# 碩論方向(⊙_⊙;) ###### tags: `ಥ_ಥ` https://2021.aclweb.org/program/accept/ https://zhuanlan.zhihu.com/p/385196031 https://dhx20150812.github.io/  善用PNG圖庫做簡報 重要的是解決了甚麼問題 ## 翻譯 0 ## Self-Supervised Learning (SSL) ## huggingface https://arxiv.org/pdf/1909.08053.pdf https://arxiv.org/pdf/2010.01057.pdf SSL其實算是近幾年迸出來的新名詞,定義上回顧一下,unsupervised代表使用沒有標註的資料,supervised代表使用了有標註的資料、而semi-supervised 代表同時使用了有標註與沒有標註的資料。而self-supervised翻成中文大概是自監督學習,也就是沒有標註資料也會自己會學習的方法。這似乎越搞越糊塗了,感覺跟unsupervised learning有87%相似。 ## Contrastive Learning (大陸很火) Contrastive learning是self-supervised learning中非常naive的想法之一。像小孩子學習一樣,透過比較貓狗的同類相同之處與異類不同之處,在即使在不知道什麼是貓、什麼是狗的情況下 (甚至沒有語言定義的情況),也可以學會分辨貓狗。 通過將數據分別與正例樣本和負例樣本在特徵空間進行對比,來學習樣本的特徵表示。Contrastive Methods主要的難點在於如何構造正負樣本。 https://zhuanlan.zhihu.com/p/141172794 http://mdeditor.tw/pl/gXpO/zh-hk https://cloud.tencent.com/developer/article/1865713 ### papers https://blog.csdn.net/dongguanting/article/details/118857820?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522162702007316780264047772%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request_id=162702007316780264047772&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~all~first_rank_v2~rank_v29-3-118857820.pc_search_result_cache&utm_term=ACL+2021+%E7%AD%86%E8%A8%98&spm=1018.2226.3001.4187 _**SimCLS: A Simple Framework for Contrastive Learning of Abstractive Summarization**_(NICE 太少頁QQ 用bart來生成 用roberta來scoring) _**Contrastive Learning for Many-to-many Multilingual Neural Machine Translation**_ ->本文的先驗假設是,如果兩句話說的是同一個意思,即使它們使用的語言不相同,那麼它們在語義空間中的表示也應該接近。所以本文的訓練目標是:文中使用了fancy的數據增強,同時使用單語和多語的數據進行對比。 ## 實體識別(NER) mRASP2: https://zhuanlan.zhihu.com/p/379328100 https://arxiv.org/pdf/2106.08977.pdf ## Adversarial Training(對抗式) ==================== 對抗例,是一種刻意製造的、讓機器學習模型判斷錯誤的輸入資 https://arxiv.org/abs/1909.11764 ->找出一個新的訓練方式 ## 用生成的方式產生dataset來讓效果改良 ### paper list #### open domain qa https://arxiv.org/pdf/2101.00288.pdf https://www.youtube.com/watch?v=ghbzVe6QLRI&ab_channel=HenryAILabs ->https://arxiv.org/pdf/2109.01156.pdf ->提到問題通常都太制式化 ,那我們就可以透過生出更多問題來增強效果(可當碩論) #### IR: https://arxiv.org/pdf/2011.13137.pdf https://arxiv.org/pdf/2105.03599.pdf (DPR的延伸 https://arxiv.org/pdf/2009.08553.pdf(根據標題去找答案的線索) Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks(https://zhuanlan.zhihu.com/p/339942960 https://arxiv.org/pdf/2009.08553.pdf #### 中文糾錯 PLOME: Pre-training with Misspelled Knowledge for Chinese Spelling Correction** #### 生成 統整的github: https://github.com/teacherpeterpan/Question-Generation-Paper-List challenges in generalization in open domain question answering 個性化生成:https://arxiv.org/pdf/2106.06169.pdf (BOB bert後面接一個bert EASY https://arxiv.org/pdf/2101.00288.pdf (Polyjuice 生成反例來提升成績 原理:針對各種風格finetune不同的dataset)被講了 https://arxiv.org/pdf/2106.00791.pdf (DYPLOC https://arxiv.org/pdf/2107.00152.pdf (利用語意圖表來進行對多句話時產生question 不錯) https://arxiv.org/pdf/2105.03023.pdf (DEXPERTS #### 非監督式風格轉換 https://arxiv.org/pdf/2105.08206.pdf(先有一個分類器可以找出風格的部分 然後再用生成 但方法效果似乎不大 資料好而已) -------------------- https://arxiv.org/pdf/2106.10502.pdf(input改成樹狀) https://arxiv.org/pdf/2105.11921.pdf #### paraphrasing https://aclanthology.org/P19-1605.pdf **ProtAugment: Intent Detection Meta-Learning through Unsupervised Diverse Paraphrasing** https://arxiv.org/pdf/2006.15020.pdf 透過zero shot翻譯成別的語言 https://arxiv.org/pdf/2106.07691.pdf https://aclanthology.org/2020.acl-main.535.pdf https://aclanthology.org/2021.findings-acl.135.pdf #### prompt https://blog.csdn.net/xixiaoyaoww/article/details/119363189?spm=1001.2014.3001.5501 https://www.youtube.com/watch?v=6RRdXnNd6XM&ab_channel=TechViz-TheDataScienceGuy https://zhuanlan.zhihu.com/p/386470305 #### other GOOGLE FINE ## 目前想法 QG or IR ACL進度: 命名實體識別的模塊化交互網絡 ## TO DO Prompting Contrastive Explanations for Commonsense Reasoning Tasks Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering CHALLENGE IN OPENDOMAIN QA
×
Sign in
Email
Password
Forgot password
or
By clicking below, you agree to our
terms of service
.
Sign in via Facebook
Sign in via Twitter
Sign in via GitHub
Sign in via Dropbox
Sign in with Wallet
Wallet (
)
Connect another wallet
New to HackMD?
Sign up