点击上方蓝字，关注我们！威斯康星大学麦迪逊分校的Dane Morgan和Maciej Polak近日发表了他们训练ChatGPT阅读学术文章、制作关键数据并检查结果准确性的研究成果，这有助于节省科研工作者宝贵的研究时间。
人工智能开发商OpenAI承诺，将通过名为ChatGPT的新型聊天机器人重塑人们的工作和学习方式。事实上，在威斯康星大学麦迪逊分校，这种大型语言模型已经在帮助材料工程师们快速、经济、高效地从科学文献中提取信息。
近年来，威斯康星大学麦迪逊分校材料科学与工程Dane Morgan教授在他的实验室中使用了一种基于机器学习算法的人工智能来评估和寻找新材料，并取得了巨大成功。与Dane Morgan密切合作的科学家Maciej Polak对人工智能可能帮助完成的其他任务进行了头脑风暴。
Polak谈道：“人工智能可以越来越多地帮助我们完成相当复杂和耗时的任务。我在想，什么是材料学家经常做的、我们希望有更多时间做的事情?最先想到的就是阅读论文获取数据。”
Polak说，材料学家经常下载，然后梳理冗长的研究论文，在其中寻找一小组数字添加到他们的数据中去。
Polak表示：“我认为我们可以把所有这些耗时的任务都交给人工智能，它可以为我们阅读这些论文，并为我们提供关键信息。”
人工智能模型，即使像ChatGPT这样强大的聊天机器人，仅仅从论文全文中寻找和提取数据，仍然超出了它们的能力范围。因此，Polak对这项技术进行了改进，要求机器人一个句子一个句子地检查，并决定每个句子是否包含相关数据，也可以说它的任务是将论文浓缩为一两个关键句子。然后，他要求人工智能以表格的形式呈现信息，最后科研人员去检查表格和句子，以确保它们是正确和相关的。这项技术可以达到大约90%的准确率，这使得研究人员能够从一组论文中快速提取数据，以创建一个关于金属玻璃临界冷却速率的数据库。
虽然这项技术已经使研究人员的论文阅读工作量减少了约99%，但Polak对改进这项技术更感兴趣。
他解释道：“最后我仍然还需要手动执行检查表格准确性。所以，我想找到一种可以完全自动化整个过程的方法。”
为了达到这一目的，团队实施了“prompt”工程（找出准确的问题和顺序），这些问题和顺序将引导人工智能提取出他们想要的信息，然后再次检查。他们应用最初的方法提取数据表，然后通过向人工智能提出一系列后续问题，以引入数据集错误的可能性。这迫使人工智能重新检查数据并标记错误。在绝大多数情况下，人工智能是能够识别错误信息的。
Polak讲道：“它可以承认自己犯了错误，这是最重要的事情，或许它不知道如何解决这个问题，但至少我们没有得到与事实不符的信息。”
Morgan认为，这种使用ChatGPT和其他大型语言模型的“prompt”工程一开始感觉就很不寻常。
Morgan解释道：“这不是传统意义上的编程，与这些机器人互动的方法是通过语言，让程序提取数据，然后让它检查它是否确定正常的句子，感觉更像是我训练孩子们得到正确答案的方式，而不是我通常训练计算机的方式，是一种让电脑做事的不同方式，它真的改变了我们对电脑功能的看法。”
Polak谈道：“以前，人们必须编写数百行代码来做这样的事情，而且结果往往不如人意。但现在我们有了像ChatGPT这样的工具，在功能上有了巨大进步。”
Morgan很快指出，将人工智能融入研究并不能取代研究生和科学家。相反，这些工具可以让科研人员从事他们以前没有时间、人力以及资金进行的项目。
Morgan表示：“我认为这些工具将改变我们进的科研的方式，就像当初谷歌改变了我们做研究的方式一样。今天，我们都是通过谷歌和其他搜索工具来探索某个领域，帮助我们查找论文和相关资源，然后我们阅读这些论文和资源提取信息及数据。现在，你可以使用这些大型语言模型之一来收集有关某个主题的信息，并使用我们一直在开发的技术，就可以在几个小时内建立一个数据库进行查阅。”
ChatGPT makes materials research much more efficient
by Jason Daley,University of Wisconsin-Madison，APRIL 20, 2023
UW–Madison's Dane Morgan and Maciej Polak have published their solution for training ChatGPT to read academic articles, tabulate key data and check the results for accuracy, thereby saving valuable research time.
The artificial intelligence developer OpenAI promises to reshape the way people work and learn with its new chatbot called ChatGPT. At the University of Wisconsin–Madison, in fact, the large language model is already aiding materials engineers, who are harnessing its power to quickly and cost-effectively extract information from scientific literature.
For several years, Dane Morgan, a professor of materials science and engineering at UW–Madison, has usedmachine learning, a type of data-based AI, in his lab to evaluate and search for new types of materials with great success. Maciej Polak, a staff scientist who works closely with Morgan, brainstormed other tasks AI might help with.
"AI can increasingly help with tasks that are quite complex and time consuming," says Polak. "And we thought, 'What is something materials scientists do very often that we wish we had more time for?' One key thing is reading papers to get data."
Polak says materials scientists often download and then comb through long research papers to search for one small group of numbers to add to theirdata sets.
"We thought we could just offload all of these time-consuming tasks onto an AI that could read those papers for us and give us that information," says Polak.
Asking chatbots, even powerful ones like ChatGPT, to simply look for and extract data from the full text of a paper remains beyond their capabilities. So Polak refined the technique, asking the bots to review sentence by sentence and decide whether each contained relevant data or not—a task that boiled papers down to one or two key sentences. He then asked the bots to present the information in a table form, at which point a human researcher could review the table and sentences to make sure they were correct and relevant. The technique yielded an accuracy rate of about 90%, allowing the researchers to extract data from a set of papers to create a database on the critical cooling rates for metallic glasses.
In February 2023, Polak, Morgan and colleagues posted a paper about the technique on thearXiv preprint server.
While the technique reduced the researchers' paper-reading workload by about 99%, Polak was interested in improving it even more.
"I was the person still doing that last, manual step—checking the accuracy of the tables," he says. "So, I wanted to find a way to fully automate this process."
To get to that point, the team engaged in "prompt" engineering—figuring out the exact questions and sequence that would cause the bot to extract and then double-check the information they wanted. They applied their initial approach to extracted the data table, and then they asked the bot a series of follow-up questions to introduce the possibility that the data set was wrong. That forced the AI to double back, recheck the data and flag mistakes. In the vast majority of cases, the AI was able to identify faulty information.
"That's the most important thing; it can admit it made a mistake," says Polak. "Maybe it doesn't know how to fix it, but at least we're not getting factually incorrect information."
The team released a separate paper on this iteration of the technique onarXiv in March 2023.
Morgan says this type of prompt engineering with ChatGPT and other large language models feels unusual at first.
"This isn't programming in the traditional sense; the method of interacting with these bots is through language," Morgan says. "Asking the program to extract data and then asking it to check if it is sure with normal sentences feels closer to how I train my children to get correct answers than how I usually train computers. It's such a different way to ask a computer to do things. It really changes how you think about what your computer can do."
Importantly, the technique doesn't require a lot of effort ordeep knowledge, according to Polak.
"Previously, people had to write hundreds of lines of code to do something like this, and the results often weren't great," Polak says. "Now we have this huge improvement in capabilities with tools like ChatGPT."
Morgan is quick to note that integrating AI into research does not replace graduate students and scientists. Instead, these tools could allow researchers to pursue projects they previously didn't have the time, money or people-power to undertake.
"I think these tools will change the way we do research, analogously to how Google changed the way we did research," Morgan says. "Today we typically explore a field by using Google and othersearch tools to help us find papers and related resources, and then we read those papers and resources to extract information and data. Now you can go to one of these large language models to collect information around a topic and, using techniques like those we've been developing, build a database for review within hours."
More information
点击文首"新材料国际情报中心"或长按下方二维码关注本公众号, 浏览更多内容。
网友评论