“AI will be developed to summarize difficult papers in an easy-to-understand manner”


by

rawpixel.com

AI works just like a 'Science Writer,' who reads esoteric papers published in specialized journals and delivers them as sentences that can be easily understood by readers without a scientific background. Research at the Massachusetts Institute of Technology (MIT) Have announced.

Rotational Unit of Memory: A Novel Representation Unit for RNNs with Scalable Applications | Transactions of the Association for Computational Linguistics | MIT Press Journals
https://www.mitpressjournals.org/doi/full/10.1622/tacl_a_00258

Rotational Unit of Memory
http://super-ms.mit.edu/rum.html

Can science writing be automated? | MIT News
http://news.mit.edu/2019/can-science-writing-be-automated-ai-0418

The research team of Rumen Dangovski and Li Jing, who are MIT graduate students, and Marin Soljači 教授, a professor of physics, have developed an AI that summarizes specialized articles into one or two sentences. The sentences that AI can generate are so short that as of April 2019, it is impossible to create articles that real science writers will produce, but when reading a large number of articles it is checked lightly and the contents It may be useful to know in advance.

Originally, the research team seems to have tried an approach with AI for the purpose of dealing with physical problems, and it was said that it did not focus on processing such as abstracts of papers. However, the research team has realized that the approach developed by itself can be applied not only to the field of physics but also to other fields including natural language processing.


by

Mike MacKenzie

In general neural networks, computers progress by 'learning' patterns for so many cases. For example, neural networks are widely used in systems that identify objects in photos or extract specifics from photos and sounds.

On the other hand, it is difficult for neural networks to pick up and associate information from a long series of data. This ability is one of the techniques required for a work like a science writer to find and summarize the required information from a long dissertation. Techniques such as long-term short-term memory (LSTM) networks that model long-range dependencies have been used to solve this problem, but natural language processing has not been as practical.

The research team devised an alternative system based on rotating vectors in multidimensional space, not the system based on matrix multiplication used in traditional neural networks. The system is named ' rotational unit of memory ' and the research team calls it 'RUM'. RUM helps neural networks remember elements and is more effective in remembering elements more accurately. RUM was originally an approach designed to solve complex physical problems such as light behavior, but eventually the research team noticed that RUM could be useful in other fields such as natural language processing as well. Yes.

In natural language processing, RUM is to represent individual words appearing in sentences as vectors in multidimensional space. When a word in a sentence is a line having a specific length and a specific direction, the sentence is represented in a theoretical space having several thousand dimensions, and the final vector is output as a sentence The team says.



When the research team read a paper on a type of

roundworm that infects animals called ' Baylisascariasis ' in the summary AI using the LSTM network, and output the summary, the following sentences were generated. This summary is very repetitive and may not be of practical accuracy.

◆ Original text:

'Baylisascariasis, kills mice, has endangered the allegheny wood rat and has caused disease like disease or severe consequences. This infection, termed' baylisascariasis, has kills mice, has the endangered the allegheny woodat the back ' , termed “baylisascariasis,” kills mice, has endangered the allegheny woodrat.



If it is converted to Japanese, it will be as follows.

◆ Japanese translation:
'Baylisascariasis' has killed mice and endangered Allegheny Wood rats with blindness and serious consequences. This infection is called 'Baylisascariasis' and kills the rat, endangering the Allegheny Wood rat with blindness and serious consequences. This infection is called 'Baylisascariasis' and has killed mice and endangered Allegheny Wood rats.


On the other hand, when I read the same paper in the summary AI based on RUM, the following was output. This summary is easier to read and less repetitive than an AI summary using an LSTM network.

◆ Original text:

Over 90 percent of raccoons in Santa Barbara play host to this parasite. Urban raccoons may infect people more than previously suspected. 7 percent of surveyed individuals tested positive for raccoon roundworm antibodies.



If it is converted to Japanese, it will be as follows.

◆ Japanese translation:
Urban raccoons may infect humans more than previously thought. Of the people surveyed, 7% tested positive for raccoon ascarid antibodies. More than 90% of raccoons living in Santa Barbara are host to this parasite.


by Alias 0591

Çağlar Gülçehre, who conducts AI research at DeepMind , a British AI development company, points out that linking related elements in remote places in time and space in AI was a very fundamental and important issue. 'I think this research does not solve all the problems, but it shows promising results for tasks such as question and answer, text summarization, and association,' said Gülçehre.

in Software,   Science, Posted by log1h_ik