Science Automated Writing: True or False?
Automated Science Writing: Myth or Nearest Future?
What do you think a scientific writer does? You might be surprised, but he reads a lot more than he writes. The task of this specialist is to study the complex scientific literature about new research and discoveries, and then to write about it in a simple, clear language. And make it so that the text was interesting to a wide audience. So, can machines do this now?
A group of scientists at the Massachusetts Institute of Technology (MIT) and others have developed a neural network – an artificial intelligence (AI) that can cope with this task. At least partially. He can read scientific articles and reproduce short conclusions (in English).
At this stage, artificial intelligence can`t replace professional, but it can be useful and facilitate the work of writers, editors, and scientists. For example, it can process large amounts of information, and thanks to its conclusions a person will be able to understand whether it is worth spending time on a detailed study of paper or there is nothing interesting there. Also, a new invention may be useful in speech recognition and machine translation.
Artificial Intelligence and Unexpected Discovery
The invention of the neural network for scientific writing was not the goal of the inventors. They worked on the development of new approaches in artificial intelligence for solving problems of physics. In the process, scientists have noticed that the result of their work can be used for other purposes. It turned out that the same technologies can solve such problems as natural language processing and other difficult computational problem.
Member of inventors’ team, a professor of physics at MIT, Marin Soljačić says that the main specialization of him and his colleagues is research on the problems of physics. For the past few years, they have worked with artificial intelligence and noticed some patterns. Namely, if they add a particular mathematical construct to the AI field, this affects the whole construct.
How does it work?
The work of the neural network can be represented as an attempt to simulate the process by which people learn new information. The system analyzes various examples and highlights the main patterns. A similar principle of operation is used for automated work with images, for example, recognition of objects in photos.
The main difficulty of the neural network in processing and comparing information from a long chain of data. And this is a key factor in working with texts. According to the researchers, they tried various techniques to solve the problem. Among other things, the methods known as long short-term memory (LSTM) and gated recurrent units (GRU) were used. But the result was still far from what is needed for real natural-language processing.
Then the scientists developed and introduced an alternative system. Without technical details, it can be called vector-based. The key element of the system is the rotational unit of memory (RUM).
Thus, the machine sees each word in the vector space in the format of a line of a certain length, pointing in a specific direction. Each next word affects the previous one by turning the vector. A certain meaning is encrypted in directions. Thus, after processing the text, the system transforms a set of vectors into words.
Scientists add that the RUM system helps to remember, that is, to save and recall information more accurately. So, in a conversation with Marin Soljačić, his colleague Mićo Tatalović noted that the capabilities of the system could help him in editorial work in the field of scientific journalism.
The team tried a few natural language processing tasks and saw a good result!
Testing Invention
As part of the study, scientists processed one using the LSTM-based neural network and through their RUM-based system. The results differed sharply in reproducing the summary. The RUM system prepared a more readable text and avoided unnecessary repetitions. Moreover, the system is significantly expanded and can handle not only abstracts but also significant research works.
Çağlar Gülçehre a researcher from the British artificial intelligence company Deepmind Technologies, who did not take part in this work, notes that not even an ideal result solves an important problem in neural networks. The central aspect is the solution with the linking of pieces of information that are separated in time and space.
For now, this invent doesn`t completely solve the problem, it shows promising results on the long-term dependency tasks such as question-answering, text summarization, and associative recall. But potentially this approach may influence a lot on the fields of natural language processing and learning.