mamoru Ko- machi. 2024a. Large language models are state-of- the-art evaluator for grammatical error correction. arXiv preprint arXiv:2403.17540. Masamune Kobayashi, Masato Mita, and Mamoru Komachi. 2024b. Revisiting meta-evaluation for grammatical error correction. arXiv preprint arXiv:2403.02674. Shota Koyama, Ryo Nagata, Hiroya Takamura, and Naoaki Okazaki. 2024. n-gram F-score for evaluat- ing grammatical error correction. In Proceedings of the 17th International Natural Language Generation Conference, pages 303–313, Tokyo, Japan. Associa- tion for Computational Linguistics. Jiayi Kuang, Jingyou Xie, Haohao Luo, Ronghao Li, Zhe Xu, Xianfeng Cheng, Yinghui Li, Xika Lin, and Ying Shen. 2024. Natural language understanding and inference with MLLM in visual question answer- ing: A survey. CoRR, abs/2411.17558. Yangning Li, Yinghui Li, Xinyu Wang, Yong Jiang, Zhen Zhang, Xinran Zheng, Hui Wang, Hai-Tao Zheng, Fei Huang, Jingren Zhou, and Philip S. Yu. 2025a. Benchmarking multimodal retrieval aug- mented generation with dynamic VQA dataset and self-adaptive planning agent. In The Thirteenth In- ternational Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenRe- view.net. Yangning Li, Tingwei Lu, Hai-Tao Zheng, Yinghui Li, Shulin Huang, Tianyu Yu, Jun Yuan, and Rui Zhang. 2024a. MESED: A multi-modal entity set expansion dataset with fine-grained semantic classes and hard negative entities. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 8697–8706. AAAI Press. Yangning Li, Qingsong Lv, Tianyu Yu, Yinghui Li, Shulin Huang, Tingwei Lu, Xuming Hu, Wenhao Jiang, Hai-Tao Zheng, and Hui Wang. 2024b. Ul- trawiki: Ultra-fine-grained entity set expansion with negative seed entities. CoRR, abs/2403.04247. Yinghui Li, Haojing Huang, Jiayi Kuang, Yangning Li, Shu-Yu Guo, Chao Qu, Xiaoyu Tan, Hai-Tao Zheng, Ying Shen, and Philip S. Yu. 2025b. Refine knowledge of large language models via adaptive contrastive learning. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net. Yinghui Li, Haojing Huang, Shirong Ma, Yong Jiang, Yangning Li, Feng Zhou, Hai-Tao Zheng, and Qingyu Zhou. 2023a. On the (in)effectiveness of large lan- guage models for chinese text correction. CoRR, abs/2307.09007. Yinghui Li, Shulin Huang, Xinwei Zhang, Qingyu Zhou, Yangning Li, Ruiyang Liu, Yunbo Cao, Hai-Tao Zheng, and Ying Shen. 2023b. Automatic context pattern generation for entity set expansion. IEEE Trans. Knowl. Data Eng., 35(12):12458–12469. Yinghui Li, Yangning Li, Yuxin He, Tianyu Yu, Ying Shen, and Hai-Tao Zheng. 2022a. Contrastive learn- ing with hard negative entities for entity set expan- sion. In SIGIR ’22: The 45th International ACM 214
+
+SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022, pages 1077–1086. ACM. Yinghui Li, Shirong Ma, Shaoshen Chen, Haojing Huang, Shulin Huang, Yangning Li, Hai-Tao Zheng, and Ying Shen. 2025c. Correct like humans: Pro- gressive learning framework for chinese text error correction. Expert Syst. Appl., 265:126039. Yinghui Li, Shirong Ma, Qingyu Zhou, Zhongli Li, Yangning Li, Shulin Huang, Ruiyang Liu, Chao Li, Yunbo Cao, and Haitao Zheng. 2022b. Learning from the dictionary: Heterogeneous knowledge guided fine-tuning for chinese spell checking. In Findings of the Association for Computational Linguistics: EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 238–249. Association for Computational Linguistics. Yinghui Li, Shang Qin, Jingheng Ye, Shirong Ma, Yangning Li, Libo Qin, Xuming Hu, Wenhao Jiang, Hai-Tao Zheng, and Philip S. Yu. 2024c. Rethinking the roles of large language models in chinese gram- matical error correction. CoRR, abs/2402.11420. Yinghui Li, Zishan Xu, Shaoshen Chen, Haojing Huang, Yangning Li, Yong Jiang, Zhongli Li, Qingyu Zhou, Hai-Tao Zheng, and Ying Shen. 2023c. Towards real-world writing assistance: A chinese character checking benchmark with faked and misspelled char- acters. CoRR, abs/2311.11268. Yinghui Li, Zishan Xu, Shaoshen Chen, Haojing Huang, Yangning Li, Shirong Ma, Yong Jiang, Zhongli Li, Qingyu Zhou, Hai-Tao Zheng, and Ying Shen. 2024d. Towards real-world writing assistance: A chinese character checking benchmark with faked and mis- spelled characters. In Proceedings of the 62nd An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 8656– 8668. Association for Computational Linguistics. Yinghui Li, Qingyu Zhou, Yangning Li, Zhongli Li, Ruiyang Liu, Rongyi Sun, Zizhen Wang, Chao Li, Yunbo Cao, and Hai-Tao Zheng. 2022c. The past mis- take is the future wisdom: Error-driven contrastive probability optimization for chinese spell checking. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3202–3213. Association for Computa- tional Linguistics. Yinghui Li, Qingyu Zhou, Yuanzhen Luo, Shirong Ma, Yangning Li, Hai-Tao Zheng, Xuming Hu, and Philip S. Yu. 2024e. When llms meet cunning texts: A fallacy understanding benchmark for large lan- guage models. In Advances in Neural Information Processing Systems 38: Annual Conference on Neu- ral Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024. Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Hai-Tao Zheng. 2022. Are we ready for a new paradigm shift? A survey on visual deep MLP. Pat- terns, 3(7):100520. Shirong Ma, Yinghui Li, Rongyi Sun, Qingyu Zhou, Shulin Huang, Ding Zhang, Yangning Li, Ruiyang Liu, Zhongli Li, Yunbo Cao, Haitao Zheng, and Ying Shen. 2022. Linguistic rules-based corpus generation for native chinese grammatical error correction. In Findings of the Association for Computational Lin- guistics: EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 576–589. As- sociation for Computational Linguistics. Nitin Madnani, Martin Chodorow, Joel Tetreault, and Alla Rozovskaya. 2011. They can help: Using crowd- sourcing to improve the evaluation of grammatical error detection systems. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 508–513. Koki Maeda, Masahiro Kaneko, and Naoaki Okazaki. 2022. Impara: Impact-based metric for gec us- ing parallel data. In Proceedings of the 29th Inter- national Conference on Computational Linguistics, pages 3578–3588. Masato Mita, Tomoya Mizumoto, Masahiro Kaneko, Ryo Nagata, and Kentaro Inui. 2019. Cross-corpora evaluation and analysis of grammatical error correc- tion models—is single-corpus evaluation enough? In Proceedings of NAACL-HLT, pages 1309–1314. Courtney Napoles, Keisuke Sakaguchi, Matt Post, and Joel Tetreault. 2015. Ground truth for grammatical error correction metrics. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Confer- ence on Natural Language Processing (Volume 2: Short Papers), pages 588–593. Courtney Napoles, Keisuke Sakaguchi, and Joel Tetreault. 2016a. There’s no comparison: Reference- less evaluation metrics in grammatical error correc- tion. arXiv preprint arXiv:1610.02124. Courtney Napoles, Keisuke Sakaguchi, and Joel Tetreault. 2016b. There’s no comparison: Reference- less evaluation metrics in grammatical error correc- tion. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Process- ing, pages 2109–2115, Austin, Texas. Association for Computational Linguistics. Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christo- pher Bryant. 2014. The conll-2014 shared task on grammatical error correction. In Proceedings of the eighteenth conference on computational natural lan- guage learning: shared task, pages 1–14. Maja Pavlovic and Massimo Poesio. 2024. The ef- fectiveness of llms as annotators: A comparative overview and empirical analysis of direct representa- tion. arXiv preprint arXiv:2405.01299. 215
+
+Libo Qin, Qiguang Chen, Xiachong Feng, Yang Wu, Yongheng Zhang, Yinghui Li, Min Li, Wanxiang Che, and Philip S Yu. 2024a. Large language models meet nlp: A survey. arXiv preprint arXiv:2405.12819. Libo Qin, Qiguang Chen, Yuhang Zhou, Zhi Chen, Yinghui Li, Lizi Liao, Min Li, Wanxiang Che, and Philip S. Yu. 2024b. Multilingual large language model: A survey of resources, taxonomy and fron- tiers. CoRR, abs/2404.04925. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, and Dario Amodei. 2019. Gpt 2; language models are unsupervised multitask learners. In 2019 by OpenAI. Alla Rozovskaya, Kai-Wei Chang, Mark Sammons, Dan Roth, and Nizar Habash. 2014. The illinois-columbia system in the conll-2014 shared task. In Proceedings of the Eighteenth Conference on Computational Nat- ural Language Learning: Shared Task, pages 34–42. Keisuke Sakaguchi, Courtney Napoles, Matt Post, and Joel Tetreault. 2016. Reassessing the goals of gram- matical error correction: Fluency instead of grammat- icality. Transactions of the Association for Computa- tional Linguistics, 4:169–182. Keisuke Sakaguchi, Matt Post, and Benjamin Van Durme. 2014. Efficient elicitation of annota- tions for human evaluation of machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 1–11. Andrea Sottana, Bin Liang, Kai Zou, and Zheng Yuan. 2023. Evaluation metrics in the era of gpt-4: reli- ably evaluating large language models on sequence to sequence tasks. arXiv preprint arXiv:2310.13800. Jiamin Su, Yibo Yan, Fangteng Fu, Han Zhang, Jingheng Ye, Xiang Liu, Jiahao Huo, Huiyu Zhou, and Xuming Hu. 2025. Essayjudge: A multi-granular benchmark for assessing automated essay scoring capabilities of multimodal large language models. arXiv preprint arXiv:2502.11916. Zhen Tan, Alimohammad Beigi, Song Wang, Ruocheng Guo, Amrita Bhattacharjee, Bohan Jiang, Mansooreh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024. Large language models for data annotation: A survey. arXiv preprint arXiv:2402.13446. Jiwei Tang, Zhicheng Zhang, Shunlong Wu, Jingheng Ye, Lichen Bai, Zitai Wang, Tingwei Lu, Jiaqi Chen, Lin Hai, Hai-Tao Zheng, et al. 2025. Gmsa: Enhancing context compression via group merg- ing and layer semantic alignment. arXiv preprint arXiv:2505.12215. Hugo Touvron, Louis Martin, Kevin Stone, Peter Al- bert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open founda- tion and fine-tuned chat models. arXiv preprint arXiv:2307.09288. Zhikun Xu, Yinghui Li, Ruixue Ding, Xinyu Wang, Boli Chen, Yong Jiang, Haitao Zheng, Wenlian Lu, Pengjun Xie, and Fei Huang. 2025. Let llms take on the latest challenges! A chinese dynamic ques- tion answering benchmark. In Proceedings of the 31st International Conference on Computational Lin- guistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025, pages 10435–10448. Association for Computational Linguistics. Yibo Yan, Shen Wang, Jiahao Huo, Jingheng Ye, Zhen- dong Chu, Xuming Hu, Philip S Yu, Carla Gomes, Bart Selman, and Qingsong Wen. 2025. Posi- tion: Multimodal large language models can signifi- cantly advance scientific reasoning. arXiv preprint arXiv:2502.02871. Jingheng Ye, Yong Jiang, Xiaobin Wang, Yinghui Li, Yangning Li, Hai-Tao Zheng, Pengjun Xie, and Fei Huang. 2024. Productagent: Benchmarking conver- sational product search agent with asking clarification questions. arXiv preprint arXiv:2407.00942. Jingheng Ye, Yinghui Li, Yangning Li, and Hai-Tao Zheng. 2023a. Mixedit: Revisiting data augmenta- tion and beyond for grammatical error correction. In Findings of the Association for Computational Lin- guistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 10161–10175. Association for Computa- tional Linguistics. Jingheng Ye, Yinghui Li, Shirong Ma, Rui Xie, Wei Wu, and Hai-Tao Zheng. 2022. Focus is what you need for chinese grammatical error correction. arXiv preprint arXiv:2210.12692. Jingheng Ye, Yinghui Li, and Haitao Zheng. 2023b. System report for ccl23-eval task 7: Thu kelab (sz)- exploring data augmentation and denoising for chi- nese grammatical error correction. In Proceedings of the 22nd Chinese National Conference on Compu- tational Linguistics (Volume 3: Evaluations), pages 262–270. Jingheng Ye, Yinghui Li, Qingyu Zhou, Yangning Li, Shirong Ma, Hai-Tao Zheng, and Ying Shen. 2023c. CLEME: Debiasing multi-reference evaluation for grammatical error correction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6174–6189, Singapore. Association for Computational Linguistics. Jingheng Ye, Shang Qin, Yinghui Li, Xuxin Cheng, Libo Qin, Hai-Tao Zheng, Ying Shen, Peng Xing, Zis- han Xu, Guo Cheng, et al. 2025a. Excgec: A bench- mark for edit-wise explainable chinese grammatical error correction. In Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 39, pages 25678–25686. Jingheng Ye, Shang Qin, Yinghui Li, Hai-Tao Zheng, Shen Wang, and Qingsong Wen. 2025b. Corrections meet explanations: A unified framework for explain- able grammatical error correction. arXiv preprint arXiv:2502.15261. 216
+
+Jingheng Ye, Shen Wang, Deqing Zou, Yibo Yan, Kun Wang, Hai-Tao Zheng, Zenglin Xu, Irwin King, Philip S Yu, and Qingsong Wen. 2025c. Position: Llms can be good tutors in foreign language educa- tion. arXiv preprint arXiv:2502.05467. Ryoma Yoshimura, Masahiro Kaneko, Tomoyuki Ka- jiwara, and Mamoru Komachi. 2020. Some: Reference-less sub-metrics optimized for manual evaluations of grammatical error correction. In Pro- ceedings of the 28th International Conference on Computational Linguistics, pages 6516–6522. Miao Yu, Junyuan Mao, Guibin Zhang, Jingheng Ye, Junfeng Fang, Aoxiao Zhong, Yang Liu, Yux- uan Liang, Kun Wang, and Qingsong Wen. 2024a. Mind scramble: Unveiling large language model psychology via typoglycemia. arXiv preprint arXiv:2410.01677. Tianyu Yu, Chengyue Jiang, Chao Lou, Shen Huang, Xiaobin Wang, Wei Liu, Jiong Cai, Yangning Li, Yinghui Li, Kewei Tu, Hai-Tao Zheng, Ningyu Zhang, Pengjun Xie, Fei Huang, and Yong Jiang. 2024b. Seqgpt: An out-of-the-box large language model for open domain sequence understanding. In Thirty-Eighth AAAI Conference on Artificial Intelli- gence, AAAI 2024, Thirty-Sixth Conference on Inno- vative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Ad- vances in Artificial Intelligence, EAAI 2014, Febru- ary 20-27, 2024, Vancouver, Canada, pages 19458– 19467. AAAI Press. Ding Zhang, Yangning Li, Lichen Bai, Hao Zhang, Yinghui Li, Haiye Lin, Hai-Tao Zheng, Xin Su, and Zifei Shan. 2025a. Loss-aware curriculum learn- ing for chinese grammatical error correction. CoRR, abs/2501.00334. Ding Zhang, Yinghui Li, Qingyu Zhou, Shirong Ma, Yangning Li, Yunbo Cao, and Hai-Tao Zheng. 2023. Contextual similarity is more valuable than charac- ter similarity: An empirical study for chinese spell checking. In IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023, pages 1–5. IEEE. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Eval- uating text generation with bert. arXiv preprint arXiv:1904.09675. Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, Feiran Huang, Sheng Zhou, Jiajun Bu, Allen Lin, James Caverlee, Fakhri Karray, Irwin King, and Philip S. Yu. 2025b. Cold-start recommendation towards the era of large language models (llms): A comprehensive survey and roadmap. CoRR, abs/2501.01945. Deqing Zou, Jingheng Ye, Yulu Liu, Yu Wu, Zishan Xu, Yinghui Li, Hai-Tao Zheng, Bingxu An, Zhao Wei, and Yong Xu. 2025. Revisiting classification taxonomy for grammatical errors. arXiv preprint arXiv:2502.11890. A
+LLM-based Edit Weighting Because of the powerful semantic comprehension abilities of LLMs (Qin et al., 2024a; Tan et al., 2024; Ye et al., 2025a; Yu et al., 2024a; Tang et al., 2025; Yan et al., 2025; Li et al., 2024e, 2022a; Du et al., 2024; Huang et al., 2024; Li et al., 2023b; Yu et al., 2024b; Li et al., 2025a; Kuang et al., 2024), recent studies (Chu et al., 2025; Ye et al., 2025c,b, 2024; Hu et al., 2024; Chen et al., 2024; Su et al., 2025; Zou et al., 2025; Xu et al., 2025; Li et al., 2025b, 2024b; Zhang et al., 2025b; Li et al., 2024a) have generated interest in employing LLMs for text assessment on various NLP tasks. Building on this idea, we use Llama-2-7B (Touvron et al., 2023) as a scorer to determine edit weights. The prompt for edit weighting is presented in Figure 3. We set the temperature to 0.1 to ensure consistent and certain results. We instruct the LLM to evaluate each edit individually to prevent interference from other grammatical errors. Edit weights vary from 1 to 5, with higher values representing a greater need for correction. We do not specify the types of edits to the LLM; instead, we allow the LLM to directly evaluate the importance of edits through its inherent language understanding abilities. An input is composed of an uncorrected sentence and a certain edit. B
+Details about GEC Meta-Evaluation B.1 Human Rankings GJG15 ranking. Grundkiewicz et al. (2015) pro- pose the first large-scale human judgement dataset for 12 participating systems of the CoNLL-2014 shared task. In this assessment, 8 native speakers are asked to rank the systems’ outputs from best to worst. Two system ranking lists are generated using Expected Wins (EW) and TrueSkill (TS), re- spectively. SEEDA ranking. Kobayashi et al. (2024b) iden- tify several limitations of the GJG15 ranking dataset, and propose a new human ranking dataset called SEEDA. SEEDA consists of corrections with human ratings along two different granularities: edit-based and sentence-based, covering 12 state- of-the-art systems, including large language mod- els (LLMs), and two human corrections with differ- 217
+
+Prompt: As an evaluator for grammatical error correction, you are tasked with assessing the importance of each error. You will be provided with two lines: the first is an uncorrected sentence, the second shows the edit. Then you output the importance score of the given edit. The scores range from 1 to 5, where a higher score reflects the greater significance of the correction, while a lower score indicates minor importance. - A score of 1 means the correction is almost negligible and unnecessary. - A score of 2 means the correction has slight influence. - A score of 3 signifies some impact by the correction. - A score of 4 means the edit is essential. - A score of 5 indicates the modification is highly important and necessary. Next, I’ll provide you a sentence with an edit. You should score each edit accordingly. The output should only be the score, with no additional explanation. Example Input: Uncorrected sentence: Nowadays the technologies were improved a lot compared to the last century. Edit: were →have Example Output (1-5): 5 Note that the output must be a number between 1 and 5. Here is the formal input: Uncorrected sentence: {uncorrected sentence} Edit: {edit} Example Output (1-5): Figure 3: Prompt of LLM-based weighting. ent focuses. Three native English speakers partic- ipate in the annotation process. Similar to Grund- kiewicz et al. (2015), the overall human rankings are derived from TrueSkill (TS) and Expected Wins (EW) based on pairwise judgments. B.2 Ranking Algorithms Our employed human judgments are originally pair- wise comparisons, i.e., humans choose the better of two available system outputs. The overall rankings are derived by using ranking algorithms, including Expected Wins (EW) and TrueSkill (TS). Expected Wins (EW) EW (Bojar et al., 2013) is a derived ranking metric that quantifies the the- oretical number of wins a participant is expected to achieve against a defined set of opponents. It is calculated by summing the probability of winning against each opponent, where these probabilities are typically derived from an existing skill rating system. EW provides a single aggregate score for ranking, useful for pre-match seeding or assessing theoretical group performance. TrueSkill (TS) TS (Sakaguchi et al., 2014) is a Bayesian skill rating system developed by Mi- crosoft Research. Unlike simpler systems, TS mod- els a participant’s skill as a probability distribution (N(µ, σ2)), where µ represents the estimated skill level and σ quantifies the uncertainty in that es- timate. Upon match outcomes, TS updates these distributions using Bayesian inference, allowing for rapid adjustments and robust ranking. A key advantage is its inherent support for multi-player or team-based matches and the explicit handling of draws. Participants are typically ranked by a conservative estimate of their skill, such as µ −3σ, which accounts for confidence. B.3 Statistics of Reference Datasets Table 7 presents the statistics of all the reference sets involved in our experiments. B.4 Baseline Metrics In our evaluation, we compare our method with the following reference-based baseline metrics, includ- ing corpus and sentence-level variants: 218
+
+Item CoNLL-2014 BN-10GEC E-Minimal E-Fluency NE-Minimal NE-Fluency # Sentence (Length) 1,312 (23.0) 1,312 (23.0) 1,312 (23.0) 1,312 (23.0) 1,312 (23.0) 1,312 (23.0) # Reference (Length) 2,624 (22.8) 13,120 (22.9) 2,624 (23.2) 2,624 (22.8) 2,624 (23.0) 2,624 (22.2) # Edit (Length) 5,937 (1.0) 36,677 (1.0) 4,500 (1.0) 8,373 (1.1) 4,964 (0.9) 11,033 (1.2) # Unchanged Chunk (Length) 11,174 (4.8) 93,496 (2.5) 8,887 (6.3) 12,823 (3.8) 10,748 (5.1) 14,086 (2.9) # Corrected/Dummy Chunk (Length) 4,994 (1.3) 26,948 (2.4) 3,963 (1.2) 6,305 (1.7) 4,221 (1.2) 6,892 (2.6) Table 7: Statistics of CoNLL-2014 (Ng et al., 2014), BN-10GEC (Bryant and Ng, 2015) and SN-8GEC (Sakaguchi et al., 2016) reference sets. We leverage ERRANT (Bryant et al., 2017) for edit extraction, and CLEME (Ye et al., 2023c) for chunk extraction. Metric CoNLL-2014 BN-10GEC E-Minimal E-Fluency NE-Minimal NE-Fluency Avg. EW TS EW TS EW TS EW TS EW TS EW TS SentGLEU γ 0.784 0.828 0.756 0.826 0.624 0.581 0.785 0.846 0.218 0.142 0.778 0.848 0.668 (⇓0.120) ρ 0.720 0.775 0.769 0.824 0.599 0.593 0.791 0.846 0.220 0.170 0.768 0.846 0.660 (⇓0.131) SentERRANT γ 0.870 0.846 0.885 0.896 0.760 0.692 0.806 0.732 0.104 -0.066 0.793 0.847 0.680 (⇓0.130) ρ 0.742 0.747 0.786 0.830 0.626 0.588 0.813 0.764 -0.003 -0.137 0.830 0.857 0.620 (⇓0.179) SentCLEME-dep γ 0.876 0.844 0.915 0.913 0.602 0.507 0.849 0.886 -0.021 -0.127 0.876 0.921 0.670 (⇓0.185) ρ 0.824 0.808 0.835 0.874 0.451 0.412 0.824 0.863 -0.181 -0.247 0.791 0.846 0.592 (⇓0.233) SentCLEME-ind γ 0.868 0.857 0.539 0.453 0.513 0.410 0.841 0.877 -0.061 -0.181 0.852 0.896 0.572 (⇓0.279) ρ 0.725 0.758 0.209 0.143 0.368 0.335 0.808 0.846 -0.167 -0.247 0.762 0.825 0.447 (⇓0.335) SentCLEME2.0-dep (Ours) γ 0.870 0.881 0.766 0.830 0.937 0.928 0.892 0.938 0.634 0.571 0.916 0.949 0.843 (⇓0.054) ρ 0.714 0.725 0.681 0.747 0.846 0.852 0.824 0.901 0.368 0.352 0.720 0.791 0.710 (⇓0.091) SentCLEME2.0-ind (Ours) γ 0.866 0.881 0.799 0.853 0.940 0.933 0.915 0.952 0.693 0.631 0.883 0.904 0.854 (⇓0.045) ρ 0.709 0.720 0.681 0.747 0.819 0.835 0.857 0.923 0.423 0.401 0.654 0.720 0.707 (⇓0.086) SentCLEME2.0-sim-dep (Ours) γ 0.926 0.937 0.797 0.861 0.914 0.902 0.908 0.952 0.607 0.550 0.918 0.947 0.852 (⇓0.054) ρ 0.907 0.912 0.808 0.863 0.808 0.813 0.885 0.945 0.527 0.505 0.896 0.940 0.817 (⇓0.051) SentCLEME2.0-sim-ind (Ours) γ 0.915 0.936 0.808 0.866 0.922 0.916 0.923 0.963 0.720 0.669 0.931 0.961 0.877 (⇓0.038) ρ 0.868 0.879 0.753 0.824 0.808 0.841 0.879 0.956 0.544 0.527 0.835 0.923 0.803 (⇓0.052) Table 8: Correlation results on GJG15 Ranking. We report the results without excluding unchanged reference sentences and the reduction compared with Table 1. We highlight the highest scores in bold and the second-highest scores with underlines. • M2 and SentM2 (Dahlmeier and Ng, 2012a) dy- namically extract the hypothesis edits with the maximum overlap of gold annotations by utiliz- ing the Levenshtein algorithm. • GLEU and SentGLEU (Napoles et al., 2015) are BLEU-like GEC metrics based on n-gram match- ing, rewarding hypothesis n-grams that align with the reference but not the source, while penaliz- ing those aligning solely with the source. GLEU is the main metric in JFLEG, an English GEC dataset that highlights holistic fluency edits. • ERRANT and SentERRANT (Bryant et al., 2017) are among the most widely recognized in grammatical error correction. They enhance the accuracy of edit extraction by employing a linguistically refined version of the Damerau- Levenshtein algorithm. • PT-M2 and SentPT-M2 (Gong et al., 2022) lever- age pre-trained language model (PLM) to evalu- ate GEC systems. The main idea is similar to M2 and ERRANT, but they can leverage the knowl- edge of pre-trained language models to score edits effectively. • CLEME and SentCLEME (Ye et al., 2023c) are proposed to provide unbiased scores for multi- reference evaluation. Furthermore, the authors present the correction independence assumption, enabling CLEME to function under either the traditional correction dependence or correction independence assumptions. For the evaluation on SEEDA, we add extra eval- uation metrics following the evaluation methods reported in Kobayashi et al. (2024b): • GoToScorer (Gotou et al., 2020): takes into ac- count the difficulty of error correction when cal- culating the evaluation score. The difficulty is 219
+
+Chunk 1 Chunk 2 Chunk 3 Chunk 4 Chunk 5 Chunk 6 Source Do one who suffered from this disease keep it a secret of infrom their relatives ? Reference Does one who suffers from this disease keep it a secret or inform their relatives ? Hypothesis Do one (5) who suffer (5) from this disease keep it a secret to inform (1) their relatives ? Hit = 0.00, Wrong = 0.55, Under = 0.45, Over = 0.00 Chunk 1 Chunk 2 Chunk 3 Chunk 4 Chunk 5 Chunk 6 Chunk 7 Chunk 8 Chunk 9 Source When we are diagonosed out with certain genetic disease , should we disclose this result to our relatives ? Ref. When we are diagnosed with certain genetic diseases , should we disclose this result to our relatives ? Hyp. When we are diagnosed out (5) with certain genetic diseases (5) , should we disclose the results (4) to their (5) relatives ? Hit = 0.50, Wrong = 0.50, Under = 0.00, Over = 0.47 Table 9: Study cases of CLEME2.0 with LLM-based weighting. We highlight TP, FPne, FPun, and FN chunks in different colors. Values in brackets are LLM-based weighting scores. calculated based on the number of systems that can correct errors. • Scribendi Score (Islam and Magnani, 2021): evaluates GEC systems in conjunction with the complexity calculated by GPT-2 (Radford et al., 2019), the labeled ranking ratio and the Leven- stein distance ratio. • SOME (Yoshimura et al., 2020): optimizes hu- man evaluation by fine-tuning BERT separately for criteria such as grammaticality, fluency, and meaning preservation. • IMPARA (Maeda et al., 2022): incorporates a quality assessment model fine-tuned using BERT parallel data and a similarity model that takes into account the effects of editing. B.5 Details of Determining Trade-off Factors A cross-validation approach was employed on the six reference sets of GJG15 to determine the opti- mal set. Five of the six reference sets were selected, and an exhaustive exploration of all trade-off fac- tors was conducted. The candidate factors were evaluated at intervals determined by a grid value of 0.05. The optimal factors were then identified and applied to the remaining reference set, yielding resultant corrections. We reiterated this process six times to ascertain the final set of trade-off factors, which exhibited the highest average correction for the remaining reference sets. C
+Extra Results C.1 Results of Full References The results without excluding unchanged reference sentences are presented in Table 8. We observe an obvious performance reduction in traditional met- rics, especially in NE-Minimal, which contains nu- merous under-corrections due to annotation by non- experts under the minimal editing guideline. We remove 470 unchanged references in E-Minimal and 612 unchanged references in NE-Minimal. In particular, SentERRANT, SentCLEME-dep, and SentCLEME-ind exhibit negative correlations in NE-Minimal, revealing their lack of robustness. Many metrics also undergo a significant decrease in E-Minimal except CLEME2.0. In the case of E- Minimal, many metrics also show a marked decline, except for CLEME2.0. Our approach achieves the highest or comparable correlations across all refer- ence sets, underscoring its robustness. C.2 Case Study of LLM-based Weighting In Table 9, we report instances of CLEME2.0 using LLM-based weighting. We notice distinct prefer- ences when comparing similarity-based and LLM- based weighting methods. In the first example, Llama-2 attributes significant weights to Chunks 1 and 2, highlighting key grammatical mistakes. Con- versely, it assigns a minor weight to Chunk 5 due to its imperfect modification. The second example shows Llama-2 attributing substantial weights to all chunks. Specifically, for Chunk 2, the hypothesis fails to remove the redundant “out," emphasizing the under-correction issue. Chunks 6 and 8 display excessive corrections, altering the sentence’s orig- inal intent and thus indicating considerable over- correction. Generally, Llama-2 tends to ascribe either very high or very low weights to modifica- tions. We speculate it is due to the small scale of the LLM we adopt, impairing its ability to distinguish grammatical errors with varying levels. 220
+
+Metric EW TS γ
+ρ γ
+ρ CLEME-dep-unw 0.638 0.654 0.681 0.709 CLEME-ind-unw 0.640 0.648 0.680 0.698 CLEME-dep-len 0.700 0.665 0.691 0.742 CLEME-ind-len 0.649 0.709 0.691 0.731 CLEME-dep-sim 0.655 0.764 0.698 0.797 CLEME-ind-sim 0.641 0.720 0.687 0.747 SentCLEME-dep-unw 0.853 0.687 0.805 0.604 SentCLEME-ind-unw 0.790 0.275 0.722 0.181 SentCLEME-dep-len 0.876 0.824 0.844 0.808 SentCLEME-ind-len 0.868 0.725 0.857 0.758 SentCLEME-dep-sim 0.888 0.692 0.844 0.648 SentCLEME-ind-sim 0.843 0.500 0.786 0.434 Table 10: Extra results of CLEME with different edit weighting techniques: unweighting (unw), length-based weighting (len), and similarity-based weighting (sim). C.3 Extra Results of CLEME with Similarity Weighting We additionally investigate the application of similarity-based weighting to CLEME (Ye et al., 2023c) and present the results on CoNLL2014 in Table 10. We find that similarity-based weighting is superior to length-based weighting for corpus- level CLEME, while the trend is reversed for Sent- CLEME, and both are better than the unweighted setting. Moreover, it should be noted that no mat- ter the weighting strategy employed, CLEME con- sistently underperforms compared to CLEME2.0. This is attributed to the fundamental disparities in design and scoring frameworks between the ver- sions. CLEME2.0 was crafted to incorporate these sophisticated weighting techniques, allowing it to better distinguish between diverse error types and deliver a more thorough and refined performance assessment. C.4 Average Correlations. To analyze our method from a global viewpoint, we present the average correlations derived from the exhaustive enumeration of possible parameter configurations. We explore all potential parameter combinations with increments of 0.05. Table 11 shows that all correlations are positive, regardless of the correction assumptions, levels of evalua- tion, or weighting techniques used. By compar- ing results from unweighted and similarity-based weighted metrics, we determine that similarity- Metric EW TS Avg. γ
+ρ γ
+ρ CLEME2.0-dep 0.461 0.423 0.483 0.457 0.456 CLEME2.0-ind 0.468 0.421 0.489 0.453 0.458 CLEME2.0-sim-dep 0.559 0.592 0.581 0.624 0.589 CLEME2.0-sim-ind 0.566 0.593 0.588 0.622 0.592 SentCLEME2.0-dep 0.374 0.305 0.362 0.290 0.333 SentCLEME2.0-ind 0.372 0.302 0.356 0.283 0.328 SentCLEME2.0-sim-dep 0.410 0.361 0.400 0.345 0.379 SentCLEME2.0-sim-ind 0.412 0.360 0.399 0.338 0.377 Table 11: Average correlations of (Sent)CLEME2.0 and (Sent)CLEME2.0-sim on CoNLL-2014. based weighting substantially enhances human cor- relation on a global level. Additionally, corpus- level metrics generally achieve higher average val- ues compared to sentence-level metrics. However, sentence-level metrics with optimal parameters can outperform their corpus-level equivalents. This im- plies that corpus-level metrics might demonstrate greater robustness concerning parameter selection. C.5 Details Results on CoNLL-2014 Table 12 presents a comprehensive evaluation of CLEME2.0 on CoNLL-2014 across all GEC sys- tems. Our method offers a clear and quantita- tive examination of detailed features of GEC sys- tems, which other automatic metrics cannot pro- vide. For instance, the CAMB system attains the top hit-correction score of 0.271 for CLEME2.0- dep, which shows that about 27.1% of edits by the system are accurate. The wrong-correction score of 0.194 indicates that 19.4% of edits are correctly placed but incorrect, the under-correction score of 0.534 indicates that 53.4% of grammatical errors are overlooked by the system, and the over- correction score of 0.470 suggests that 47.0% of the edits are unnecessary. As a result, developers and researchers can pin- point the aspects of their systems that require en- hancement. Furthermore, users can select GEC systems that best meet their requirements. For in- stance, users might opt for a system with a minimal under-correction score in high-stakes situations, as they expect to detect every possible grammatical mistake even though the system might make some unnecessary edits. 221
+
+Metric AMU CAMB CUUI IITB INPUT IPN NTHU PKU POST RAC SJTU UFC UMC CLEME2.0-dep TP 380 584 471 22 0
+39 330 246 412 254 85 32 260 sim 9.20 12.66 7.58 0.39 0.00 0.77 5.79 6.69 8.80 6.68 1.50 0.42 5.29 FP 817 1307 964 67 0
+488 905 709 1145 782 272 18 789 sim 16.03 30.92 16.06 1.80 0.00 11.93 24.56 14.36 19.25 11.98 6.49 0.25 18.26 FPne 276 418 311 34 0
+149 302 254 316 259 76 12 245 sim 4.08 6.55 3.68 0.75 0.00 4.61 5.89 4.06 4.60 3.80 2.30 0.17 3.83 FPun 541 889 653 33 0
+339 603 455 829 523 196 6
+544 sim 11.95 24.36 12.38 1.05 0.00 7.33 18.67 10.30 14.64 8.18 4.19 0.08 14.43 FN 1360 1150 1357 2057 1782 2886 1388 1454 1354 1487 1668 2087 1461 sim 34.25 28.45 36.21 78.39 48.24 83.10 46.53 36.15 34.48 38.00 56.28 51.27 39.60 Hit 0.188 0.271 0.220 0.010 0.00 0.013 0.163 0.126 0.198 0.127 0.046 0.015 0.132 sim 0.194 0.266 0.160 0.005 0.00 0.009 0.100 0.143 0.184 0.138 0.025 0.008 0.109 Wrong 0.137 0.194 0.145 0.016 0.00 0.048 0.150 0.130 0.152 0.130 0.042 0.006 0.125 sim 0.086 0.138 0.078 0.009 0.00 0.052 0.101 0.0866 0.096 0.078 0.038 0.003 0.079 Under 0.675 0.534 0.634 0.973 1.00 0.939 0.687 0.744 0.650 0.744 0.912 0.979 0.743 sim 0.721 0.597 0.763 0.986 1.00 0.939 0.799 0.771 0.720 0.784 0.937 0.989 0.813 Over 0.452 0.470 0.455 0.371 0.00 0.643 0.488 0.476 0.532 0.505 0.549 0.12 0.519 sim 0.474 0.559 0.524 0.478 0.00 0.577 0.615 0.490 0.522 0.438 0.524 0.116 0.613 Score 0.483 0.508 0.497 0.431 0.45 0.408 0.463 0.450 0.479 0.505 0.434 0.450 0.453 sim 0.503 0.520 0.484 0.425 0.45 0.408 0.439 0.474 0.491 0.438 0.424 0.448 0.452 SentCLEME2.0-dep TP 376 580 467 22 0
+39 327 244 409 251 84 32 259 sim 9.14 12.63 7.52 0.39 0.00 0.76 5.72 6.65 8.75 6.59 1.48 0.42 5.23 FP 821 1311 968 67 0
+488 908 711 1148 785 273 18 790 sim 16.49 31.25 16.50 1.85 0.00 13.00 24.83 14.38 19.36 12.34 7.13 0.26 18.47 FPne 286 431 320 22 0
+132 310 262 326 271 81 10 255 sim 4.60 7.51 4.27 0.44 0.00 2.62 6.58 4.58 5.06 4.02 1.28 0.15 4.39 FPun 535 880 648 45 0
+356 598 449 822 514 192 8
+535 sim 11.89 23.74 12.23 1.42 0.00 10.39 18.24 9.80 14.30 8.32 5.85 0.12 14.07 FN 1600 1374 1577 1972 1982 1940 1660 1712 1587 1744 1900 1980 1714 sim 43.65 35.92 45.22 57.46 58.31 54.69 46.92 46.02 43.09 46.05 55.32 58.35 48.02 Hit 0.136 0.210 0.163 0.008 0.00 0.013 0.119 0.088 0.142 0.089 0.032 0.012 0.091 sim 0.131 0.205 0.142 0.007 0.00 0.011 0.104 0.088 0.129 0.086 0.027 0.008 0.087 Wrong 0.080 0.129 0.090 0.005 0.00 0.038 0.095 0.076 0.088 0.071 0.023 0.002 0.070 sim 0.063 0.102 0.066 0.004 0.00 0.033 0.079 0.059 0.070 0.051 0.020 0.001 0.059 Under 0.500 0.392 0.479 0.675 0.687 0.639 0.496 0.538 0.486 0.551 0.637 0.678 0.546 sim 0.519 0.419 0.517 0.673 0.684 0.645 0.524 0.553 0.509 0.567 0.641 0.680 0.557 Over 0.248 0.419 0.293 0.031 0.00 0.242 0.304 0.235 0.342 0.232 0.121 0.006 0.267 sim 0.241 0.421 0.294 0.030 0.00 0.224 0.302 0.224 0.331 0.203 0.119 0.005 0.267 Score 0.498 0.513 0.507 0.467 0.466 0.447 0.481 0.475 0.495 0.477 0.469 0.471 0.476 sim 0.502 0.520 0.504 0.467 0.466 0.449 0.479 0.481 0.494 0.484 0.467 0.469 0.479 CLEME2.0-ind TP 388 596 487 22 0
+39 338 248 420 255 85 32 262 sim 9.47 13.11 7.99 0.40 0.00 0.81 6.13 6.80 9.07 6.91 1.54 0.47 5.49 FP 809 1295 948 67 0
+488 897 707 1137 781 272 18 787 sim 14.74 28.11 14.42 1.91 0.00 11.82 22.93 13.03 17.62 11.23 6.46 0.25 16.99 FPne 408 627 449 34 0
+234 447 388 487 406 134 12 366 sim 6.32 10.62 5.51 0.86 0.00 4.79 9.50 7.30 7.12 5.56 2.41 0.17 6.14 FPun 401 668 499 33 0
+254 450 319 650 375 138 6
+421 sim 8.42 17.49 8.91 1.05 0.00 7.03 13.43 5.73 10.50 5.67 4.05 0.08 10.85 FN 1029 778 984 1497 1530 1382 1045 1129 989 1135 1398 1506 1136 sim 26.88 20.31 27.94 53.23 41.31 50.21 36.83 28.40 26.59 29.30 40.63 41.49 31.88 Hit 0.213 0.298 0.254 0.014 0.000 0.024 0.185 0.141 0.222 0.142 0.053 0.021 0.149 sim 0.222 0.298 0.193 0.007 0.000 0.015 0.117 0.160 0.212 0.165 0.035 0.011 0.126 Wrong 0.224 0.313 0.234 0.022 0.000 0.141 0.244 0.220 0.257 0.226 0.083 0.008 0.207 sim 0.148 0.241 0.133 0.016 0.000 0.086 0.181 0.172 0.166 0.133 0.054 0.004 0.141 Under 0.564 0.389 0.513 0.964 1.000 0.835 0.571 0.640 0.522 0.632 0.865 0.972 0.644 sim 0.630 0.461 0.674 0.977 1.000 0.900 0.702 0.668 0.622 0.701 0.911 0.985 0.733 Over 0.335 0.353 0.348 0.371 0.000 0.482 0.364 0.334 0.417 0.362 0.387 0.12 0.401 sim 0.348 0.424 0.397 0.454 0.000 0.557 0.462 0.289 0.393 0.313 0.506 0.11 0.483 Score 0.472 0.486 0.490 0.432 0.450 0.389 0.448 0.434 0.461 0.431 0.431 0.453 0.439 sim 0.503 0.508 0.490 0.426 0.450 0.400 0.428 0.463 0.489 0.479 0.425 0.449 0.446 SentCLEME2.0-ind TP-sim 9.16 12.59 7.73 0.40 0.00 0.75 5.93 6.67 8.77 6.67 1.50 0.47 5.21 FP-sim 15.83 29.93 15.62 1.76 0.00 12.58 24.30 14.17 18.94 12.00 6.84 0.27 17.76 FPne-sim 7.20 12.38 6.58 0.70 0.00 5.27 10.94 8.38 8.37 6.25 2.70 0.19 6.81 FPun-sim 8.63 17.54 9.03 1.07 0.00 7.31 13.36 5.80 10.57 5.75 4.14 0.08 10.95 FN-sim 31.54 22.55 32.06 47.73 48.90 43.66 33.43 33.87 30.37 33.61 45.12 48.29 36.24 Hit 0.155 0.239 0.189 0.010 0.000 0.016 0.137 0.100 0.165 0.106 0.036 0.015 0.105 sim 0.154 0.240 0.174 0.009 0.000 0.014 0.125 0.100 0.155 0.103 0.033 0.012 0.102 Wrong 0.159 0.261 0.178 0.015 0.000 0.110 0.192 0.165 0.192 0.162 0.059 0.005 0.147 sim 0.134 0.229 0.147 0.013 0.000 0.094 0.170 0.144 0.164 0.129 0.051 0.004 0.127 Under 0.403 0.268 0.373 0.627 0.647 0.563 0.390 0.447 0.375 0.450 0.574 0.635 0.449 sim 0.429 0.299 0.415 0.629 0.647 0.580 0.425 0.467 0.407 0.475 0.586 0.639 0.471 Over 0.183 0.315 0.227 0.023 0.000 0.171 0.224 0.163 0.266 0.165 0.086 0.004 0.206 sim 0.183 0.320 0.230 0.023 0.000 0.169 0.229 0.159 0.264 0.150 0.089 0.005 0.211 Score 0.485 0.486 0.493 0.466 0.468 0.428 0.461 0.453 0.474 0.458 0.461 0.474 0.461 sim 0.493 0.498 0.496 0.466 0.468 0.432 0.462 0.461 0.478 0.469 0.462 0.473 0.466 Table 12: Detailed evaluation results across GEC systems on CoNLL-2014. We report True Positives (TPs), False Positives (FPs), False Negatives (FNs), and True Negatives (TNs) with or w/o similarity-based weighting (sim). 222