Challenge (Proof of concept)
In the media industry, the move towards automatic generation of news articles is accelerating. However, many existing technologies generate articles by embedding words and numerical values in prepared template statements, and we rely on human for design. For this reason it was a time-consuming process to apply to multiple fields.
In collaboration with business persons in Japanese media industry, NTT DATA has initiated a demonstration experiment on technology to automatically generate weather news documents with relatively simple patterns. We built a mechanism that could collect weather cloud data from the past 4 years released by the Meteorological Agency and weather news script actually read out by the announcer are set and learning is done by deep learning. As a result of evaluating weather news manuscripts generated from this method, Japanese grammar has reached a level that does not cause discomfort even when people read it, and some correction is necessary in the correctness of meaning, but we have confirmed that we can create documents with the same contents as weather telegrams in general.
Furthermore, we are planning to proceed with new demonstration experiments in fields that involve massive amounts of data, such as corporate financial results announcements and sports articles.