[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-1-en-105":3,"doc-seo-194628-105":53,"doc-detail-194628-en":126},{"code":4,"msg":5,"data":6},0,"success",[7,14,19,24,29,34,39,44,49],{"id":8,"doc_module":9,"doc_module_name":10,"category_name":11,"show_sort_weight":12,"slug":13},11,1,"Template","Presentations",90,"presentations",{"id":15,"doc_module":9,"doc_module_name":10,"category_name":16,"show_sort_weight":17,"slug":18},12,"Resumes",80,"resumes",{"id":20,"doc_module":9,"doc_module_name":10,"category_name":21,"show_sort_weight":22,"slug":23},14,"Invoices",70,"invoices",{"id":25,"doc_module":9,"doc_module_name":10,"category_name":26,"show_sort_weight":27,"slug":28},15,"Posters",60,"posters",{"id":30,"doc_module":9,"doc_module_name":10,"category_name":31,"show_sort_weight":32,"slug":33},16,"Social Media",50,"social-media",{"id":35,"doc_module":9,"doc_module_name":10,"category_name":36,"show_sort_weight":37,"slug":38},17,"Forms",40,"forms",{"id":40,"doc_module":9,"doc_module_name":10,"category_name":41,"show_sort_weight":42,"slug":43},18,"Letters",30,"letters",{"id":45,"doc_module":9,"doc_module_name":10,"category_name":46,"show_sort_weight":47,"slug":48},21,"Paper Templates",5,"papers-templates",{"id":50,"doc_module":9,"doc_module_name":10,"category_name":51,"show_sort_weight":4,"slug":52},158,"General","general-158",{"code":4,"msg":54,"data":55},"ok",{"site_id":56,"language":57,"slug":58,"title":59,"keywords":60,"description":61,"schema_data":62,"social_meta":119,"head_meta":121,"extra_data":123,"updated_unix":125},105,"en","2023acl-long380","2023.acl-long.380","","A study evaluates grammatical error correction performance using multiple transformer-based model variants and decoding setups. Results compare systems across English and Chinese benchmarks, reporting precision, recall, and F0.5 for detection and correction behavior. Experiments include template consistency, prediction versus gold evaluation settings, and ablations over backbone architectures (ELECTRA, GECToR, BART, T5, Transformer variants). Additional configuration sweeps cover learning rates, token limits, optimizers, warmup, loss functions, dropout, and beam size to quantify impact on accuracy.",{"@graph":63,"@context":118},[64,80,101],{"@type":65,"itemListElement":66},"BreadcrumbList",[67,71,74,77],{"item":68,"name":69,"@type":70,"position":9},"https://docshare.wps.com","Home","ListItem",{"item":72,"name":10,"@type":70,"position":73},"https://docshare.wps.com/template/",2,{"item":75,"name":51,"@type":70,"position":76},"https://docshare.wps.com/template/general/",3,{"item":78,"name":59,"@type":70,"position":79},"https://docshare.wps.com/template/2023acl-long380/194628/",4,{"url":78,"name":59,"@type":81,"image":82,"author":87,"headline":59,"publisher":90,"fileFormat":93,"inLanguage":57,"description":61,"dateModified":94,"datePublished":95,"encodingFormat":93,"isAccessibleForFree":96,"interactionStatistic":97},"DigitalDocument",{"url":83,"@type":84,"width":85,"height":86},"https://docshare.wps.com/thumbnails/2023acl-long380/194628.png","ImageObject",442,249,{"name":88,"@type":89},"Patrick","Person",{"url":68,"name":91,"@type":92},"DocShare","Organization","application/pdf","2026-09-28","2026-09-03",true,{"@type":98,"interactionType":99,"userInteractionCount":76},"InteractionCounter",{"@type":100},"ViewAction",{"@type":102,"mainEntity":103},"FAQPage",[104,110,114],{"name":105,"@type":106,"acceptedAnswer":107},"What metrics are used to evaluate the models?","Question",{"text":108,"@type":109},"The document reports precision (P), recall (R), and F0.5 for evaluation, with separate results across detection and correction-related settings.","Answer",{"name":111,"@type":106,"acceptedAnswer":112},"Which datasets are used for English and Chinese experiments?",{"text":113,"@type":109},"English results reference corpora such as cLang-8, WI/LOCNESS, and CoNLL-14, while Chinese results reference NLPCC18 and MuCGEC.",{"name":115,"@type":106,"acceptedAnswer":116},"What kinds of ablation experiments are performed?",{"text":117,"@type":109},"Experiments vary template settings (e.g., default templates), template consistency, and detection label options (such as 2-class vs 4-class), along with different backbone architectures.","https://schema.org",{"og:url":78,"og:type":120,"og:title":59,"og:site_name":91,"og:description":61},"article",{"robots":122,"canonical":78},"index,follow",{"doc_id":124,"site_id":56},194628,1790437498,{"code":4,"msg":5,"data":127},{"doc_id":124,"user_id":128,"nickname":88,"user_avatar":129,"doc_module":9,"category_id":50,"category_name":51,"doc_title":59,"doc_description":61,"doc_content":130,"file_id":131,"file_url":132,"file_type":133,"file_size":134,"view_count":76,"is_deleted":4,"is_public":9,"is_downloadable":9,"audit_status":9,"page_count":25,"language":135,"language_code":57,"site_id":56,"html_lang":57,"table_of_contents":136,"faqs":137,"seo_title":138,"seo_description":61,"update_tm":139,"read_time":47},549758146520,"https://ap-avatar.wpscdn.com/avatar/80002397d8c0411e94?_k=1775819394049821470","|  | He prefer study in home. |\n| --- | --- |\n|  | |\n|  | Detection  Correction\u003Cbr>TemplateGEC\u003Cbr>|\n|  |  |\n|  |  |\n|  | |\n|  | He prefers to study at home. |\n\n\n| It\u003Cbr>is difficult\u003Cbr>answer\u003Cbr>at\u003Cbr>the question\u003Cbr>. |  | Error Parts 􀀉􀀂\u003Cbr>C C C I C I C C\u003Cbr>C C C I I C C C Gold label 􀀁􀀄 ERRANT  (answer at) \u003Cbr>Error Parts 􀀉􀀄 | \u003Cbr>\u003Cbr>  Prediction of 􀀃􀀇 \u003Cbr> Seq2Seq Model  ℒg: Loss of 􀀃􀀄 \u003Cbr> Template 􀀃􀀄\u003Cbr> \u003CS1> answer at \u003Csep>  It is difficult \u003CS1> the question . \u003Cbr>\u003Cbr>answer\u003Cbr>the question\u003Cbr>Detection Prefix 􀀈􀀄 Modified Source 􀀋′􀀄\u003Cbr>.\u003Cbr>|\n| --- | --- | --- | --- |\n|  |  |  |  |\n|  |  |  |  |\n\n\n| System | NLPCC18-Test (ZH) |  |  | BEA-Dev (EN) |  |  | CoNLL14-Test 1 (EN) |  |  | CoNLL14-Test 2 (EN) |  |  |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n|  | P | R | F0.5 | P | R | F0.5 | P | R | F0.5 | P | R | F0.5 |\n| ELECTRA(Yuan et al., 2021) | - | - | - | 72.8 | 46.9 | 65.6 | 55.2 | 39.8 | 51.2 | 76.4 | 40.1 | 64.7 |\n| GECToR (Omelianchuk et al., 2020)   | - | - | - | 75.4 |  52.6  | 69.4   | 55.8 | 38.9 | 51.3 | 77.4 | 38.8 | 64.6  |\n| ELECTRA (Our Reproduced) | 70.1 | 37.5 | 59.7 |  73.7  | 41.4 | 63.8 |  57.1  | 36.4  | 51.3 |  75.9  | 34.8 | 61.4 |\n\n\n| Configuration | English | Chinese |\n| --- | --- | --- |\n| Architecture | Transformer-large BART-large T5-large | Transformer-large BART-large |\n| Epochs | 30 20 5 | 30 10 |\n| Max Tokens | 16384 4096 2048 | 8192 2048 |\n| Learning Rate | 5×10−4 1×10−5 1×10−3 | 5×10−4 3×10−5 |\n| Optimizer | Adam (Kingma and Ba, 2015) Adafactor\u003Cbr>(β1 = 0 .9,β2 = 0 .98,ϵ = 1 × 10−6) (Shazeer and Stern, 2018) | Adam (Kingma and Ba, 2015)\u003Cbr>(β1 = 0 .9,β2 = 0 .98,ϵ = 1 × 10−6) |\n| Warmup | 4000 8000 4000 | 2000 2000 |\n| Loss Function | label smoothed cross entropy (label-smoothing=0.1) | (Szegedy et al., 2016) |\n| Dropout | 0.1 0.3 0.3 | 0.1 0.3 |\n| Beam Size | 5 5 5 | 12 12 |\n\n\n| Language | Corpus | Train | Dev | Test |\n| --- | --- | --- | --- | --- |\n| English | cLang-8 | 2,372,119 | - | - |\n| English | WI, LOCNESS | - | 4,384 | 4,477 |\n| English | CoNLL-14 | - | - | 1,312 |\n| Chinese | NLPCC18 | 1,377,172 | - | 2,000 |\n| Chinese | MuCGEC | - | 2,467 | - |\n\n\n| System | Proposed Methods |  | Detection Label |  | NLPCC18-Test (ZH) |  |  | BEA-Test (EN) |  |  | CoNLL14-Test (EN) |  |  |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n|  | Template | Consistency | Train | Test | P | R | F0.5 | P | R | F0.5 | P | R | F0.5 |\n| GECToR | ✕ | ✕ | - | - | - | - | - | 79.2 | 53.9 | 72.4 | 77.5 | 40.1 | 65.3 |\n| Multi-encoder | ✕ | ✕ | - | - | - | - | - | 73.3 | 61.5 | 70.6 | 71.3 | 44.3 | 63.5 |\n| T5-large | ✕ | ✕ | - | - | - | - | - | - | - | 72.1 | - | - | 66.1 |\n| Type-Driven | ✕ | ✕ | - | - | - | - | - | 81.3 | 51.6 | 72.9 | 78.2 | 42.7 | 67.0 |\n| SynGEC | ✕ | ✕ | - | - | 50.0 | 33.0 | 45.3 | 75.1 | 65.5 | 72.9 | 74.7 | 49.0 | 67.6 |\n| Transformer\u003Cbr>| ✕ | ✕ | - | - | 36.1 | 19.9 | 31.0 | 56.2 | 51.5 | 55.2 | 59.3 | 39.9 | 54.0 |\n|  | ✓ | ✕ | Pred | Pred | 37.2 | 23.9 | 33.5 | 60.0 | 51.7 | 58.1 | 61.1 | 40.0 | 55.3 |\n|  | ✓ | ✓ | Gold+Pred | Pred | 42.0 | 22.2 | 35.6 | 67.8 | 50.7 | 63.5 | 64.7 | 38.9 | 57.1 |\n| BART\u003Cbr>| ✕ | ✕ | - | - | 48.8 | 33.5 | 44.7 | 70.4 | 60.0 | 68.0 | 67.1 | 47.1 | 61.9 |\n|  | ✓ | ✕ | Pred | Pred | 52.2 | 27.9 | 44.5 | 71.7 | 61.5 | 69.4 | 67.6 | 48.5 | 62.6 |\n|  | ✓ | ✓ | Gold+Pred | Pred | 54.5 | 27.4 | 45.5 | 74.8 | 61.0 | 71.6 | 69.7 | 46.7 | 63.5 |\n| T5\u003Cbr>| ✕ | ✕ | - | - | - | - | - | 74.2 | 66.5 | 72.5 | 71.8 | 50.8 | 66.3 |\n|  | ✓ | ✕ | Pred | Pred | - | - | - | 74.6 | 64.4 | 72.3 | 72.4 | 50.7 | 66.7 |\n|  | ✓ | ✓ | Gold+Pred | Pred | - | - | - | 76.8 | 64.8 | 74.1 | 74.8 | 50.0 | 68.1 |\n\n\n| System | Proposed Methods |  | Detection Label |  | NLPCC18-Test (ZH) |  |  | BEA-Dev (EN) |  |  | CoNLL14-Test (EN) |  |  |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n|  | Template | Consistency | Train | Test | P | R | F0.5 | P | R | F0.5 | P | R | F0","cbCaikC4kwA1ZP0p","https://ap.wps.com/l/cbCaikC4kwA1ZP0p","pdf",383159,"English","# Experimental Setup\n## Architectures and Systems\n## Datasets and Evaluation\n## Configuration Details\n## Results and Ablations","[{\"question\":\"What metrics are used to evaluate the models?\",\"answer\":\"The document reports precision (P), recall (R), and F0.5 for evaluation, with separate results across detection and correction-related settings.\"},{\"question\":\"Which datasets are used for English and Chinese experiments?\",\"answer\":\"English results reference corpora such as cLang-8, WI/LOCNESS, and CoNLL-14, while Chinese results reference NLPCC18 and MuCGEC.\"},{\"question\":\"What kinds of ablation experiments are performed?\",\"answer\":\"Experiments vary template settings (e.g., default templates), template consistency, and detection label options (such as 2-class vs 4-class), along with different backbone architectures.\"}]","2023.acl-long.380 | PDF",1788440904]