[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-513646-105":59,"doc-detail-513646-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","simple-dynamic-word-embeddings-for-mapping-perceptions-in-the-public-sphere","Simple dynamic word embeddings for mapping perceptions in the public sphere","","Word embeddings trained on large-scale historical corpora can illuminate human biases and stereotypes that reinforce social inequalities. This paper proposes a unified dynamic embedding approach that learns attribute-specific word embeddings in a shared framework. The method is applied to COHA to study 20th-century gender and ethnic occupation biases, and to a 119-million-word talk radio transcript corpus to examine refugee-related perceptions and associated linguistic biases.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/simple-dynamic-word-embeddings-for-mapping-perceptions-in-the-public-sphere/513646/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/simple-dynamic-word-embeddings-for-mapping-perceptions-in-the-public-sphere/513646.png","ImageObject",300,407,{"name":92,"@type":93},"AfterMidnight","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-08","2026-10-06",true,{"@type":102,"interactionType":103,"userInteractionCount":44},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What problem does the paper address about word embeddings and bias?","Question",{"text":112,"@type":113},"It addresses how biases and stereotypes can be captured through word embeddings, while also noting that many existing approaches train separate vector spaces for each attribute instead of leveraging shared training data effectively.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How does the paper’s dynamic embedding model work at a high level?",{"text":117,"@type":113},"It learns an attribute-invariant embedding for each word plus an attribute-specific offset, combining them to produce embeddings that vary with a chosen attribute context.",{"name":119,"@type":110,"acceptedAnswer":120},"Which datasets are used to test the proposed model?",{"text":121,"@type":113},"The model is applied to COHA for gender and ethnic occupation bias analysis, and to a talk radio transcript corpus totaling 119 million words to study biases against refugees.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},513646,1791385292,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":44,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":144,"read_time":41,"preview_page_count":19,"preview_urls":145},962090896061,"https://ap-avatar.wpscdn.com/davatar_9964176cb1d06d4a9deccf72a44ae3dc","Simple dynamic word embeddings for mapping perceptions in the public  \nsphere  \nNabeel Gillani  \nMIT  \n[ngillani@mit.edu](ngillani@mit.edu)  \nRoger Levy  \nMIT  \n[rplevy@mit.edu](rplevy@mit.edu)  \nThis is a modiﬁed version of a paper that originally appeared in the 2019 NAACL workshop on NLP+CSS. Since its initial publication, we discovered implementation errors that invalidate the results presented in the original version of the paper. These errors are corrected in this version. A corrigendum at the end of this paper summarizes the original errors and how we corrected them.  \nAbstract  \nWord embeddings trained on large-scale historical corpora can illuminate human biases and stereotypes that perpetuate social inequalities. These embeddings are often trained in separate vector space models deﬁned according to different attributes of interest. In this paper, we develop a uniﬁed dynamic embedding model that learns attribute-speciﬁc word embeddings. We apply our model to investigate i) 20th century gender and ethnic occupation biases embedded in the Corpus of Historical American English (COHA), and ii) biases against refugees embedded in a novel corpus of talk radio transcripts containing 119 million words produced over one month across 83 stations and 64 cities. Our results shed preliminary light on scenarios when dynamic embedding models may be more suitable for representing linguistic biases than individual vector space models, and vice-versa.  \n1 Introduction  \nLanguage has long been described as both a cause and reﬂection of our psycho-social contexts (Lewis and Lupyan, 2018) . Recent work using word embeddings—low-dimensional vector representations of words trained on large datasets to capture key semantic information—has demonstrated that language encodes several gender, racial, and other biases that correlate with both implicit biases (Caliskan et al., 2017) and historical trends (Garg et al., 2018) .  \nThese studies have validated the use of word embeddings to measure a range of psychological and social contexts, yet in most cases, they do not leverage the full power of available datasets. For example, the historical biases presented in (Garget al., 2018) are computed using decade-speciﬁc word embeddings produced by training different Word2Vec (Mikolov et al., 2013) models on a large corpus of historical text from that decade. The authors then use a Procrustes alignment to project embeddings from different models into the same vector space so they can be compared across decades (Hamilton et al., 2016) . While this approach is reasonable when there are large-scale datasets available for a given attribute of interest (e.g. decade), it requires an additional optimization step and disregards valuable training data that could be pooled and leveraged across attribute values to help with both training and regularization—especially when data is sparse.  \nIn this paper, we present a uniﬁed dynamic word embedding model that jointly trains linguistic information alongside any categorical variable of interest describing its context (-e.g. year, geography, income bracket, etc.) . We apply this model to two datasets: i) the Corpus of Historical American English (COHA (Davies, 2010)) to analyze gender and ethnic occupation biases, and ii) a novel data corpus of 119 million words spoken on talk radio (Beeferman and Roy, 2018) during a onemonth period in late 2018 across 64 US cities to explore perceptions about refugees. Our results shed preliminary light on scenarios when dynamic embedding models may be more suitable for representing linguistic biases than individual vector space models, and vice-versa.  \n2 Model  \nWe describe our model and implementation below.  \n2.1 Overview  \nOur dynamic embedding for word w is deﬁned as  \nE (w; A) = 􀀍w + 􀀆 a2A 􀀌aw (1)  \nwhere 􀀍w is an attribute-invariant embedding of w computed across the entire corpus, 􀀌aw is the offset for w with respect to attribute a across the set of attributes A we are interes","cbCaioinYQEHYSBp","https://ap.wps.com/l/cbCaioinYQEHYSBp","pdf",647612,12,"English","# Abstract\n# 1 Introduction\n# 2 Model\n## 2.1 Overview\n## 2.2 Implementation","[{\"question\":\"What problem does the paper address about word embeddings and bias?\",\"answer\":\"It addresses how biases and stereotypes can be captured through word embeddings, while also noting that many existing approaches train separate vector spaces for each attribute instead of leveraging shared training data effectively.\"},{\"question\":\"How does the paper’s dynamic embedding model work at a high level?\",\"answer\":\"It learns an attribute-invariant embedding for each word plus an attribute-specific offset, combining them to produce embeddings that vary with a chosen attribute context.\"},{\"question\":\"Which datasets are used to test the proposed model?\",\"answer\":\"The model is applied to COHA for gender and ethnic occupation bias analysis, and to a talk radio transcript corpus totaling 119 million words to study biases against refugees.\"}]","Simple dynamic word embeddings for mapping perceptions in the public sphere | PDF",1791288361,[146,150,154,158],{"page_no":8,"url":147,"width":148,"height":149},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/db1fc3ce8455f3412d219d17cbcb0b35.webp",983,1390,{"page_no":14,"url":151,"width":152,"height":153},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/5658db53d9f3cc78d8204ca61486d07f.webp",982,1388,{"page_no":81,"url":155,"width":156,"height":157},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/ce0e6e1c8c5ab0b61f4e69b07df1fbab.webp",981,1387,{"page_no":19,"url":159,"width":160,"height":161},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/b4412cb46016bb4391198133846a4395.webp",980,1385]