A team of scientists has developed a technique that automatically makes written sentences more polite.
Why it matters: As the authors themselves note in the paper, it is "imperative to use the appropriate level of politeness for smooth communication in conversations." And what better to determine the appropriate level of politeness than an unfeeling machine-learning algorithm?
What's new: In a paper presented this week at the annual meeting of the Association for Computational Linguistics, researchers from Carnegie Mellon University analyzed a dataset of 1.39 million sentences, each of which was labeled with a politeness score.
- Using what is called a "tag and generate" approach, those sentences labeled as impolite were tagged, and then new text was generated to make the phrase nicer.
- So a phrase like "send me the data" would be automatically restructured to read "could you please send me the data?"
- Sometimes the system veered into outright editorializing, changing a sentence like "their chips are ok" to "their chips are great," which I must say may be more polite but could inadvertently end up sending someone to a mediocre Mexican restaurant.
Of note: The researchers used the "Enron Corpus" as a dataset — hundreds of thousands of emails exchanged by Enron employees and preserved by the federal government during its investigation of the now-defunct energy firm.
- A copy of the Enron Corpus was purchased by a computer scientist and released to researchers, who have since used it for numerous machine-learning studies.
- But really, why not train an AI for social niceties on the internal messages of a company that carried out one of the greatest frauds in American history?
The bottom line: This paper is really, really fascinating, and I am absolutely not just being polite.
Go deeper: Rooting out AI bias
