Now an algorithm developed by Salesforce researchers seems to show how the computer could eventually take the trouble of document summaries. In order to do this, he uses a series of tricks in the subject of computerized learning which produces – surprisingly – pieces of precise and coherent texts. And while there is still a lot of work to do, it is clear that it is giving interesting clues about how this task could eventually be automated.
The Salesforce algorithm produces much better summaries than any other program developed for this task, according to an automated tool that measures the accuracy of what has been summarized.
According to Richard Socher, chief salesforce scientist, this scheme is a big improvement over earlier systems like never before. Socher, a specialist in computer learning and natural language processing, had a company called MetaMind, which was acquired by Salesforce in 2016.
Software still has a long way to go to match the human ability to capture the essence of a document, of a text. Summarizing a text perfectly might require genuine intelligence, including common sense and certainly a degree of mastery in a language.
The analysis of a language is still one of the great challenges of AI but it is a challenge that has a huge commercial potential. The power to condense information and extract the most relevant could be commercially beneficial without a doubt.
Caiming Xiong, a Salesforce scientist who contributed to this work, says the algorithm developed, though imperfect, could summarize news articles daily or give synopses of the emails a person receives. The latter could be useful even to the company Salesforce.
The algorithm uses a combination of approaches to achieve improvements in summarizing a note. The system learns from examples of good summaries, which is called supervised learning, but also uses a kind of artificial attention to text that is fed and output. This helps to not produce many repeated phrases, a common problem in algorithms that make abstracts.
