AI能写论文了!华人本科生发明AI论文生成器
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p26-sign.toutiaoimg.com/pgc-image/9496338c78e04dd0bb1aaacf32cd438f~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=%2B9scsqew8jYdmejEZzEttRrcmtc%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">【新智元导读】</strong>AI写论文达到了几近完善的程度!伦斯勒理工学院大四学生王清昀等<span style="color: black;">科研</span>人员最新<span style="color: black;">研发</span>PaperRobot,能够从产生点子、写摘要、写结论到写“<span style="color: black;">将来</span><span style="color: black;">科研</span>”,<span style="color: black;">乃至</span>它还能为你写出下一篇论文的题目。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">还在为写论文想不出好点子而发愁吗?</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">不消</span>愁了!伦斯勒理工学院、斯坦福大学等的<span style="color: black;">科研</span>人员最新<span style="color: black;">研发</span>的PaperRobot,<span style="color: black;">供给</span>从<strong style="color: blue;">产生idea、写摘要、写结论到写“<span style="color: black;">将来</span><span style="color: black;">科研</span>”的一站式服务</strong>!<span style="color: black;">乃至</span>它还能为你写出下一篇论文的题目,从此<strong style="color: blue;">论文无忧</strong>。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">这篇题为<strong style="color: blue;">PaperRobot: Incremental Draft Generation of Scientific Ideas</strong>的论文已被ACL 2019录取,<span style="color: black;">近期</span>在推特上引起<span style="color: black;">海量</span>关注。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/0c73c4881e5143b681b1c5089f256fc8~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=vXOn9FO3lSsLMWBXPn0RUExrWGE%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">谷歌大脑<span style="color: black;">专家</span>David Ha(hardmaru)<span style="color: black;">评估</span>:“<strong style="color: blue;">May a thousand (incremental) ideas bloom. </strong>”</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">大四华人一作发明AI「论文生成」神器</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">论文作者来自伦斯勒理工学院、DiDi实验室、伊利诺伊大学香槟分校、北卡罗来纳大学教堂山分校和斯坦福大学。其中,<span style="color: black;">第1</span>作者<strong style="color: blue;">Qingyun Wang (王清昀)</strong>是伦斯勒理工学院的大四本科生(今年8月<span style="color: black;">起始</span>讲进入UIUC读计算机科学PhD)。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">这不是王清昀<span style="color: black;">朋友</span><span style="color: black;">第1</span>次<span style="color: black;">科研</span>AI写论文,早在2017年他的“<strong style="color: blue;">论文摘要生成</strong>”<span style="color: black;">科研</span><span style="color: black;">亦</span>曾<span style="color: black;">导致</span><span style="color: black;">热榜</span>。王清昀<span style="color: black;">朋友</span>中学在杭州第二中学就读,从小<span style="color: black;">便是</span>“发明小达人”,取得专利的发明就有2个。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/b0a8302c81be4777b3bcf78148437dd2~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=A%2F7Y76DIR4P88tvAngCN7U2l6iA%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">论文地址:</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">https://arxiv.org/pdf/1905.07870.pdf</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">PaperRobot是<span style="color: black;">怎么样</span>自动写论文的呢?简单<span style="color: black;">来讲</span>,它从以前的论文中提取背景知识图谱,产生新的科学思想,最后写出论文的关键要素。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">它的工作流程<span style="color: black;">包含</span>:</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">(1)对<span style="color: black;">目的</span><span style="color: black;">行业</span>的<strong style="color: blue;"><span style="color: black;">海量</span>人类撰写的论文</strong>进行深入的理解,并<strong style="color: blue;">构建全面的背景知识图</strong>(knowledge graphs, KGs);</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">(2)<span style="color: black;">经过</span>结合从图<span style="color: black;">重视</span>力(graph attention)和上下文文本<span style="color: black;">重视</span>力(contextual text attention),<strong style="color: blue;">从背景知识库KG中预测</strong><strong style="color: blue;">链接</strong><strong style="color: blue;">,从而产生新想法</strong>;</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">(3)基于memory-attention网络,<strong style="color: blue;">逐步写出一篇新论文的<span style="color: black;">有些</span>关键要素</strong>:从输入标题和预测的<span style="color: black;">关联</span>实体,生成一篇<strong style="color: blue;">摘要</strong>;从摘要生成<strong style="color: blue;">结论</strong>和<span style="color: black;">将来</span><strong style="color: blue;">工作</strong>;最后从<span style="color: black;">将来</span>工作生成<strong style="color: blue;">下一篇论文的标题</strong>。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">科研</span>者对这个AI论文生产机进行了图灵测试:</p>PaperRobot生成生物医学<span style="color: black;">行业</span>论文的摘要、结论和<span style="color: black;">将来</span>工作部分,<span style="color: black;">同期</span>展示人类写作的同<span style="color: black;">行业</span>论文,<span style="color: black;">需求</span>一名生物医学<span style="color: black;">行业</span>的专家进行比较。结果<span style="color: black;">表示</span>,分别就摘要、结论和<span style="color: black;">将来</span>工作部分而言,在30%、24%和12%的<span style="color: black;">状况</span>下人类专家认为AI生成的比人类写作的更好。<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">至于这批AI<span style="color: black;">科研</span>人员<span style="color: black;">为何</span><span style="color: black;">选取</span>生物医学<span style="color: black;">行业</span>来做实验,<span style="color: black;">原由</span>很简单:生物医学论文<span style="color: black;">非常多</span>,非常多!<span style="color: black;">她们</span>尝试了用自己<span style="color: black;">行业</span>(NLP)来做实验,结果并<span style="color: black;">不睬</span>想(NLP的论文语料还<span style="color: black;">不足</span>多)。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">接下来,新智元对这篇论文进行了译介:</p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;">简单3步,图网络+<span style="color: black;">重视</span>力机制,AI写论文<span style="color: black;">乃至</span>比人类好</strong></h1>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">咱们</span>的<span style="color: black;">目的</span>是打造一个论文<span style="color: black;">设备</span>人PaperRobot,来加速科学<span style="color: black;">发掘</span>和生产,它的<span style="color: black;">重点</span>任务如下。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">阅读现有的论文。</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">论文太多了。<span style="color: black;">专家</span>们很难跟上井喷式的论文增长速度。例如,在生物医学<span style="color: black;">行业</span>,平均每年有超过50万篇论文被<span style="color: black;">发布</span>,仅2016年就有超过120万篇新论文<span style="color: black;">发布</span>,总论文数超过2600万篇(Van Noorden, 2014)。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">然而,人类的阅读能力几乎是不变的。2012年,美国<span style="color: black;">专家</span>估计,<span style="color: black;">她们</span>平均每年只能阅读264篇论文(5000篇论文中只读1篇),这个数字与<span style="color: black;">她们</span>在2005年进行的<span style="color: black;">一样</span>调查中报告的数据一致。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">PaperRobot自动阅读所有可用的论文,构建背景知识图(KG),其中节点<span style="color: black;">暗示</span>实体/概念,边<span style="color: black;">暗示</span>这些实体之间的关系。</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">在本<span style="color: black;">科研</span>中,<span style="color: black;">咱们</span>采用的是<span style="color: black;">海量</span>已<span style="color: black;">发布</span>的生物医学论文,提取实体及其关系来构建背景知识图。<span style="color: black;">咱们</span>应用了Wei等人(2013)中提出的实体和关系提取系统,提取了3类实体(<span style="color: black;">疾患</span>,化学和基因)。<span style="color: black;">而后</span>,<span style="color: black;">咱们</span>进一步将所有实体链接到CTD(比较遗传毒理学数据库),提取出133个子类型的关系,如标记/机制、治疗和<span style="color: black;">加强</span>表达。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">图3是一个示例。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/4a4ade3e4f43460680ecd41cad20bb44~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=lYLKWQzXLHmZgDqWQOJ6EvpbXeA%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">图3:生物医学知识提取与链接预测示例(虚线<span style="color: black;">暗示</span>预测的链接)</p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;">产生新的想法</strong></h1>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">科学<span style="color: black;">发掘</span><span style="color: black;">能够</span>看作是在知识图中创建新的节点或链接(links)。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">创建新节点<span style="color: black;">一般</span><span style="color: black;">寓意</span>着<span style="color: black;">经过</span>一系列真实的实验室实验<span style="color: black;">发掘</span>新的实体(如新的蛋白质),这对PaperRobot<span style="color: black;">来讲</span>可能太难了。<span style="color: black;">然则</span>,<span style="color: black;">运用</span>背景知识图<span style="color: black;">做为</span>起点,自动地创建新的边是<span style="color: black;">更易</span>的。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">Foster等人(2015)的<span style="color: black;">科研</span><span style="color: black;">显示</span>,640万篇生物医学和化学论文中,60%以上是增量式的工作。这启发<span style="color: black;">咱们</span><span style="color: black;">经过</span>预测背景知识图(KGs)中的新链接来自动地<span style="color: black;">增多</span>新想法和新假设。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">咱们</span>提出了一种新的实体<span style="color: black;">暗示</span><span style="color: black;">办法</span>,结合了KG结构和非结构化上下文文本来进行链接预测。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">如上面的图3所示,虚线<span style="color: black;">暗示</span>了预测的链接,<span style="color: black;">因为</span>钙和锌在上下文文本信息和图结构上都<span style="color: black;">类似</span>,<span style="color: black;">咱们</span>预测了钙的两个新邻居:CD14分子和神经纤毛蛋白2(neuropilin 2),它们是初始背景知识图中锌的邻居。</p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;">写一篇关于新想法的新论文</strong></h1>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">最后一步是把新想法清晰地传达给读者,这是一件非常困难的事情;事实上,许多<span style="color: black;">专家</span>都是糟糕的作家(Pinker, 2014)。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">运用</span>一个新颖的memory-attention网络架构,基于输入的标题和预测的<span style="color: black;">关联</span>实体,PaperRobot自动写出了一篇新论文的摘要,<span style="color: black;">而后</span>进一步写出了结论部分和<span style="color: black;">关联</span>工作部分,最后,为后续论文写了新标题。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">这个流程如图1所示。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/d4c39289c29540f89d4dae2cdcbfbcb6~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=yRStXBKyVsicpzuvHCxhKP%2BMhxU%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">图1: PaperRobot论文写作流程</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">咱们</span><span style="color: black;">选取</span>生物医学<span style="color: black;">做为</span><span style="color: black;">咱们</span>的<span style="color: black;">目的</span><span style="color: black;">行业</span>,<span style="color: black;">由于</span>这一<span style="color: black;">行业</span>有<span style="color: black;">海量</span>的可用论文。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">图灵测试<span style="color: black;">显示</span>,PaperRobot生成的输出内容有时比人工编写的内容更受欢迎;<span style="color: black;">况且</span>大<span style="color: black;">都数</span>论文摘要只<span style="color: black;">必须</span><span style="color: black;">行业</span>专家进行少量编辑,就<span style="color: black;">能够</span>变得信息丰富、条理清晰。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">让<span style="color: black;">咱们</span><span style="color: black;">瞧瞧</span>AI写的摘要:</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">Bac<span style="color: black;">公斤</span>round:</strong> <strong style="color: blue;">Snail</strong> is a multifunctional protein that plays an important role in the pathogenesis of<strong style="color: blue;"> prostate cancer</strong>. However, it has been shown to be associated with poor prognosis. The purpose of this study was to investigate the effect of negatively on the expression of<strong style="color: blue;">maspin</strong> in <strong style="color: blue;">human nasopharyngeal carcinoma</strong>cell lines. Methods: Quantitative real-time PCR and western blot analysis were used to determine whether the demethylating agent was investigated by quantitative<strong style="color: blue;">RT-PCR</strong> (qRT-PCR) and <strong style="color: blue;">Western blotting</strong>. Results showed that the binding protein plays a significant role in the regulation of <strong style="color: blue;">tumor</strong> growth and progression.</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">PaperRobot的整体框架如图2所示。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/086b738c12134fc5b50296d7338e96e7~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=vftWkJjjuptaD2YxwzmR5e7TCmA%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">表1<span style="color: black;">表示</span>了从<span style="color: black;">全部</span>过程生成的示例。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/b4abf5566c1744098e850174362e2e25~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=mYra2DfGaVf%2FrNyVVJZDq3Xd4Lc%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">表1:人类写的论文与AI系统写的论文的比较(粗体字<span style="color: black;">暗示</span>与主题<span style="color: black;">关联</span>的实体;斜体<span style="color: black;">暗示</span>人工编辑)</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">(关于<span style="color: black;">每一个</span><span style="color: black;">过程</span>的算法的<span style="color: black;">仔细</span>介绍,请阅读原始论文。)</p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;">实验过程及结果</strong> </h1>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">数据收集</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">咱们</span>从PMC开放存取子集中收集了生物医学论文。为人类书面论文引用一篇论文来构建新标题预测的ground truth,<span style="color: black;">咱们</span>假设论文A的标题是从论文B的“结论和<span style="color: black;">将来</span>工作”中生成的。<span style="color: black;">咱们</span>从1,687,060篇论文中构建了背景知识图,其中<span style="color: black;">包含</span>30,483个实体和875,698个关系。表2所示为<span style="color: black;">仔细</span>数据统计。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/4141a71674904fc5bfa8a7a742cf6fee~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=hEZS%2B500CPaVMBf35Y4p%2FNWnEaU%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">表2 论文写作统计结果</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">自动<span style="color: black;">评定</span></strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">以前的<span style="color: black;">关联</span><span style="color: black;">科研</span><span style="color: black;">显示</span>,自动<span style="color: black;">评定</span>长文本生成是一项重大挑战。在故事生成之后,<span style="color: black;">咱们</span><span style="color: black;">运用</span>METEOR来量度<span style="color: black;">文案</span>主题与给定标题的<span style="color: black;">关联</span>性,并<span style="color: black;">运用</span>困惑度(perplexity)来进一步<span style="color: black;">评定</span>语言模型的质量。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">咱们</span>的模型的困惑度评分是基于在PubMed上的论文(500,000篇题材,50,000篇摘要,50,000个结论和<span style="color: black;">将来</span>工作)中学习的语言模型评出的,这些论文在<span style="color: black;">咱们</span>的实验中<span style="color: black;">无</span>用于训练或测试。结果如表3所示。<span style="color: black;">咱们</span>的框架优于以前的所有<span style="color: black;">办法</span>。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/97376d860420405db4c5e80b8e24857a~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=a1iNALVOh4UfcoimhxC3Qp8fNY4%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">表3 对诊断任务论文写作的自动<span style="color: black;">评定</span>结果</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">图灵测试</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">由生物医学专家(非母语人士)和非专家(母语人士)对模型进行图灵测试。测试中<span style="color: black;">需求</span><span style="color: black;">每一个</span>人类对系统输出的字符串和人类创作的字符串,并选出质量更高的字符串。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/908915d04a68400a8e308a3c0af365f5~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=A1%2BwjebsEjGZe0SVzqfyFGtXAJ4%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">表4 对模型的图灵测试结果(%)。百分比<span style="color: black;">暗示</span>人类裁判<span style="color: black;">选取</span><span style="color: black;">咱们</span>的模型输出结果的频率。<span style="color: black;">倘若</span>输出字符串(如摘要)基于相同的输入字符串(如标题),输入<span style="color: black;">要求</span>标记为“相同”,否则标记为“<span style="color: black;">区别</span>”。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">能够</span>看到,在专家的<span style="color: black;">选取</span>中,PaperRobot生成的摘要入选率比人类撰写的摘要入选率最多高出30%,“结论和<span style="color: black;">将来</span>工作”部分最多高24%,新标题最多高出12%。<span style="color: black;">行业</span>内专家的表现并未<span style="color: black;">显著</span>优于非专家,<span style="color: black;">由于</span>这两类人倾向于关注<span style="color: black;">区别</span>方面:专家侧重于内容(实体,主题等),而非专家侧重于语言。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;">人类后期编辑</strong></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">为了<span style="color: black;">测绘</span>PaperRobot<span style="color: black;">做为</span>写作助手的有效性,<span style="color: black;">咱们</span>在<span style="color: black;">第1</span>次迭代中随机<span style="color: black;">选取</span>了系统生成的50篇论文摘要,并<span style="color: black;">需求</span><span style="color: black;">行业</span>内的专家对其进行编辑,直到专家认为编辑后摘要<span style="color: black;">拥有</span>足够的信息性和连贯性。 <span style="color: black;">而后</span>由BLEU,ROUGE和TER<span style="color: black;">经过</span>比较人类编辑前后的摘要质量给出评分,如表5所示。专家花了大约40分钟。完<span style="color: black;">成为了</span>50篇摘要的编辑。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/109aed79303f472a98d11799ec03c155~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=%2BPKA7Dgter79IyLPpykYY79wbDo%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">有些</span>后期编辑后的示例。<span style="color: black;">能够</span>看到大<span style="color: black;">都数</span>编辑内容都是形式上的变化。</p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;">华人本科生一作,发明小达人</strong></h1>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/989c9f5383814862a5940abef3ed496a~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=RDu6rJAG6vOg9KPWO9JemaqoiLc%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">论文一作Qingyun Wang (王清昀)是伦斯勒理工学院的大四本科生,主修计算机科学与数学双学位。今年8月<span style="color: black;">起始</span>他将在伊利诺伊大学厄巴纳香槟分校读博,主修计算机科学。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">王清昀对自然语言处理很感兴趣,专研自然语言生成、信息提取和对话系统,本科<span style="color: black;">时期</span>已<span style="color: black;">发布</span>多篇<span style="color: black;">关联</span>论文。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/99130d0649634bdbb0faf5a3001e8a27~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=2peA8DctU8wUINVJcWGBmRAhEFw%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">令人意外的是,王清昀简历中还列举了2项专利,分别是“遥控方便桌”和“家用废油制皂<span style="color: black;">安装</span>”,都是中学时期取得的,其中《遥控方便桌》<span style="color: black;">得到</span>第27届浙江省创新大赛一等奖。</p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/pgc-image/d8623ca12f994c3399bc3f51a1f7540f~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1723954949&x-signature=Jp8vD7F4U04x5zknJxJfu%2Fk3pRE%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">中学时期的王清昀<span style="color: black;">朋友</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">看来,王<span style="color: black;">朋友</span>从小<span style="color: black;">便是</span>发明达人啊。AI写论文机<span style="color: black;">不消</span>说<span style="color: black;">亦</span>是一大造福人类的好发明,期待王<span style="color: black;">朋友</span>继续改进。</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">参考链接:</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">https://arxiv.org/pdf/1905.07870.pdf</p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">http://www.hz2hs.net.cn/news/allinfo/1251.html</p>
</div>
“NB”(牛×的缩写,表示叹为观止) 回顾过去一年,是艰难的一年;展望未来,是辉煌的一年。 你的努力一定会被看见,相信自己,加油。
页:
[1]