In an attempt to simplify the process of summarising complex scientific research papers, researchers at the Allen Institute for Artificial Intelligence have released a new AI-based tool that summarises the text from scientific papers and present it in a few sentences.
Considering scientific research papers are complex to understand because of the language it is presented in, it becomes a challenge for many who are willing to work on the same or trying to be updated with scientific literature. And, that is why the researchers from Allen Institute for Artificial Intelligence came out with this new AI-based model — Semantic Scholar — that automatically generates a single-sentence summary using GPT-3 style techniques. This helps in locating the right paper and deciding whether to dedicate time to read that complex paper or not, stated by the official website.
Credit: Semantic Scholar // The NextWeb
Researchers released this new feature is available in beta for over 10 million papers in the computer science domain. According to the news media, currently, visitors will be able to see these TLDR (Too Long; Didn’t Read) summaries of the papers on Allen Institute’s Semantic Scholar search engine.
These AI-generated summaries will help the readers make quick, informed decisions about the relevancy of the scientific https://analyticsindiamag.com/arxiv-makes-all-its-research-papers-available-on-kaggle-to-boost-machine-learning-developments/paper and where to invest the time in further reading. These summaries explain the work in various contexts, such as sharing a paper on social media.
According to the news media, the AI collates the most important parts of the paper from the key sections of abstract, introduction, and conclusion to create the summary.
According to the research paper by the institute — to build this model, the researchers first pre-trained the model on the English language and then created a SciTLDR data set of 3200 documents, including over 5,400 summaries of computer science papers. Further, the model has been trained on more than 20,000 titles of research papers to reduce dependency on domain knowledge while writing a synopsis. Once the training was done, the model was able to summarise documents over 5,000 average words in the document with just 21 average words in summary, which created a compression ratio of 238.1.