exbert-project/exbert
183
1<!DOCTYPE html>2<html lang="en">3 4<head>5 <!-- Global site tag (gtag.js) - Google Analytics -->6 <script async src="https://www.googletagmanager.com/gtag/js?id=UA-149275901-1"></script>7 <script>8 window.dataLayer = window.dataLayer || [];9 function gtag() { dataLayer.push(arguments); }10 gtag('js', new Date());11 12 gtag('config', 'UA-149275901-2');13 </script>14 <meta charset="UTF-8">15 <title>exBERT</title>16 <!--<link rel="stylesheet" href="bootstrap.min.css">-->17 <link href="https://maxcdn.bootstrapcdn.com/font-awesome/4.6.3/css/font-awesome.min.css" rel="stylesheet"18 integrity="sha384-T8Gy5hrqNKT+hzMclPo118YTQO6cYprQmhrYwIiQ/3axmI1hQomh7Ud2hPOy8SP1" crossorigin="anonymous">19 <link href="https://maxcdn.bootstrapcdn.com/bootswatch/3.3.6/cosmo/bootstrap.min.css" rel="stylesheet"20 integrity="sha384-OiWEn8WwtH+084y4yW2YhhH6z/qTSecHZuk/eiWtnvLtU+Z8lpDsmhOKkex6YARr" crossorigin="anonymous">21 <!-- <link rel="stylesheet" type="text/css" href="css/index.scss">-->22 <link href="https://fonts.googleapis.com/css?family=IBM+Plex+Sans:300,400&display=swap" rel="stylesheet">23 24 25 <link rel="shortcut icon" href="./img/exBERT_favicon.png" type="image/png">26 <style>27 body {28 font-family: 'IBM Plex Sans', sans-serif;29 30 }31 32 h1,33 h2,34 h3 {35 font-family: inherit;36 }37 38 h3 {39 font-weight: normal;40 color: white;41 }42 43 a {44 color: black;45 }46 47 a:hover {48 color: black;49 font-weight: 500;50 }51 </style>52 <meta name="title" content="exBERT" />53 <meta name="description" content="A visual exploration of BERT's latent space" />54 <link rel="image_src" href="http://lstm.seas.harvard.edu/small_teaser.png" />55</head>56 57<body>58 <div class="container">59 <div class="page-header">60 <h1>61 <img src="./img/exBERT.png" height="50" />62 - A Visual Analysis of BERT63 </h1>64 <p class="lead">65 <a target="_blank" href="">Benjamin Hoover</a>,66 <a target="_blank" href="http://hendrik.strobelt.com">Hendrik Strobelt</a>,67 <a target="_blank" href="http://scholar.harvard.edu/gehrmann/home">Sebastian Gehrmann</a>68 </p>69 </div>70 71 <p>72 Large language models can produce powerful contextual representations that lead to improvements across many73 NLP tasks.74 Since these models are typically guided by a sequence of learned self attention mechanisms and may comprise75 undesired inductive biases, it is paramount to be able to explore what the attention has learned.76 While static analyses of these models lead to targeted insights, interactive tools are more dynamic and can77 help humans better gain an intuition for the model-internal reasoning process.78 </p>79 80 <p>81 We present <b>exBERT</b> , an interactive tool named after the popular BERT language model, that provides82 insights into the meaning of the contextual representations by matching a human-specified input to similar83 contexts in a large annotated dataset.84 By aggregating the annotations of the matching similar contexts, <b>exBERT</b> helps intuitively explain85 what each attention-head has learned.86 </p>87 88 <p> Large language models can produce powerful contextual representations that lead to improvements across many89 NLP tasks. Though these models can comprise undesired inductive biases, it is challenging to identify what90 information they encode in their learned representations. </p>91 92 <p> Since the model-internal reasoning process is often guided by a sequence of learned self-attention93 mechanisms, it is paramount to be able to explore what the attention has learned. While static analyses for94 this can lead to targeted insights, interactive tools can be more dynamic and help humans gain an intuition95 for the model-internal reasoning process. We present exBERT, a tool that helps to gain insights into the96 meaning of the contextual representations. exBERT matches a human-specified input to similar contexts in a97 large annotated dataset. By aggregating these annotations across all similar contexts, exBERT can help to98 explain what each attention-head has learned. </p>99 100 <p> Thanks to101 <a target="_blank" href="https://www.parc.com/about-parc/our-people/jesse-vig/">Jesse Vig</a>102 for feedback. Please let us know what you think by commenting below! </p>103 104 <div style="height: 20px"></div>105 106 107 <div class="row">108 <div class="col-md-4">109 <div class="text-center" style="background-color: rgb(81,116,144);">110 <a href="exBERT.html">111 <h3><i class="fa fa-rocket" aria-hidden="true" style="font-size: 80%; "></i> Live Server </h3>112 </a>113 </div>114 115 116 </div>117 <div class="col-md-4">118 <div class="text-center" style="background-color: rgb(203,122,130);padding: inherit">119 <a href="https://github.com/bhoov/transformer-viz">120 <h3><i class="fa fa-github" aria-hidden="true" style="font-size: 80%;"></i> Source Code </h3>121 </a>122 </div>123 </div>124 <div class="col-md-4">125 <div class="text-center" style="background-color: rgb(152,152,152);padding: inherit">126 <a href="">127 <h3><i class="fa fa-book" aria-hidden="true" style="font-size: 80%;"></i> Pre-Publication </h3>128 </a>129 </div>130 </div>131 </div>132 133 134 <div class="row">135 <div class="col-md-8 col-md-offset-2">136 <div class='col-sm'>137 <img id='tool-screenshot' src="./img/exBERT_overview.png" alt="Tool screenshot" height="600px">138 </div>139 </div>140 </div>141 142 <hr>143 144 <div class="row">145 <div class="col-md-4">146 <div class="text-center" style="background-color: rgb(81,116,144);">147 <a href="exBERT.html">148 <h3><i class="fa fa-rocket" aria-hidden="true" style="font-size: 80%; "></i> Live Server </h3>149 </a>150 </div>151 152 153 </div>154 <div class="col-md-4">155 <div class="text-center" style="background-color: rgb(203,122,130);padding: inherit">156 <a href="https://github.com/bhoov/transformer-viz">157 <h3><i class="fa fa-github" aria-hidden="true" style="font-size: 80%;"></i> Source Code </h3>158 </a>159 </div>160 </div>161 <div class="col-md-4">162 <div class="text-center" style="background-color: rgb(152,152,152);padding: inherit">163 <a href="">164 <h3><i class="fa fa-book" aria-hidden="true" style="font-size: 80%;"></i> Pre-Publication </h3>165 </a>166 </div>167 </div>168 </div>169 170 171 <div class="vspace-40"></div>172 173 <div class="row">174 <div class="col-12 text-center gray-1">175 <p></p>176 <p>We care about your privacy, but know that your activity on the site may be monitored. For more177 information, check out the links below.</p>178 </div>179 </div>180 181 <div class="row">182 <div class="col-lg-12 text-center gray-1">183 <a target="_blank" href="http://www.ibm.com/legal/us/en/?lnk=flg-tous-usen" class="gray-1">Terms of184 Use</a> |185 <a target="_blank" href="http://www.ibm.com/privacy/us/en/?lnk=flg-priv-usen" class="gray-1">Privacy</a>186 </div>187 </div>188 189 190 <hr>191 192 <div class="row justify-content-md-center">193 <!-- <div class="col-xs-offset-2 col-lg-4 text-center" style="font-size: 16pt;"> -->194 <div class="col-lg-6 text-center" style="font-size: 16pt;">195 <span style="white-space: nowrap;"><b>IBM</b> Research, Cambridge </span><br>196 <b>MIT-IBM</b> Watson <b>AI Lab</b>197 </div>198 <div class="col-xs-3 text-center">199 <a target="_blank" href="http://nlp.seas.harvard.edu/"><img class="img-responsive"200 src="./img/harvardnlp_logo.png"></a>201 </div>202 </div>203 204 205 206 207 <!-- DISQUS THREAD -->208 <div id="disqus_thread"></div>209 <script>210 211 /**212 * RECOMMENDED CONFIGURATION VARIABLES: EDIT AND UNCOMMENT THE SECTION BELOW TO INSERT DYNAMIC VALUES FROM YOUR PLATFORM OR CMS.213 * LEARN WHY DEFINING THESE VARIABLES IS IMPORTANT: https://disqus.com/admin/universalcode/#configuration-variables*/214 /*215 var disqus_config = function () {216 this.page.url = PAGE_URL; // Replace PAGE_URL with your page's canonical URL variable217 this.page.identifier = PAGE_IDENTIFIER; // Replace PAGE_IDENTIFIER with your page's unique identifier variable218 };219 */220 (function () { // DON'T EDIT BELOW THIS LINE221 var d = document, s = d.createElement('script');222 s.src = 'https://exbert.disqus.com/embed.js';223 s.setAttribute('data-timestamp', +new Date());224 (d.head || d.body).appendChild(s);225 })();226 </script>227 <noscript>Please enable JavaScript to view the <a href="https://disqus.com/?ref_noscript">comments powered by228 Disqus.</a></noscript>229 230 </div>231 232 233 <!-- <script id="dsq-count-scr" src="//exbert.disqus.com/count.js" async></script> -->234</body>235 236</html>