写在前面
大家好,我是刘聪NLP。
最近在做Prompt的相关,整理NAACL2022中Prompt相关的论文,并把论文中对应的Github链接也附上了。并且一些自己特别感兴趣的论文,已经标星”*“,希望对大家有所帮助。
后面也简单谈了谈对Prompt的看法,不是吐槽,不是吐槽,只是一些自己的真实感受。
一起学起来吧~~~
官方论文列表:https://2022.naacl.org/program/accepted_papers/
Prompt相关论文
[1] *Template-free Prompt Tuning for Few-shot NER
Github: https://github.com/rtmaww/EntLM/
[2]IDPG: An Instance-Dependent Prompt Generation Method
[3]Automatic Multi-Label Prompting: Simple and Interpretable Few-Shot Classification
Github: https://github.com/HanNight/AMuLaP
[4]ProQA: Structural Prompt-based Pre-training for Unified Question Answering
[5]On Transferability of Prompt Tuning for Natural Language Processing
Github: https://github.com/thunlp/Prompt-Transferability
[6]PROMPT WAYWARDNESS: The Curious Case of Discretized Interpretation of Continuous Prompts
Github: https://github.com/Alrope123/prompt-waywardness
[7]Learning to Transfer Prompts for Text Generation
Github: https://github.com/RUCAIBox/Transfer-Prompts-for-Text-Generation
[8]*Learning To Retrieve Prompts for In-Context Learning
Github: https://github.com/OhadRubin/EPR
[9]Do Prompt-Based Models Really Understand the Meaning of Their Prompts?
Github: https://github.com/awebson/prompt_semantics
[10]Go Back in Time: Generating Flashbacks in Stories with Event Plots and Temporal Prompts
Github: https://github.com/PlusLabNLP/flashback_gen
[11]Probing via Prompting and Pruning
[12]Contrastive Learning for Prompt-based Few-shot Language Learners
Github: https://github.com/yiren-jian/LM-SupCon
[13]Using Natural Sentence Prompts for Understanding Biases in Language Models
[14]Zero-Shot Event Detection Based on Ordered Contrastive Learning and Prompt-Based Prediction
[15]LiST: Lite Prompted Self-training Makes Efficient Few-shot Learners
Github: https://github.com/microsoft/LiST
[16]Exploring the Universal Vulnerability of Prompt-based Learning Paradigm
Github: https://github.com/leix28/prompt-universal-vulnerability
[17]Prompt Augmented Generative Replay via Supervised Contrastive Learning for Lifelong Intent Detection
[18]RGL: A Simple yet Effective Relation Graph Augmented Prompt-based Tuning Approach for Few-Shot Learning
[19]On Measuring Social Biases in Prompt-Based Learning
[20]Few-Shot Self-Rationalization with Natural Language Prompts
Github: https://github.com/allenai/feb
[21]SEQZERO: Few-shot Compositional Semantic Parsing with Sequential Prompts and Zero-shot Models
Github: https://github.com/amzn/SeqZero
[22]PromptGen: Automatically Generate Prompts using Generative Models
聊聊Prompt
说句实话,到目前为止,我还是「有些失望的」,和我想象的并不一样。在工业落地方面,我们总是希望模型具有解决小样本、快速训练、快速推理的能力。那么prompt解决了吗?其实并没有。
-
现在说的解决few-shot甚至zero-shot,基本上都是居于超大参数的模型,推理速度可想而知,加入我是机器问答场景,让用户等个几秒钟,是完全不可以接受的。
-
很多论文都是自动模板、自动答案空间映射,并且还是全量参数在训练,训练时长不减反增,真的是心累。有哪些时间,你说我标点数据香不香。
-
反正模型都很大,不是large就是xxlarge,或者一切软模版也是增加了部分参数,或多或少,反正推理时间是没变短。
-
在真实情况下,NER任务中每个类别标个50-100条数据,分类任务中每个类比标个200-300条,都可以达到一个不错的效果,对于公司来说,标这点数据还是可以接受的,那么在这种数据规模下,prompt又能提高多少呢?
-
在5-shot下,不用prompt是30,用了是40,确实提高了10个点,但是40的精读,模型能上线吗?
但是,我们依然要抱有希望,并且要冷静地看待新的技术。我并没有否定Prompt的价值,只是目前我现在很难在资源有限的情况下把它很好地落地,也许是自己的能力有限。
整理不易,请多多点赞,关注,有问题的朋友也欢迎加我微信「logCong」、公众号「NLP工作站」、知乎「刘聪NLP」私聊,交个朋友吧,一起学习,一起进步。
我们的口号是“生命不止,学习不停”。