Artificial intelligence to make our emails more polite? Yes please, and thank you!


Email etiquette is a funny thing. With no universal rules regarding the use of emojis, lol's, exclamation marks, and sign offs, we're all winging it until we're told otherwise. However, at its most basic, our emails need to be polite, which is perhaps even more important a rule than remembering to say ‘Dear xxx' and ‘Speak soon'.

Despite the fact that most of us make a conscious effort to be well-mannered in our messages, unfortunately, smiles and friendly tones often get lost in text-based communications. Therefore, you might think that you're making a polite request, but there's a good chance that you may accidentally come across ominous, or worse, rude.

Luckily, we need fret no more. Researchers at Carnegie Mellon University have developed an automated computer system that harnesses artificial intelligence (AI) to change sentences and make them more polite. Think Grammarly, but more… Mannerly?

However, it's not just a matter of inputting pleases and thank yous. Rather, the engine makes messages a lot less direct, incorporating “us” and “we” where reasonable to indicate a shared burden.

For example, rather than saying “Yes, delete it”, the message would change to the effect of “Yes, let's delete it” or “Yes, we can delete it.”

The CMU team used a perhaps unexpected, albeit relevant, dataset: half a million emails exchanged by employees at Enron, which were made public following the company's high-profile accounting fraud scandal in 2001. In turn, CMU had a sufficient dataset at the heart of its engine, enabling the team to move onto the next task: defining politeness. Given the wide scope of how politeness differs between cultures, the CMU researchers restricted their work to North American English only.

The team also had to determine the frequency and distribution of words in polite and impolite sentences. In doing so, they were able to develop a “tag and generate” pipeline to perform politeness transfers. Firstly, it tags impolite or neutral words, before the text generator replaces them with a more well-mannered counterpart – all without changing the meaning of the sentence.

However, this initially meant that the system was relying heavily adding polite words to sentences to make them more well-mannered. For example, “please help me” was considered polite, but “please, please, please help me” even more so.

After some ironing out, the scoring system became more realistic, and the changes became subtler. Since then, the team has released its dataset for use by other researchers to further the study of politeness.

Of course, this is a great stride in the AI and language technology arenas, but we have a teeny tiny point to make: let's hope that outsourcing politeness doesn't become some kind of norm. As the saying goes, “it's nice to be nice,” and robots shouldn't have to take care of that on our behalf! However, if it becomes a widely used tool, it'll be a great way to teach us about polite language use in any setting and will help to reinforce courteousness.