


2021 is a very productive year for natural language processing (NLP) and machine learning (ML). Now it is time to count the papers in the field of NLP and ML last year.
MAREK REI, a researcher in machine learning and natural language processing from the University of Cambridge, summarized and analyzed classic papers in 2021 and summarized the statistics of ML and NLP publications in 2021. The major conferences and journals in the intelligence industry were analyzed, including ACL, EMNLP, NAACL, EACL, CoNLL, TACL, CL, NeurIPS, AAAI, ICLR, and ICML.
The analysis of the paper is completed using a series of automated tools, which may not be perfect and may contain some flaws and errors. For some reason, some authors started publishing their papers in obfuscated form to prevent any form of content duplication or automated content extraction, and these papers were excluded from the analysis process.
Now let’s take a look at the MAREK REI statistical results.
Based on academic conference statistics
The number of submissions to most conferences continues to rise and break records. ACL appears to be an exception, with AAAI almost leveling off and NeurIPS still growing steadily.
Based on institutional statistics
The leading research institution in the number of papers published in 2021 is undoubtedly Google ; Microsoft ranked second; CMU, Stanford University, Meta and MIT ranked closely behind, and Tsinghua University ranked seventh. Microsoft, CAS, Amazon, Tencent, Cambridge, Washington, and Alibaba stand out with a sizable proportion of papers at NLP conferences, while other top organizations seem to focus primarily on the ML field.
From the data of 2012-2021, Google published 2170 papers and ranked first, surpassing the 2013 papers published by Microsoft . CMU published 1,881 papers, ranking third.
Most institutions continue to increase their annual publication numbers. The number of papers published by Google used to grow linearly, and now this trend has eased, but it still publishes more papers than before; CMU had a plateau last year, but has made up for it this year; IBM seems to be the only company that publishes slightly more papers Declining institutions.
By author statistics
Next, let’s take a look at 2021 Researchers who publish the most papers per year. Sergey Levine (Assistant Professor of Electrical Engineering and Computer Science, University of California, Berkeley) published 42 papers, ranking first; Liu Tieyan (Microsoft), Zhou Jie (Tsinghua University), Mohit Bansal (University of North Carolina at Chapel Hill), Graham Neubig (CMU) also ranks relatively high in the number of papers published.
Throughout 2012-2021, the papers published by Sergey Levine ranked first. Last year he ranked sixth. This year It jumped to the first place; Yoshua Bengio (Montreal), Graham Neubig (CMU), Zhang Yue (Westlake University), Zhou Ming (Chief Scientist of Innovation Works), Ting Liu (Harbin Institute of Technology) and others also ranked relatively high in terms of the number of papers they published. .
Sergey Levine sets a new record by a considerable margin; Mohit Bansal’s number of papers also increases significantly, 2021 Published 31 papers in 2020, the same as Graham Neubig; Yoshua Bengio's number of papers decreased in 2020, but is now rising again.
Statistics of papers published as the first author
Researchers who publish the most papers are usually postdocs and supervisors. In contrast, people who publish more papers as first authors are usually people who do actual research.
Ramit Sawhney (Technical Director of Tower Research Capital) published 9 influential papers in 2021, Jason Wei (Google) and Tiago Pimentel (PhD student at Cambridge University) published respectively 6 influential papers were published.
From the 2012-2021 distribution, Ivan Vulić (University of Cambridge) and Zeyuan Allen-Zhu (Microsoft) are both first authors Published 24 influential papers, tied for first place; Yi Tay (Google) and Li Jiwei (Shannon Technology) ranked second, having published 23 and 22 influential papers as first authors respectively. papers on NeurIPS; Ilias Diakonikolas (University of Wisconsin-Madison) has published 15 NeurIPS papers as the first author.
Based on national statistics
Number of publications by country in 2021, United States The number of publications is the largest, with China and the UK ranking second and third respectively. In the United States and the United Kingdom, NeurIPS accounts for the largest proportion, while AAAI accounts for the largest proportion in China.
The vertical coordinates from top to bottom are 500, 1000, 1500, 2000, 2500, and so on
Almost all top-ranked countries continue to increase their number of publications and set new records in 2021. The increase was the largest for the United States, further extending its lead.
In the United States, Google, Microsoft and CMU once again lead the list in terms of number of publications.
##In China, Tsinghua University, Chinese Academy of Sciences and Peking University published the most papers in 2021.
Through visualization, these Organizations are clustered together primarily based on geographic proximity, with companies in the middle.
We can also draw drawings containing specific keys word proportion of papers and track changes in this proportion over time.
The word “neural” seems to be on a slight downward trend, although you can still find it in 80% of papers. At the same time, the proportions of "recurrent" and "convolutional" are also declining, and the word "transformer" appears in more than 30% of papers.
If you look at the word "adversarial" alone, we will find that it is very common in ICLR, and almost half of the papers mention it. The proportion of "adversarial" in ICML and NeurIPS seems to have peaked before, while AAAI has not. In the past few years, the term "transformer" has become very popular. It is particularly widely used in NLP papers, with over 50% of published papers containing it, and its popularity is steadily increasing across all ML conferences.
The above is the detailed content of 2021 ML and NLP academic statistics: Google ranks first, and reinforcement learning expert Sergey Levine tops the list. For more information, please follow other related articles on the PHP Chinese website!

译者 | 布加迪审校 | 孙淑娟目前,没有用于构建和管理机器学习(ML)应用程序的标准实践。机器学习项目组织得不好,缺乏可重复性,而且从长远来看容易彻底失败。因此,我们需要一套流程来帮助自己在整个机器学习生命周期中保持质量、可持续性、稳健性和成本管理。图1. 机器学习开发生命周期流程使用质量保证方法开发机器学习应用程序的跨行业标准流程(CRISP-ML(Q))是CRISP-DM的升级版,以确保机器学习产品的质量。CRISP-ML(Q)有六个单独的阶段:1. 业务和数据理解2. 数据准备3. 模型

人工智能(AI)在流行文化和政治分析中经常以两种极端的形式出现。它要么代表着人类智慧与科技实力相结合的未来主义乌托邦的关键,要么是迈向反乌托邦式机器崛起的第一步。学者、企业家、甚至活动家在应用人工智能应对气候变化时都采用了同样的二元思维。科技行业对人工智能在创建一个新的技术乌托邦中所扮演的角色的单一关注,掩盖了人工智能可能加剧环境退化的方式,通常是直接伤害边缘人群的方式。为了在应对气候变化的过程中充分利用人工智能技术,同时承认其大量消耗能源,引领人工智能潮流的科技公司需要探索人工智能对环境影响的

Wav2vec 2.0 [1],HuBERT [2] 和 WavLM [3] 等语音预训练模型,通过在多达上万小时的无标注语音数据(如 Libri-light )上的自监督学习,显著提升了自动语音识别(Automatic Speech Recognition, ASR),语音合成(Text-to-speech, TTS)和语音转换(Voice Conversation,VC)等语音下游任务的性能。然而这些模型都没有公开的中文版本,不便于应用在中文语音研究场景。 WenetSpeech [4] 是

条形统计图用“直条”呈现数据。条形统计图是用一个单位长度表示一定的数量,根据数量的多少画成长短不同的直条,然后把这些直条按一定的顺序排列起来;从条形统计图中很容易看出各种数量的多少。条形统计图分为:单式条形统计图和复式条形统计图,前者只表示1个项目的数据,后者可以同时表示多个项目的数据。

arXiv论文“Sim-to-Real Domain Adaptation for Lane Detection and Classification in Autonomous Driving“,2022年5月,加拿大滑铁卢大学的工作。虽然自主驾驶的监督检测和分类框架需要大型标注数据集,但光照真实模拟环境生成的合成数据推动的无监督域适应(UDA,Unsupervised Domain Adaptation)方法则是低成本、耗时更少的解决方案。本文提出对抗性鉴别和生成(adversarial d

数据通信中的信道传输速率单位是bps,它表示“位/秒”或“比特/秒”,即数据传输速率在数值上等于每秒钟传输构成数据代码的二进制比特数,也称“比特率”。比特率表示单位时间内传送比特的数目,用于衡量数字信息的传送速度;根据每帧图像存储时所占的比特数和传输比特率,可以计算数字图像信息传输的速度。

数据分析方法有4种,分别是:1、趋势分析,趋势分析一般用于核心指标的长期跟踪;2、象限分析,可依据数据的不同,将各个比较主体划分到四个象限中;3、对比分析,分为横向对比和纵向对比;4、交叉分析,主要作用就是从多个维度细分数据。

在日常开发中,对数据进行序列化和反序列化是常见的数据操作,Python提供了两个模块方便开发者实现数据的序列化操作,即 json 模块和 pickle 模块。这两个模块主要区别如下:json 是一个文本序列化格式,而 pickle 是一个二进制序列化格式;json 是我们可以直观阅读的,而 pickle 不可以;json 是可互操作的,在 Python 系统之外广泛使用,而 pickle 则是 Python 专用的;默认情况下,json 只能表示 Python 内置类型的子集,不能表示自定义的


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

Dreamweaver Mac version
Visual web development tools

MantisBT
Mantis is an easy-to-deploy web-based defect tracking tool designed to aid in product defect tracking. It requires PHP, MySQL and a web server. Check out our demo and hosting services.

Notepad++7.3.1
Easy-to-use and free code editor

SAP NetWeaver Server Adapter for Eclipse
Integrate Eclipse with SAP NetWeaver application server.

SublimeText3 Mac version
God-level code editing software (SublimeText3)
