If you train your algorithm with whatever raw data, you would get whatever result. Even a model perfectly analysing the given situation becomes useless when not being adequately trained. In this specific case, the problem is clear: that tool was designed to deal with a different type of scenarios. Coming up with names for objects by training the program with many other names of equivalent objects makes perfect sense. Trying to figure out the best title for an article by analysing a big number of past titles