Science communication is a topic I’ve touched on many times on this blog, with the sector having an ongoing battle in its efforts to communicate the work it does efficiently and effectively. In a recent paper, a team from MIT describe their efforts at automating the task of providing a plain-English summary of any given scientific paper.
Various methods have been tested to provide passable bodies of text for news reporting and so on, but none have ever really cracked the nut. The MIT team believe that they’ve done so via a system that modifies traditional neural networks so that they rely on vectors rotating in a multidimensional space, which is something known as rotational unit of memory (RUM).
The system attempts to represent each word in a body of text via a vector in multidimensional space. Subsequent words then move this director in a particular direction, and after all words have been moved, the final vector is translated back into a string of words.
“RUM helps neural networks to do two things very well,” the researchers explain. “It helps them to remember better, and it enables them to recall information more accurately.”
Machine translation
The approach had originally been developed for work with various physics related problems, but the team quickly realized it could also be applied to natural language processing. After testing it on the summation of research articles, the early results were promising enough to continue development.
The team tested their system against existing AI-based systems, and there was a clear difference in the text output from each system. For instance, the traditional systems tended to produce repetitive and technical summaries, whereas the RUM-based system produced a much more readable summary that avoided pointless repetition of phrases.
The team built upon this early success, with the tool advanced so that it can ingest entire papers rather than just the abstract, with accurate summaries of lengthy and technical texts achieved. It’s clear that a lot of work still needs to be done, especially to produce summaries that are understandable to various audiences whilst maintaining the genesis of the research itself, but it’s another sign of the progress being made in AI-based editing and content production tools.
