Politeness theory, set out by Penelope Brown and Stephen C. Levinson in 1987, holds that politeness is produced by a small set of recognisable linguistic moves rather than by general warmth. A 2013 study scored more than 10,000 real requests by hand and found the moves hold up, with one result the theory had not anticipated: the politeness of the word please depends on where in the sentence it appears.<\/p>\n\n\n\n
Most advice about sounding polite in writing is advice about tone, which is not something you can point at. Politeness theory takes the opposite view, that politeness is assembled from specific devices that can be listed, counted and checked. In 2013 a team at Stanford and the Max Planck Institute tested that claim on a corpus of real requests, and published the results as A computational approach to politeness with application to social factors<\/a>.<\/p>\n\n\n\n Penelope Brown and Stephen C. Levinson published Politeness: some universals in language usage<\/em> in 1987, building on their 1978 account of politeness phenomena. Their central distinction is between positive politeness, which builds a good relationship with the person you are addressing, and negative politeness, which works by minimising, or appearing to minimise, the imposition you are placing on them.<\/p>\n\n\n\n The 2013 paper explains why requests are the natural place to test this. A request imposes on the person receiving it, and can leave them in social difficulty if they are unwilling or unable to comply, so requests draw heavily on negative politeness: being indirect, as in Would you mind, or apologising for the imposition, as in I am terribly sorry, but.<\/p>\n\n\n\n The authors assembled requests from two large online communities where members routinely ask things of each other: Wikipedia, where the requests concern editing and administration, and Stack Exchange, where they cover subjects from programming to gardening and cycling. They then labelled more than 10,000 utterances using Amazon Mechanical Turk, which they describe as the largest corpus with politeness annotations.<\/p>\n\n\n\n The design is narrow on purpose. They annotated only requests of exactly two sentences, where the second sentence is the request itself and ends with a question mark. That gives an annotator enough context to judge the request without reading a whole thread. Each marker was then extracted automatically, using the Stanford Dependency Parser on the sentence structure together with purpose built word lists.<\/p>\n\n\n\n Two reference points make the numbers readable. A politeness score of zero is what a random sample of requests scores, and 25 per cent is the share of a random sample that falls in the top quarter of politeness scores. Anything above those two figures is politer than average.<\/p>\n\n\n\n This is the finding the theory had not predicted, and it is the most useful thing on the page. Please in the middle of a sentence is polite, scoring 0.49 with 57 per cent of such requests in the top quartile, because it sits inside indirect constructions such as Could you please say more. Please at the start of a sentence is impolite, scoring 0.30 below zero with only 22 per cent in the top quartile, because it signals a direct instruction such as Please do not remove warnings, which can make the politeness marker itself look insincere.<\/p>\n\n\n\n The word you behaves the same way. Starting a sentence with it, as in You have reverted yourself, scored 0.30 below zero with 17 per cent in the top quartile. Mid sentence, you is usually part of the indirect forms the authors group with negative politeness, as in Would you or Could you.<\/p>\n\n\n\n The practical reading is that politeness in writing is positional, not just lexical. Moving one word changes how the sentence lands, which is why advice to add please is not reliably advice to sound politer.<\/p>\n\n\n\n Expressing gratitude was the strongest marker in their table, scoring 0.87 with 78 per cent of those requests in the top quartile. Deference followed at 0.78 and 70 per cent, then being indirect with a phrase such as By the way at 0.63 and 58 per cent, a counterfactual modal such as Could you or Would you at 0.47 and 52 per cent, and apologising at 0.36 and 53 per cent. A greeting scored 0.43 and 45 per cent.<\/p>\n\n\n\n At the other end, opening with a direct instruction scored 0.43 below zero with 9 per cent in the top quartile, and asserting a fact against the other person, as in In fact you did link it, scored 0.38 below zero with 13 per cent. A bare direct question scored 0.27 below zero. All of these are measured against that zero and 25 per cent baseline.<\/p>\n\n\n\n The authors then built a classifier from these features and report that it reaches close to human performance across both communities. The human reference points they publish are 86.72 per cent on the Wikipedia requests and 80.89 per cent on the Stack Exchange requests, against a random baseline of 50 per cent, which is worth holding on to: people themselves agree on the politeness of a written request about four times in five.<\/p>\n\n\n\n Applied at scale, the classifier produced the paper’s social result. Polite Wikipedia editors were more likely to reach high status through elections, but once elevated they became less polite, and on Stack Exchange users at the top of the reputation scale were less polite than those at the bottom.<\/p>\n\n\n\n The corpus is public requests in two online communities, not private conversation. Every annotated item is a two sentence request ending in a question mark, which excludes most of what people actually write. The labels come from crowd workers judging a request in isolation rather than from the person it was sent to. The authors call their analysis of variation by gender and community preliminary, and they note in a footnote that the human labelled data was too small to control for the role of the user in the Stack Exchange reputation result.<\/p>\n\n\n\n None of that undermines the strategy level findings, which is why this page reports them with their scores attached. It does mean the numbers describe how requests of a particular shape were read by strangers, and not how your message landed with the person who received it.<\/p>\n\n\n\n If politeness is positional, then a judgement about how you come across cannot be made from word choice alone, which is the same problem from a different angle as what your texts sound like to everyone else<\/a>. It is also why a baseline matters: as the Pollyanna hypothesis<\/a> shows for positivity, and as communication accommodation theory<\/a> shows for style, a single number means little without something to compare it against. The same gap opens up in the other direction when a sender guesses at how their own wording landed, which is the subject of why text messages get misunderstood<\/a>.<\/p>\n\n\n\n Portraitor reads one named participant in a conversation you choose. Its privacy policy describes the output as inferences about that participant, communication style among them, each carrying an estimate of confidence, and says that names, email addresses and telephone numbers are removed in your browser before anything is transmitted. What a chat analysis is honest about matters more than what it claims, which is the subject of what a chat analysis cannot tell you<\/a>. If you want that read done on a real thread, analysing a WhatsApp chat with AI<\/a> sets out what is involved, and Portraitor<\/a> shows the flow itself.<\/p>\n\n\n\n Every marker in the 2013 table is a property of wording, or of where a word sits in the sentence. Portraitor’s privacy policy says that names, email addresses and telephone numbers are removed on your device, in your browser or in the iPhone app, before anything is transmitted, and that only the masked text is sent. It states the consequence in its own words: the step removes who is speaking, and it does not remove what was discussed.<\/p>\n\n\n\n Those two facts together are why a positional reading is possible on an anonymised file at all. A please at the start of a sentence is still at the start of the sentence after a name has become a placeholder, and gratitude, deference, an apology and a counterfactual modal are all still where the writer put them. What the file gives up is the identity attached to a turn, not the shape of the turn.<\/p>\n\n\n\n What survives is the marker, not a verdict. The 2013 labels came from crowd workers reading a two sentence public request in isolation, and the privacy policy describes Portraitor’s own output as inferences about one participant, communication style among them, each carrying an estimate of confidence. Neither the paper nor the policy offers a reading of how a particular message landed with the person who received it.<\/p>\n\n\n\n Not on its own. In the 2013 study, please in the middle of a sentence scored 0.49 on politeness while please at the start of a sentence scored 0.30 below zero, against a baseline of zero for a random request. Position changed the effect of the identical word.<\/p>\n<\/details>\n\n\n\n Positive politeness builds a relationship with the person you are addressing, through gratitude, greetings or warm wording. Negative politeness reduces the imposition you are making on them, through indirectness, hedging or an apology. The distinction comes from Brown and Levinson in 1987.<\/p>\n<\/details>\n\n\n\n Penelope Brown and Stephen C. Levinson, in Politeness: some universals in language usage, published in 1987 and developed from their 1978 work on politeness phenomena. The 2013 computational study tested their framework on real request data and confirmed its broad outlines.<\/p>\n<\/details>\n\n\n\n On requests of the kind studied, reasonably well. The 2013 classifier came close to the human reference points the authors report, 86.72 per cent on Wikipedia requests and 80.89 per cent on Stack Exchange requests against a random baseline of 50 per cent. That is a judgement about one request in a public community, not about a person.<\/p>\n<\/details>\n\n\n\n No. The study measures how a request was read by strangers, not whether it was appropriate. Opening with a direct instruction scored lowest in their table, 0.43 below zero, but between people who know each other well directness often carries no cost at all.<\/p>\n<\/details>\n\n\n\nWhere the theory comes from<\/h2>\n\n\n\n
How politeness was actually measured<\/h2>\n\n\n\n
The same word, two opposite effects<\/h2>\n\n\n\n
Which moves actually scored highest<\/h2>\n\n\n\n
The limits the authors state<\/h2>\n\n\n\n
What it means for reading your own messages<\/h2>\n\n\n\n
Which of these markers survive an anonymised export<\/h2>\n\n\n\n
Frequently asked questions<\/h2>\n\n\n\n
Does saying please make a message politer?<\/summary>\n
What is the difference between positive and negative politeness?<\/summary>\n
Who came up with politeness theory?<\/summary>\n
Can software judge how polite a message is?<\/summary>\n
Is being direct always impolite?<\/summary>\n