WIf you have to prepare a text for publication, according to an old editor’s adage, check what happens if you simply delete the first and last sentence. In many cases, this alone makes the text better. Less is often more. Such maxims are valid precisely because people tend to do the opposite. Whether it’s a recipe or a specialist publication: Anyone who is asked to improve something is more likely to do so by adding – be it spices or footnotes – than by omitting.
A research group from Birmingham has now also demonstrated this phenomenon in language. In the journal “Cognitive Science” they diagnose an “addition bias” in English: Words with the meaning of a quantitative increase (to add, more) occur significantly more often than those of diminution (to subtract, less), and in expressions made up of pairs of opposites, the word increase always comes first – examples of which come to mind from other languages as well (plus minus, more or less). In addition, the semantic fields of verbs of change tend to overlap with those of increase (augment, amend) than of restricting. And the German word “supplement” also sounds like improvement, although one can also add nonsense.
The Birmingham word statistics are current because they were found not only in texts by human authors, but also in the products of large language models, in the examined case GPT-3. Could this be avoided with suitable training material for the learning machines? Hardly likely. Finally, language models should simulate human speech. You also have to be able to generate new sentences without new data or ideas if necessary – in other words, to talk more than there is to say.