Look at the paper published by this project team in AC2016, which is actually a question of sentence selection. The main job is to select important sentences from the text live broadcast of sports events and combine them into a news report of the game.

Experimental data set: 150 live text commentary scripts of football matches)+150 official news reports of football matches.

Data characteristics: In the live text section, there are 242 sentences in a single game, accounting for 4,590 words; Regular news section, with an average of 32 sentences per field, 1 185 words.

Experimental process: This problem is a supervised learning process. The real-time text part is used as the input of training set and test set. The input feature vector X is a number of dimensions set manually, which is used for vector representation of each sentence, and the label Y is generated by regular news. Therefore, a learning ranking model is trained, which can generate a score for each sentence in the live text broadcast to represent its importance, select a certain number of sentences according to the score, and then combine them into a news by removing redundancy.

Therefore, the phrase "the goddess of failure throws an olive branch to it" is indeed problematic. First of all, the live text broadcast was written by people, and people certainly wouldn't write such sentences. Secondly, since AI essentially writes news by selecting sentences from the live broadcast, of course, it will not include this sentence for no reason.

This is just based on the idea of this article. Maybe they have added some other methods in practical application, but at present, the so-called "artificial intelligence replacing editors or journalists" is all scared by friends in the press. After all, the trained intelligence is only the secondary processing of existing materials.