[Most-ai-contest] Ensemble analysis of date-duration questions

[kysu]

>因為AMODE是date-duration的問題，計算時的單位常常與問題不一致

 

Do you mean that the desired math-formula will include a unit which is inconsistent with that is asked in the question? Or, an incorrect answer is generated due to involving two quantities with unmatched units? If it is the latter case, maybe we can add a checking-procedure to filter out those unmatched candidates. Could you please give an example with more details? Thanks.

 

KY

 

From: 郭家銍 [mailto:jerrykuo7727 at gmail.com] 
Sent: Wednesday, December 25, 2019 9:32 PM
To: Most-ai Contest <Most-ai-contest at iis.sinica.edu.tw>; kysu <kysu at iis.sinica.edu.tw>
Subject: [Most-ai-contest] Ensemble analysis of date-duration questions

 

蘇老師您好：

 

目前在date-duration上實驗了用三個不同的模型作ensemble

對於AMODE是date-duration的問題，EM會從42.0下降到39.1

對於ATYPE是date-duration的single-span問題，EM則從57.0上升到60.7

 

初步的分析，是因為ensemble會讓模型預測的答案與問題更加吻合

(e.g. question問「幾年」，則ensemble後的N-best預測幾乎都是「XX年」)

 

這對single-span問題有正面效益，但對AMODE是date-duration時卻有害

因為AMODE是date-duration的問題，計算時的單位常常與問題不一致

以問題問「幾年？」的例子來說，可能出現以下的計算形式：

1037年 + 20歲 = 1057年

24歲 - 18歲 = 6年

 

Respectfully yours,

郭家銍

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.iis.sinica.edu.tw/pipermail/most-ai-contest/attachments/20191225/ef6f556e/attachment.html>