關於資料集分數重現差異 about w2ner

[Tomcathhh]

我也是，在使用源代码基础上复习Genia数据集得到的F1 为 80.87，与论文中有差异，请问是不是还用了什么trick？

[vicaasas]

我有注意到作者說可以調參，也試著調了幾個，如dilation [1,2,3] 改成 [1] 或是 不加 Biaffine 分數都沒有變好，不確定有沒有其他調參的做法，也想問一下 先生是否有找到其他作法 可以增高分數

[Tomcathhh]

我发现源代码使用的应该是biobert，但是使用这个预训练模型效果不如使用pubmedbert

[vicaasas]

作者還有一篇論文叫 TOE https://arxiv.org/pdf/2211.00684.pdf 不知道對先生有沒有幫助
只是分數一樣沒辦法重現，但"可能"會比 w2ner 高一點

[Tomcathhh]

谢谢你

[Tomcathhh]

你换成mutibiaffine试试 获取Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: vicaasas ***@***.***> Sent: Thursday, January 4, 2024 4:13:33 PM To: ljynlp/W2NER ***@***.***> Cc: LI ZHE ***@***.***>; Comment ***@***.***> Subject: Re: [ljynlp/W2NER] P於Y料集分�抵噩F差� (Issue #114) 我有注意到作者f可以{�苍�著{了��，如dilation [1,2,3] 改成 [1] 或是 不加 Biaffine 分�刀�]有�好，不_定有�]有其他{�⒌淖龇ǎ蚕��一下 先生是否有找到其他作法 可以增高分� ― Reply to this email directly, view it on GitHub<#114 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A26EABSLOWKNYPPJJALVWP3YMZQC3AVCNFSM6AAAAABASVHUP6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNZWGY4DCNBYHA>. You are receiving this because you commented.Message ID: ***@***.***>

[vicaasas]

mutibiaffine 是?
有教學實作嗎

[Tomcathhh]

An Embarrassingly Easy but Strong Baseline for Nested Named Entity Recognition
这篇论文

[vicaasas]

感謝，先生是加了這個分數就重現了嗎

[Tomcathhh]

不是，这只是我的想法 获取Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: vicaasas ***@***.***> Sent: Monday, January 8, 2024 3:47:58 PM To: ljynlp/W2NER ***@***.***> Cc: LI ZHE ***@***.***>; Comment ***@***.***> Subject: Re: [ljynlp/W2NER] P於Y料集分�抵噩F差� (Issue #114) 感x，先生是加了@��分�稻椭噩F了� ― Reply to this email directly, view it on GitHub<#114 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A26EABTOUWNEY7DPXWKB243YNOQC5AVCNFSM6AAAAABASVHUP6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBQGUYTKMBUGA>. You are receiving this because you commented.Message ID: ***@***.***>

[Tomcathhh]

因为我也没能复现出分数 获取Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: vicaasas ***@***.***> Sent: Monday, January 8, 2024 3:47:58 PM To: ljynlp/W2NER ***@***.***> Cc: LI ZHE ***@***.***>; Comment ***@***.***> Subject: Re: [ljynlp/W2NER] P於Y料集分�抵噩F差� (Issue #114) 感x，先生是加了@��分�稻椭噩F了� ― Reply to this email directly, view it on GitHub<#114 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A26EABTOUWNEY7DPXWKB243YNOQC5AVCNFSM6AAAAABASVHUP6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBQGUYTKMBUGA>. You are receiving this because you commented.Message ID: ***@***.***>

[vicaasas]

先生給的那篇論文，我大致看了一下....，先生會覺得需要重現分數才能投上期刊嗎

[Tomcathhh]

我觉得肯定是，所以我现在很头疼，不然体现不出我们改的东西有增益 获取Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: vicaasas ***@***.***> Sent: Monday, January 8, 2024 4:07:07 PM To: ljynlp/W2NER ***@***.***> Cc: LI ZHE ***@***.***>; Comment ***@***.***> Subject: Re: [ljynlp/W2NER] P於Y料集分�抵噩F差� (Issue #114) 先生o的那篇�文，我大致看了一下....，先生�X得需要重F分�挡拍芡渡掀诳� ― Reply to this email directly, view it on GitHub<#114 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A26EABWTPHP3DLNEIADRJTTYNOSKXAVCNFSM6AAAAABASVHUP6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBQGUZTINZTGI>. You are receiving this because you commented.Message ID: ***@***.***>

[zhangyipeng1]

pubmedbert

您好，请问您用的pubmedbert是Hugging Face中的哪一个？NeuML的那个吗？还有就是您说的效果提升大概有多大呢？万分感谢！

[Tomcathhh]

我不太记得了，现在好像PubMedbert名字改了，你仔细看一下描述，里面有写的。在我的设备上大概比biobert有1%提升，你也可以试试别的比如scibert之类的