[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-123303-en":3,"doc-seo-123303-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},123303,2336464648746,"Skyler","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","Human-Machine Rationale Alignment for Better Machine Learning Interpretability - Dissertation","Recent advances in deep neural network models achieve strong results in computer vision, natural language processing, and data analysis, yet many systems still behave as black boxes. Revealing how a model reasons has therefore become essential for safe and effective development. This dissertation studies human–machine rationale alignment through three complementary perspectives, including measuring dissonance between human and machine explanations, training interpretable-by-design models supervised by human rationales, and translating knowledge from non-interpretable models into human-understandable rules. Extensive evaluations cover sentiment analysis, fact-checking, and reading comprehension.","Human-Machine Rationale Alignment for Better Machine Learning Interpretability  \nVon der Fakultät für Elektrotechnik und Informatik der Gottfried Wilhelm Leibniz Universität Hannover zur Erlangung des Grades  \nDoktor der Naturwissenschaften  \n(abgekürzt: Dr. rer. nat.)  \ngenehmigte Dissertation von  \nM. Sc. Zijian Zhang  \ngeboren am 23 . Mai 1992  \nin Tangshan, Hebei, China  \nii  \nReferent: Prof. Dr. techn. Wolfgang Nejdl Korreferent: Prof. Dr. Avishek Anand Vorsitz: Prof. Dr. Henning Wachsmut  \nTag der Promotion: 16 . Juni 2025  \nAbstract  \nThe recent advancements in deep neural network models have yielded significant achievements in various cognitive tasks across fields such as computer vision (CV), natural language processing (NLP), and data analysis. Despite their impressive performance, these models often operate as black boxes, posing substantial risksand challenges in their development. Consequently, interpreting machine learning by revealing its reasoning process has become a critical area of research. Aligning human users’ reasoning processes with machine learning (ML) models is now vital for interpreting and comprehending these models.  \nThis thesis details my doctoral research on aligning human and machine reasoning, encompassing three distinct perspectives.  \n1. Differences between human and machine rationales – In this thesis’s first research, we explore the differences in reasoning between humans and machines when faced with the same specific cognitive task. This research starts by designing an interactive application to collect humans’ rationalizations on image classification. The application asks the subjects to rank the features according to their importance to the classification. We compare these rationales with machines’ rationales identified by post-hoc feature attribution approaches. Another group of subjects then evaluates the absolute goodness of the ranking using a per-feature revealing game, where the order of revealing follows the ranking from both machines and humans. The results of this research expose a significant dissonance between human and machine rationalization and stress the importance of aligning human and machine reasoning. Chapter 3 describes this piece of research.  \n2. Training an interpretable-by-design model to rationalize like humans – We then extend the research scope by training an interpretable-by-design model to bridge the reasoning gap between humans and machines. The model performs its original NLP classification task well and rationalizes the prediction by evidence extracted from the input. The extractive rationales are supervised and thus aligned with humans’ annotation of the rationalization. I present this research in Chapter 4, while Chapters 4.7.1 and 4.7.2 explore its application to interpretable machine learning.  \n3. Translating rationales of non-interpretable models to a human-understandable form  \n– The third piece of our research focuses on translating machine-acquired knowledge  \niv  \ninto a format understandable to humans. Specifically, we verify that the deep models are vulnerable to overfitting the shortcuts between the input sequencesand the label in the NLP task, but ignore the collective semantics of the whole sentence during training. This approach translates the models’ decision-making process to human-understandable rules, based on which human inspectors can identify spurious sequence-label correlations between input sequences and labels. Such identified spurious correlations result in the model’s good prediction of the data within the distribution but increase the model’s vulnerability against outof-distribution data when deployed in the field. Chapter 5 presents this piece of research.  \nThis thesis examines whether the ML model engages in causal reasoning akin to humans or merely internalizes statistical correlations, potentially leading to stereotypical errors in practical applications. It also presents several applications empowered by the interpretab","cbCaicK5QpWJXbMZ","https://ap.wps.com/l/cbCaicK5QpWJXbMZ","pdf",8606470,1,154,"English","en",105,"# Abstract\n## Human-machine rationale alignment overview\n## Research perspective 1: differences between human and machine rationales\n## Research perspective 2: training an interpretable-by-design model to rationalize like humans\n## Research perspective 3: translating rationales of non-interpretable models","[{\"question\":\"Why does machine learning interpretability require aligning human and machine reasoning?\",\"answer\":\"Deep models often act like black boxes, making their decisions risky and hard to trust. Aligning reasoning processes helps users interpret and understand model predictions by making explanations more comparable to human rationales.\"},{\"question\":\"How does the dissertation study differences between human and machine rationales?\",\"answer\":\"It designs an interactive application to collect human feature rankings for image classification and compares them with machine rationales from post-hoc feature attribution. A per-feature revealing game then evaluates how well the revealed ranking matches the absolute goodness of both groups.\"},{\"question\":\"What is the goal of translating rationales from non-interpretable models?\",\"answer\":\"The work converts model decision-making into human-understandable rules. It focuses on how models can rely on spurious sequence-label shortcuts, and the translated rules enable inspectors to identify those harmful correlations for improving robustness out of distribution.\"}]","Human-Machine Rationale Alignment for Better Machine Learning Interpretability - Dissertation | PDF",1785815838,388,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"human-machine-rationale-alignment-for-better-machine-learning-interpretability-dissertation","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/human-machine-rationale-alignment-for-better-machine-learning-interpretability-dissertation/123303/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-04",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why does machine learning interpretability require aligning human and machine reasoning?","Question",{"text":75,"@type":76},"Deep models often act like black boxes, making their decisions risky and hard to trust. Aligning reasoning processes helps users interpret and understand model predictions by making explanations more comparable to human rationales.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the dissertation study differences between human and machine rationales?",{"text":80,"@type":76},"It designs an interactive application to collect human feature rankings for image classification and compares them with machine rationales from post-hoc feature attribution. A per-feature revealing game then evaluates how well the revealed ranking matches the absolute goodness of both groups.",{"name":82,"@type":73,"acceptedAnswer":83},"What is the goal of translating rationales from non-interpretable models?",{"text":84,"@type":76},"The work converts model decision-making into human-understandable rules. It focuses on how models can rely on spurious sequence-label shortcuts, and the translated rules enable inspectors to identify those harmful correlations for improving robustness out of distribution.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]