[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-191531-105":3,"detail-sidebar-cat-1-en-105":80,"doc-detail-191531-en":126},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":73,"head_meta":75,"extra_data":77,"updated_unix":79},105,"en","intent-detection-and-active-learning-based-correction-id-alc-algorithm-active-learning-alc-correction","Intent Detection and Active Learning based Correction (ID-ALC) Algorithm - Active Learning ALC Correction","","The document presents an intent detection framework combining active learning with correction. It defines the ID-ALC pipeline using labelled data, unlabelled data, a blind test set, and a predefined confidence threshold. A correction procedure applies a neural model to score unlabelled samples, rejects low-confidence cases, and re-labels them via majority voting across multiple traditional classifiers, followed by manual annotation and iterative retraining. Multiple datasets and multilingual settings are used to report accuracy, F1, and comparisons against several baseline and LLM-assisted methods.",{"@graph":14,"@context":72},[15,34,55],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/template/","Template",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/template/general/","General",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/template/intent-detection-and-active-learning-based-correction-id-alc-algorithm-active-learning-alc-correction/191531/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/intent-detection-and-active-learning-based-correction-id-alc-algorithm-active-learning-alc-correction/191531.png","ImageObject",442,249,{"name":42,"@type":43},"Gelato","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-09-13","2026-09-03",true,{"@type":52,"interactionType":53,"userInteractionCount":30},"InteractionCounter",{"@type":54},"ViewAction",{"@type":56,"mainEntity":57},"FAQPage",[58,64,68],{"name":59,"@type":60,"acceptedAnswer":61},"How does the ID-ALC pipeline use labelled, unlabelled, and test data?","Question",{"text":62,"@type":63},"It starts with labelled data (cycle 0), applies the model to unlabelled data, and evaluates on blind test data. The procedure iteratively updates labelled sets using active learning and correction outputs, then retrains and predicts on the test set.","Answer",{"name":65,"@type":60,"acceptedAnswer":66},"What triggers correction of an unlabelled sample in ALC?",{"text":67,"@type":63},"A predefined threshold (TH) is applied to the model’s confidence score. If the score is below TH, the sample is corrected using majority voting over multiple classifiers and then incorporated into the majority voting set for retraining.",{"name":69,"@type":60,"acceptedAnswer":70},"How are low-confidence rejected samples handled?",{"text":71,"@type":63},"Rejected samples are added to a separate set (R set) and then manually annotated. The final labelled set becomes the union of the previous labelled data, the majority voting set, and the manually annotated rejected samples.","https://schema.org",{"og:url":32,"og:type":74,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":76,"canonical":32},"index,follow",{"doc_id":78,"site_id":7},191531,1788409124,{"code":4,"msg":81,"data":82},"success",[83,88,93,98,103,108,113,118,123],{"id":84,"doc_module":22,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},11,"Presentations",90,"presentations",{"id":89,"doc_module":22,"doc_module_name":25,"category_name":90,"show_sort_weight":91,"slug":92},12,"Resumes",80,"resumes",{"id":94,"doc_module":22,"doc_module_name":25,"category_name":95,"show_sort_weight":96,"slug":97},14,"Invoices",70,"invoices",{"id":99,"doc_module":22,"doc_module_name":25,"category_name":100,"show_sort_weight":101,"slug":102},15,"Posters",60,"posters",{"id":104,"doc_module":22,"doc_module_name":25,"category_name":105,"show_sort_weight":106,"slug":107},16,"Social Media",50,"social-media",{"id":109,"doc_module":22,"doc_module_name":25,"category_name":110,"show_sort_weight":111,"slug":112},17,"Forms",40,"forms",{"id":114,"doc_module":22,"doc_module_name":25,"category_name":115,"show_sort_weight":116,"slug":117},18,"Letters",30,"letters",{"id":119,"doc_module":22,"doc_module_name":25,"category_name":120,"show_sort_weight":121,"slug":122},21,"Paper Templates",5,"papers-templates",{"id":124,"doc_module":22,"doc_module_name":25,"category_name":29,"show_sort_weight":4,"slug":125},158,"general-158",{"code":4,"msg":81,"data":127},{"doc_id":78,"user_id":128,"nickname":42,"user_avatar":129,"doc_module":22,"category_id":124,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":130,"file_id":131,"file_url":132,"file_type":133,"file_size":134,"view_count":22,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":89,"language":135,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":136,"faqs":137,"seo_title":138,"seo_description":12,"update_tm":79,"read_time":33},19241457091524,"https://us-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","| Dataset | \\# Intent\u003Cbr>Class | Total data | Type | \\# Lab | \\#Unlab |  |  | \\#Test |  |  |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n|  |  |  |  |  | \\#Ind | \\#OOD | \\#Tot | \\#Ind | \\#OOD | \\#Tot |\n| SNIPS | 7 (5 + 2 ) | 14484 | 1 | 2990 | 5661 | 3449 | 9110 | 1679 | 705 | 2384 |\n|  |  |  | 2 | 2990 | 5629 | 3481 | 9110 | 1709 | 675 | 2384 |\n|  |  |  | 3 | 2990 | 5689 | 3421 | 9110 | 1692 | 692 | 2384 |\n| FB | 12 (9 + 3) | 43323 | EN | 10000 | 9978 | 15022 | 25000 | 3849 | 4474 | 8323 |\n|  |  | 8643 | ES | 2000 | 2321 | 1678 | 3999 | 1512 | 1132 | 2644 |\n|  |  | 5083 | TH | 1000 | 1203 | 1797 | 3000 | 318 | 765 | 1083 |\n| ATIS | 17 (10 + 7 ) | 5871 | - | 1501 | 2688 | 682 | 3370 | 802 | 198 | 1000 |\n\n\n| Algorithm 1 Intent Detection and Active Learning based Correction (ID-ALC) Algorithm |\n| --- |\n| 1: Input\u003Cbr>2: L Labelled Data (Cycle 0)\u003Cbr>3: U Unlabelled Data\u003Cbr>4: T Blind Test Data\u003Cbr>5: TH A predefined Threshold\u003Cbr>6: MV Majority Voting Count\u003Cbr>7: Paramters\u003Cbr>8: M Neural Model\u003Cbr>9: L, Urem , M ← ID(L, U , M, T)\u003Cbr>10: L, Urem , M ← ALC(L, Urem , M, T, TH, MV) |\n\n\n| Algorithm 3 Active Learning ALC(L, Urem , M, T, TH, MV) | based | Correction: |\n| --- | --- | --- |\n| 1: procedure AL CORRECTION\u003Cbr>2: Apply Model M on Urem to predict intent class label and respective probability score\u003Cbr>3: for Each sample in Urem do\u003Cbr>4: if score \u003C TH then\u003Cbr>5: Apply Classifiers-RF, LR, Bg, KNN and LDA\u003Cbr>6: Label the sample with Majority Voting predictions (Vote Count ≤MV)\u003Cbr>7: Add Sample and Label to Majority Voting Set (MVS)\u003Cbr>8: Add Rejected Samples to R set\u003Cbr>9: end if\u003Cbr>10: end for\u003Cbr>11: Manually Annotate R set\u003Cbr>12: L = L ∪ MVS ∪ R\u003Cbr>13: Urem = Urem − MVS − R\u003Cbr>14: Retrain M on L and predict on T\u003Cbr>15: Repeat Steps 2 to 14 for K times\u003Cbr>16: end procedure |  |  |\n\n| Dataset | Method | Threshold | Acc (%) | F1 (%) |\n| --- | --- | --- | --- | --- |\n| SNIPS | MSP | 0.7 | 86.72 | 89.27 |\n|  | LOF | - | 81.21 | 83.44 |\n|  | DOC | - | 90.3 | 89.7 |\n| ATIS | MSP | 0.3 | 85.74 | 86.32 |\n|  | LOF | - | 78.2 | 78.05 |\n|  | DOC | - | 88.88 | 87.19 |\n| FB | MSP | 0.7 | 88.63 | 89.11 |\n|  | LOF | - | 81.49 | 82.37 |\n|  | DOC | - | 93.42 | 93.5 |\n\n\n| Class | Accuracy (%) |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- | --- |\n|  | SNIPS | ATIS | FB-EN | FB-ES | FB-TH | Avg |\n| RF | 95.4 | 89.8 | 91.2 | 90.6 | 88.3 | 91.1 |\n| Bg | 97.9 | 87.8 | 92.3 | 89.5 | 88.4 | 91.2 |\n| SVM | 87.4 | 85.8 | 86.42 | 83.5 | 80.4 | 84.7 |\n| LR | 97.8 | 96.5 | 94.7 | 95.9 | 97.4 | 96.5 |\n| XGB | 87.6 | 75.3 | 86.5 | 84.0 | 85.1 | 83.7 |\n| NB | 85.1 | 86.9 | 83.6 | 80.0 | 81.6 | 83.4 |\n| LDA | 97.8 | 97.02 | 95.4 | 96.7 | 97.6 | 96.9 |\n| KNN | 95.1 | 89.6 | 90.5 | 82.3 | 92.4 | 90.0 |\n| Gau | 90.2 | 79.1 | 80.2 | 84.0 | 85.7 | 83.9 |\n\n\n| Method | FB-EN |  | FB-ES |  | FB-TH |  | ATIS |  | SNIPS |  |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n|  | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 |\n| SEEN* | 84.9 | 86.2 | 75.4 | 79.8 | 71.9 | 77.5 | 84.3 | 71.4 | 85.8 | 86.2 |\n| Zero-Shot-OOD* | 86.7 | 87.4 | 62.1 | 70.1 | 68.6 | 72.1 | 84.8 | 81.7 | 87.4 | 88.5 |\n| SENC-MaS* | 82.3 | 83.9 | 69.4 | 70.5 | 64.6 | 72.8 | 80.1 | 69.4 | 82.0 | 81.4 |\n| SENNE* | 61.4 | 60.5 | 56.1 | 58.9 | 58.2 | 59.4 | 53.2 | 51.5 | 52.7 | 55.8 |\n| IFSTC | 84.7 | 82.0 | 80.4 | 54.3 | 86.1 | 87.2 | 71.1 | 71.4 | 70.4 | 71.1 |\n| TARS | 71.2 | 72.3 | 61.1 | 60.2 | 81.1 | 80.1 | 78.4 | 72.4 | 79.6 | 81.1 |\n| LLaMA2-FT | 92.2 | 78.1 | 89.1 | 71.5 | 83.5 | 79.0 | 81.6 | 73.3 | 90.6 | 90.1 |\n| LLaMA2-FT+ALC(2) | 92.7 | 82.4 | 91.3 | 76.2 | 86.3 | 84.1 | 82.5 | 77.5 | 92.6 | 92.9 |\n| ChatGPT 3.5T-Pr | 70.5 | 69.7 | 70.4 | 64.5 | 51.7 | 52.1 | 68.8 | 51.9 | 70.8 | 72.2 |\n| ID (0) | 61.9 | 72.4 | 66.6 | 74.2 | 68.1 | 80.2 | 82.6 | 81.1 | 73.7 | 80.2 |\n| ID + RA | 66.7 | 52.4 | 56.6 | 52.3 | 48.1 | 60.7 | 52.6 | 51.9 | 73.7 | 78.7 |\n| ID (1) + KM | 95.7 | 92.3 | 87.6 | 94.9 | 89.3 | 88.6 | 90.3 | 86.4 | 94.7 | 93.0 |\n| ID (1) ","cbCaihHzgXhohOnZ","https://ap.wps.com/l/cbCaihHzgXhohOnZ","pdf",509699,"English","# Intent Detection and Active Learning based Correction (ID-ALC) Algorithm\n## Active Learning ALC(L, Urem, M, T, TH, MV) Correction Procedure\n## Experimental Results and Performance Comparisons","[{\"question\":\"How does the ID-ALC pipeline use labelled, unlabelled, and test data?\",\"answer\":\"It starts with labelled data (cycle 0), applies the model to unlabelled data, and evaluates on blind test data. The procedure iteratively updates labelled sets using active learning and correction outputs, then retrains and predicts on the test set.\"},{\"question\":\"What triggers correction of an unlabelled sample in ALC?\",\"answer\":\"A predefined threshold (TH) is applied to the model’s confidence score. If the score is below TH, the sample is corrected using majority voting over multiple classifiers and then incorporated into the majority voting set for retraining.\"},{\"question\":\"How are low-confidence rejected samples handled?\",\"answer\":\"Rejected samples are added to a separate set (R set) and then manually annotated. The final labelled set becomes the union of the previous labelled data, the majority voting set, and the manually annotated rejected samples.\"}]","Intent Detection and Active Learning based Correction (ID-ALC) Algorithm - Active Learning ALC Correction | PDF"]