Since we started Zeta Alpha, using AI to help people make sense of the contents of very large document collections has been one of our goals. With more than 2300 papers, the ICLR conference is a very nice prototyping ground. In the last few months, our research intern Pavlos Zakkas has been working very hard on multi-document summarization. Now that ICLR 2023 is around, and we have published our manually curated conference guide (here), we're happy to unveil the first results of our work in this direction. Proudly presenting the world's first fully automated conference summary.
Image by DALL-E: Robots creating an ICLR2023 summary.
To digest a conference, we crawl all documents and create neural vector representations. Then we first cluster all documents in the conference, select representative documents for each cluster based on bibliometric signals, and apply multiple stages of chatGPT based multi-document summarization. Add an introduction and an end using meta-data from the conference, and Bob's your uncle. A more detailed overview will be described in a separate blog post soon. Further improvements can for sure be made, but for now we are just too excited to present the first fully automated conference review, three days before the conference starts. Would you like us to personalize it for you or do the same for your document collection? Here it is. Enjoy Discovery!
While this is the first fully automated conference summary, our ICLR coverage has just started! To keep up with the latest trends in AI, make sure to follow us on Twitter @zetavector to stay up to date with everything that's happening there!