<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40"><head><meta http-equiv=Content-Type content="text/html; charset=utf-8"><meta name=Generator content="Microsoft Word 15 (filtered medium)"><style><!--
/* Font Definitions */
@font-face
{font-family:新細明體;
panose-1:2 2 5 0 0 0 0 0 0 0;}
@font-face
{font-family:"Cambria Math";
panose-1:2 4 5 3 5 4 6 3 2 4;}
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
@font-face
{font-family:"\@新細明體";
panose-1:2 1 6 1 0 1 1 1 1 1;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0cm;
margin-bottom:.0001pt;
font-size:12.0pt;
font-family:"新細明體",serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:blue;
text-decoration:underline;}
a:visited, span.MsoHyperlinkFollowed
{mso-style-priority:99;
color:purple;
text-decoration:underline;}
p.msonormal0, li.msonormal0, div.msonormal0
{mso-style-name:msonormal;
mso-margin-top-alt:auto;
margin-right:0cm;
mso-margin-bottom-alt:auto;
margin-left:0cm;
font-size:12.0pt;
font-family:"新細明體",serif;}
span.EmailStyle18
{mso-style-type:personal-reply;
font-family:"Calibri",sans-serif;
color:#1F497D;}
.MsoChpDefault
{mso-style-type:export-only;
font-family:"Calibri",sans-serif;}
@page WordSection1
{size:612.0pt 792.0pt;
margin:72.0pt 90.0pt 72.0pt 90.0pt;}
div.WordSection1
{page:WordSection1;}
--></style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]--></head><body lang=ZH-TW link=blue vlink=purple><div class=WordSection1><p class=MsoNormal><span lang=EN-US style='font-family:"Calibri",sans-serif;color:#1F497D'>Thanks a lot for your great effort. Could you please show us those remaining errors (so that we can check if they can be further improved)? Thanks.<o:p></o:p></span></p><p class=MsoNormal><span lang=EN-US style='font-family:"Calibri",sans-serif;color:#1F497D'><o:p> </o:p></span></p><p class=MsoNormal><span lang=EN-US style='font-family:"Calibri",sans-serif;color:#1F497D'>KY<o:p></o:p></span></p><p class=MsoNormal><span lang=EN-US style='font-family:"Calibri",sans-serif;color:#1F497D'><o:p> </o:p></span></p><p class=MsoNormal><b><span lang=EN-US style='font-size:11.0pt;font-family:"Calibri",sans-serif'>From:</span></b><span lang=EN-US style='font-size:11.0pt;font-family:"Calibri",sans-serif'> most-ai-contest-bounces@iis.sinica.edu.tw [mailto:most-ai-contest-bounces@iis.sinica.edu.tw] <b>On Behalf Of </b></span><span style='font-size:11.0pt'>郭家銍</span><span lang=EN-US style='font-size:11.0pt;font-family:"Calibri",sans-serif'><br><b>Sent:</b> Thursday, April 2, 2020 3:45 PM<br><b>To:</b> Most-ai Contest <Most-ai-contest@iis.sinica.edu.tw><br><b>Subject:</b> [Most-ai-contest] Date-duration</span><span style='font-size:11.0pt'>模組</span><span lang=EN-US style='font-size:11.0pt;font-family:"Calibri",sans-serif'> ver4.0<o:p></o:p></span></p><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p><div><div><p class=MsoNormal>各位好：<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><p class=MsoNormal><span lang=EN-US><a href="https://drive.google.com/file/d/1bWgBWuuLtLUHiHnUEiC9j7Q32jYxior8/view?usp=sharing" target="_blank">https://drive.google.com/file/d/1bWgBWuuLtLUHiHnUEiC9j7Q32jYxior8/view?usp=sharing</a><o:p></o:p></span></p><div><p class=MsoNormal>以上是最新版的<span lang=EN-US>date-duration</span>模組<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal>目前的<span lang=EN-US>EM</span>成績如下：<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal>訓練集<span lang=EN-US> - 76.58<o:p></o:p></span></p></div><div><p class=MsoNormal>驗證集<span lang=EN-US> - 84.62<o:p></o:p></span></p></div><div><p class=MsoNormal>測試集<span lang=EN-US> - 77.42<o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal>註：部分<span lang=EN-US>rules</span>會利用到當前日期，而上述<span lang=EN-US>EM</span>是模擬初賽時的日期做測試<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal> 打包好的模組是直接使用<span lang=EN-US>datetime.today()</span>回傳的當前日期，因此<span lang=EN-US>EM</span>會稍微降低<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal>模組有以下三點改善：<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal>一、<span lang=EN-US>NN</span>模型的訓練採用較高的<span lang=EN-US>dropout(0.5)</span>與<span lang=EN-US>curriculum learning<o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US> step 1. </span>混和<span lang=EN-US>DRCD/Lee/Kaggle/ASR/DROP/FGC</span>的訓練集，預訓練一個<span lang=EN-US>epoch<o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US> step 2. </span>承上，訓練集拔掉<span lang=EN-US>ASR</span>與<span lang=EN-US>DROP</span>這兩個品質較差的<span lang=EN-US>dataset</span>，<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal> 在混合<span lang=EN-US>DRCD/Lee/Kaggle/FGC</span>的驗證集上訓練至<span lang=EN-US>F1 earlystop (patience = 0)<o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US> step 3. </span>單獨在<span lang=EN-US>FGC</span>訓練集上<span lang=EN-US>finetune</span>，直到<span lang=EN-US>FGC</span>驗證集的<span lang=EN-US>EM earlystop (patience = 5)<o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal> 以上方法雖然沒辦法讓<span lang=EN-US>single-span</span>成績再提高<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal> 但可以避免模型預測值被<span lang=EN-US>top-1</span>的答案給<span lang=EN-US>dominate</span>住<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal>二、承上<span lang=EN-US>NN</span>模型，把可行的<span lang=EN-US>rule</span>皆添加進規則庫中<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal>三、修正規則庫中的規則衝突與一些<span lang=EN-US>bug<o:p></o:p></span></p></div><div><p class=MsoNormal>（規則衝突無法修正時，採用能讓整體<span lang=EN-US>EM</span>較高者，<span lang=EN-US>e.g. </span>年齡取四捨五入會比無條件進位對更多題）<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal>我目前正在以同樣的手法，趕工<span lang=EN-US>arithmetic</span>模組<span lang=EN-US><o:p></o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p></div><div><p class=MsoNormal><span lang=EN-US>Best regards,<o:p></o:p></span></p></div><div><p class=MsoNormal>郭家銍<span lang=EN-US><o:p></o:p></span></p></div></div></div></body></html>