CMU researchers develop an automatic politeness engine for text-based communications

If you’re a typically terse communicator who could probably benefit from a little more civility in your everyday communications, a new Carnegie Mellon University research project could be the answer. A team at CMU created an automated way to improve the politeness of written requests and communications, which could have a number of potential applications — including eventually providing the basis for a sort of Grammarly, but designed for writing tone instead of adherence to grammar rules.

The politeness transfer engine that the CMU research team (including Language Technology Institutes PhD student Shrimai Prabhumoye, as well as master’s students Aman Madaan, Amrith Setlur and Tanmay Parekh) developed is based on similar style transfer mechanisms you may be more familiar with from photography AI projects, where software can apply the style of one photograph to any other. This project used a data set of half a million emails exchanged by Enron employees, which were made public as part of legal proceedings against the company resulting from its corruption and fraud scandals.

Despite the company’s wrongdoing, many of the emails exchanged between employees were — unsurprisingly, if you’ve ever worked in a large corporation — laden with common niceties and politely formed requests and responses. These proved a good basis from which to train a computational linguistics algorithm that could then be used to take either basic or impolite requests, like “Show me last month’s reports,” and turn that into something with a little more basic human kindness and decency, like “Could you please send me the reports from last month?”

It might seem like the task is relatively simple — append a “please” and “thank you” to any phase and you’re mostly there, right? In fact, the researchers say that it actually involved a lot more subtlety, because in fact when we are striving to be polite we do a lot more, like rephrase what’s actually an order to be a request, as in the example above.

The automated method that the CMU team developed works only in North American English, as employed in a formal (i.e. workplace) setting for now, and there will be lots of work required in order to localize it as regional and linguistic differences regarding what’s considered polite vary greatly. But even in its current form, it could provide a lot of benefit when used for automated customer service chatbots, for instance, or in autosuggesting text in an email client.

There’s clearly an interest in this from companies that make significant use of automated text suggestions — like Apple, which provided support for this research alongside the Air Force Research Laboratory, the Office of Naval Research, the National Science Foundation and Nvidia.