Last month, Springer Nature announced the publication of their first machine-generated book — an experimental proof of the efficacy and impacts of algorithmically curated scholarly resources. In the age of “robot reporters” and auto-generated novels, Springer intends to lead the way in seriously examining the value of machine learning to aid readers at all levels with comprehending vast volumes of academic literature. This publication — a book synthesizing 150 other Springer books that address the topic of lithium ion batteries — has the hefty goal of piloting the ability of such machine-learning technology to save us the time of reading dozens of resources in order to grasp a new topic.
This news has me wondering: Will students adequately learn complex concepts from these resources, with the necessary depth of understanding we have come to expect from many years of rigorous reading and study? Will lay-readers of such books sufficiently comprehend the relevant context in which new concepts fit? How do we avoid perpetuating human biases, power structures, and assumptions that inevitably become baked into both scientific research and the information technologies that support it?
I recently had the opportunity to address these questions with two leading forces that made this new machine-generated ebook possible: Henning Schoenenberger, Director Product Data & Metadata Management at Springer Nature, and Prof. Dr. Christian Chiarcos, Applied Computational Linguistics lab at Goethe University Frankfurt/Main.
What was the goal or inspiration for this project?
Henning Schoenenberger: Springer Nature is aiming to shape the future of book publishing and reading. Progress in natural language processing is advancing fast, and new technologies around Artificial Intelligence offer promising opportunities for generating scientific content automatically with the help of algorithms. Hence, we decided to develop and publish our first machine-generated research book. This prototype is designed for all interested audiences, such as researchers, master and PhD students, reviewers, academic writers and decision-makers in science education. By providing a structured excerpt from a potentially huge set of papers, it is supposed to deliver an overview of a specific subject area or topic, saving time and effort. With this prototype we would also like to initiate a public debate on the opportunities, implications and potential risks of machine-generated content in scholarly publishing. The risk is growing with the complexity of the technology. The more we continue to look into Deep Learning approaches the more we also have to refine the reviewing process by subject matter experts in order to ensure the accuracy of the content.
So, how would you characterize the problem you’re looking to solve with this machine-generated book?
Henning Schoenenberger: Our first machine-generated book proposes a potential solution to the problem of managing information overload efficiently. It allows for readers to gain an overview of a given field of research in a short amount of time, instead of reading through hundreds of published articles. At the same time, if needed, readers are always able to identify and click through to the underlying original source in order to dig deeper and further explore the subject. Instead of using results from search engines which may often be hard to qualify, readers can rely on qualified information published on Springer Nature’s content platform SpringerLink which stands up to scientific scrutiny. Machine-generated books, such as our prototype, can assist anyone who, for example, has to write a literature survey or requires a quick and focused start into a certain topic.
What metrics will be used to assess how well this book solves that problem? Are you looking at learning outcomes or other ways to measure researchers’ ability to consume / understand these topics more quickly or effectively?
Henning Schoenenberger: We are evaluating the success of this project through ongoing user research as well as analyzing the feedback from the research community. At the moment, our focus lies especially on finding the most useful parameter settings for the algorithm pipeline, to allow for an optimized consumption process. Parameters we use are for example, page counts, the number of clusters and sections, the length of summaries, the number of word features etc. Of course, the parameter settings vary from discipline to discipline, and are also dependent on the scope and size of a given topic. Our expectation so far as learning outcomes for readers is a pragmatic one: It should speed up the literature digestion process. At the same time it should support readers identifying the underlying original sources and click through and be able to further explore the subject where necessary.
What about peer review? How did you validate the end results were accurate and of sufficient quality for publication?
Henning Schoenenberger: We are aware that the quality of machine-generated content can only be as good as the underlying sources which have been used to curate it. Hence we have decided to only use peer-reviewed, robust research from our content platform SpringerLink for this prototype. Through referencing all source documents with hyperlinks, readers are at any time able to identify the underlying source. We decided not to manually polish or copy-edit any of the texts, as we want to highlight the current status and remaining boundaries of machine-generated content. Springer Nature editors and experts in the field of Chemistry supervised the iterative content creation process and provided guidance and feedback regarding the content output on a regular basis.
How did you avoid human bias built into the algorithm?
Christian Chiarcos: The current implementation consists of several components, each of which have their own characteristics. These components include, for example, the generation of the table of contents, respectively, the overall structuring of the book in chapters, sections, and so on. The underlying algorithm in this case is fully unsupervised: Based on a given set of parameters such as number of chapters and sections, and a particular similarity metric, similar papers are grouped together and clustered. From each cluster, the most representative publications are selected according to predefined parameters.
Is there a scalable business model in machine-generated book publishing?
Henning Schoenenberger: I do expect that machine-generated content will become a scalable business model at some point. However, as with many technological innovations, we also acknowledge that machine-generated research texts may become an entirely new kind of content with specific features not yet fully foreseeable. As a global publisher, it is our responsibility to take potential implications of machine-generated content into consideration and work on providing a framework for machine-generated research content. That being said, it would be highly presumptuous to claim we knew exactly where this journey would take us in the future.
Would you expect composition, editing, or other publishing costs to be remarkably more or less for machine-generated books production vs. traditional book publishing costs?
Henning Schoenenberger: At a first glance, you should expect cost reduction when it comes to producing machine-generated content. However, this topic is more complex than it might seem, and in fact we are just starting to explore this field. It might be the case that the review process and the iterative quality checks which have to be built into the publication cycle, eventually eat up the time that we gained through facilitating a faster content generation process. Therefore, it is too early for us to provide an informed answer to this question. As a global publisher, we are always looking for options to optimize our publication processes, to make them faster and more efficient and hence more valuable for our customers.
What has been the biggest surprise in this project so far? Has anything come out of this experiment that you didn’t expect?
Christian Chiarcos: When we started this project, we had a range of well-understood technologies at hand, but it was completely unclear how to evaluate a machine-generated book. In a machine learning context, people usually have a certain amount of manually created gold data to test whether the system performs as expected. For book generation, nothing similar is in place, and I doubt that we could find the resources to manually create, let’s say, ten 250-page overview books about lithium-ion batteries or any particular topic that we could possibly use as gold or training data. While we chose the technologies for the different modules according to our intuitions about the expectations of the audience, it was thus largely unclear how the researchers in this field, and the general public, would perceive such a prototype. So far, the reactions have actually been way more positive than I personally anticipated – or feared, if you will.
Whether the machine-generated book on lithium-ion batteries is a good read is very hard to judge. Most likely not, as its prose is certainly not better nor more readable than that of the original text – and much of the content is very technical. But people are attracted by the convenience of structuring and compressing a large body of literature into a single book, and the technology is mostly seen as a fruitful endeavor – despite some fears about the future of scientific authors, which I personally don’t see endangered by machine-generated publications of this kind any time soon, but rather supported by this type of technology.
Another thing that really surprised me was the granularity of the feedback we got from subject matter experts consulted during the generation process. This indicates that users would like to have a more direct interaction with the system in order to explore the effect of the parameters it provides, and possibly, to revise some of the decisions the system made with respect to chapter structuring, style or degree of compression. This may be a promising direction for future research in this fascinating field.
Researchers play a crucial role in the scholarly publishing ecosystem. With Springer Nature’s first machine-generated book, do you want to introduce a new book format, and does this mean that human authors can be made redundant?
Henning Schoenenberger: If the technology turns out to be reliable, we plan to increase the use and creation of machine-generated content. However, it is not our intention to disregard the high value of human-created content. Research articles and books written by researchers and authors will continue to play a crucial role in scientific publishing. Artificial Intelligence is not yet able to generate anything similar to a full-scope and meaningful research article. Algorithms still have a hard time rememberinf what was said three pages before – due to a lack of contextual understanding – and to build a storyline that appeals to readers, although the latest research in this field is quite promising.
We foresee that in future there will be a wide range of options to create content – from entirely human-created content, a variety of blended man-machine text generation to entirely machine-generated text.
What comes next in this project?
Henning Schoenenberger: For the first prototype, we decided to focus on a current chemistry topic. We are planning to publish prototypes in other subject areas as well, including the Humanities and Social Sciences, with special emphasis on an interdisciplinary approach, acknowledging how difficult it often is to keep an overview across the disciplines. The current implementation will be subject to ongoing refinement – based on user research and advances of the technology – and we will use the prototype on Lithium-Ion Batteries as a basis to explore further development of the product.