Discourse markers: what oh, well and so do to a turn

A discourse marker is a word like oh, well, so, actually or I mean that adds nothing to what a sentence states and instead tells the reader how to take it. Deborah Schiffrin named and systematised the class in a book published in 1987. Researchers still disagree about which words belong in it, and agree that what the members share is commenting on the message rather than adding to it.

Take the oh off the front of a message and every content word is still in place. The sentence reports the same thing it reported before. What moves is the instruction wrapped around it.

In a review of the literature published in 2010, Jean Fox Tree writes that a speaker can say oh to indicate that the immediately following words are not meant to be connected to the immediately preceding words, which would be the usual state of affairs in the absence of an oh. Two letters carry that whole job, and they are the two letters a writer would drop first.

Schiffrin put oh, well, so and eight others in one class

Deborah Schiffrin, of Georgetown University, published Discourse Markers with Cambridge University Press, whose page for the book gives 12 February 1987 as the print publication date. The publisher’s own description, a secondary account of what the book argues, names eleven items: the particles oh, well, now, then, you know and I mean, and the connectives so, because, and, but and or.

That list is why a bare and at the start of a message sits in the same class as an oh. Both are doing something to the turn rather than in it. A separate measure runs on the same raw material from the other direction and counts the function words two people share.

The same description says the book’s core is a comparative analysis of markers in conversational discourse Schiffrin collected during sociolinguistic fieldwork, and that the study concludes markers provide contextual coordinates which aid in the production and interpretation of coherent conversation at both local and global levels of organization. Contextual coordinates is the useful phrase: the marker is not part of the message, it is a position fix for it.

Eleven years later Andreas Jucker and Yael Ziv edited Discourse Markers: Descriptions and theory for John Benjamins, a volume the publisher describes as thirteen papers drawing on English, Finnish, Hebrew, Korean and Japanese. Fox Tree’s 2010 review credits it with settling the name of the class.

The lists never matched, and one thing every list agrees on

Fox Tree’s review appeared in Language and Linguistics Compass and is readable as her own manuscript on her university page. A fundamental characteristic of discourse markers, she writes, is that they operate beyond the propositional content of the communication, and there is widespread disagreement about what constitutes a discourse marker and what discourse markers do.

The review sets the positions of the 1980s side by side. Lawrence Schourup in 1985 focused on like, well and you know; Schiffrin in 1987 on her eleven; Britt Erman the same year on you know, you see and I mean. Well itself, the review notes, is an obvious marker to some and left out by others, naming David Wilkins among those who exclude it, and the question of what counts is still not resolved.

The functions attributed to markers contradict each other too. The review records that markers have been called turn-initiators and turn-relinquishers, grammatically superfluous and grammatically functional, equated with hedges and fillers and kept distinct from them. Some researchers concluded from that spread that markers are interchangeable or meaningless. Against that, the review cites a content analysis of folk definitions of um and uh, like, and you know, which found people did hold intuitive understandings that told the words apart.

What survives the argument is a negative definition. The review states that researchers do agree that discourse markers focus on the way communication is negotiated rather than on its content, an idea it says travels under the names metalanguage, procedural meaning, negotiating strategies and collateral signals.

The same words have been published under at least nine other labels. The review lists pragmatic expressions, pragmatic markers, pragmatic devices, pragmatic particles, discourse particles, discourse connectives, phatic connectives, ritualized speech and interjections. Searching any one of those names finds a fraction of the work.

Four kinds of marker, sorted by whether anyone is waiting

In a 2015 paper in Discourse Studies, also posted as her accepted manuscript, Fox Tree splits markers four ways: attitudinal, tailored, temporally-sensitive and cohesive. The split is built to predict something countable, namely the frequency with which discourse markers occur in spoken versus written formats.

Attitudinal markers carry a stance: the paper names actually, really and oh in its attitudinal form. Cohesive markers tell the reader how to relate what comes before the marker to what comes after, and the paper names well and some ohs, giving well as indicating that seemingly irrelevant following information is relevant after all. The tailored and temporally-sensitive kinds are like, you know and the fillers um and uh, whose work depends on a listener waiting while the sentence is still being assembled.

That dependency is the whole prediction. Attitude and cohesion are useful in any medium. Warning somebody that a delay is coming is only useful if somebody is standing there.

Corpus work had already given individual markers individual jobs. Karin Aijmer’s English Discourse Particles, published by John Benjamins in 2002, draws on the London-Lund Corpus, and its table of contents names chapters after the readings it argues for: now as the topic-changer, oh and ah as interjections, just as the interpersonal particle, sort of as the adjuster and actually as the expectation marker.

How readers scored a post that opened with oh

The first experiment in the 2015 paper took debate posts and asked readers what they made of them. The material came from a corpus of debates collected from 4forums.com, and the subset used was every response post beginning with a discourse marker. Seventeen turned up in that turn-initial slot, well most often at 605 posts, then so at 583 and and at 582, down to you think at seven. A sample of 2,576 posts with none of those markers in their first ten words served as an unmarked control, giving 6,462 items, each rated by between five and seven anonymous annotators recruited through Amazon Mechanical Turk.

Each rater moved four sliders running from minus 5 through zero to plus 5, for disagreement against agreement with the post being answered, attack against support, feelings against facts, and nasty against nice, then answered one yes or no question: was the respondent using sarcasm.

The sarcasm figures separate the markers cleanly. Readers called 15 percent of unmarked posts sarcastic. Posts opening with oh drew 35 percent and posts opening with really 26 percent, while posts opening with actually drew 11 percent, below the unmarked rate. Posts opening with well drew 17 percent, which the paper reports as no different from unmarked.

The sliders point the same way. Unmarked posts averaged minus 1.12 on agreement, while really averaged minus 2.37 and actually minus 1.85, both further towards disagreement. On feelings against facts, unmarked posts sat at plus 0.37, oh at minus 0.43 towards feelings and actually at plus 1.07 towards facts. The paper reports that oh preceded some of the most feeling-oriented, nasty and sarcasm-identified posts in the set, and that of all seventeen markers only I dunno preceded more feeling-oriented posts.

Well is the one that does none of this. The paper reports that well was neither more disagreeing, more attacking, more feeling-oriented, more nasty, nor more likely to be called sarcastic than unmarked posts, and that it was the only one of the four markers under test not linked to attitudinal or emotional judgements, exactly what the paper had predicted for a marker whose job is cohesion and nothing else.

This experiment compared different posts that happened to begin with different words. It did not take one post and remove its opening marker, so the figures show the company each marker keeps rather than the size of the change deleting it would make. The paper also notes that because items were picked by a computer program some may be ordinary uses rather than marker uses, and that you know was left out of this experiment for that reason.

The paper’s reading of actually is that in turn-initial position it mitigates face-threatening acts and is used for non-sarcastic disagreement, which puts a single adverb in the same territory as politeness theory and its account of face work.

In instant messaging most markers thin out, two do not

The second experiment compared the two modes directly. Students in two seminar courses, taught in 2006 and 2008, each collected five to ten minutes of spoken conversation among friends and five to ten minutes of instant messaging, giving 44 spoken samples and 44 written ones, averaging 994 words spoken and 580 written. Only marker uses were counted, so the like in I like that did not count.

Rates were expressed as a percentage of all words. Like fell from 2.1 percent of spoken words to 0.2 percent of written ones, the fillers um and uh from 0.9 percent to 0.1 percent, and you know from 0.4 percent to 0.1 percent. Well barely moved, from 0.4 percent to 0.3 percent, and neither did oh, from 0.7 percent to 0.5 percent. In the written dialogues oh was the most common of the five at half of 1 percent, followed by well.

The two that held their rate are the two the paper classes as cohesive. The three that collapsed are the tailored and temporally-sensitive ones, whose job depends on somebody waiting. The paper concludes from the pair of experiments that markers vary in frequency between spoken and written domains but are used similarly across them, with particular communicative functions that make them non-interchangeable.

The 2010 review puts overall rates on the same gap, citing other work rather than measuring it: about 1 percent of words in instant messaging were markers, rising to 2.5 percent if yeah is counted as one, against 4.8 percent in conversation on a comparable count, which it calls almost twice as high. The review attributes the difference to the relative costs of typing a discourse marker versus saying one, and to the minimal amount of private reviewing and revision time the text preparation window affords.

That is the part worth keeping. In speech a marker costs nothing and cannot be taken back. In a message box it costs a keystroke and survives a chance to delete it, and people type it anyway. The review names I mean as the marker that helps fix a mistake once it is made, one route by which a chat repairs itself, and its summary puts the central role of markers as helping people achieve grounding in unplanned, unrehearsed communication, the business of building shared understanding as a conversation goes along.

Masking has no placeholder for a word that names nobody

An export keeps the markers, because it keeps the characters that were typed. Portraitor’s home page lists WhatsApp, Telegram, Messenger and Instagram as the platforms it accepts and says the uploader takes .txt, .html, .json or .zip files.

What happens next is set out on the privacy policy, version 1.3, last updated 30 September 2026. Before any part of your conversation is transmitted, it says, your browser removes names, email addresses, telephone numbers, postal addresses and credentials and replaces them with anonymous placeholders, entirely on your own device. The policy then states its own limit: this removes who is speaking, and it does not remove what was discussed.

A bare well is none of the five categories that sentence names, and neither is so, actually or I mean. The masking step is defined by what it looks for, and a word with no referent gives it nothing to find. That is a statement about what the policy says masking does and nothing further: it makes no claim about discourse markers, and neither it nor the Portraitor home page says anything is counted or reported about them. For the rest of what masking does and does not reach there is a note on what masking misses, and for whose words get read at all, on picking yourself as the subject.

Why the marker cannot tell you what the sender meant by it

Every count above inherits a definition, and the definitions do not agree. The 2010 review says plainly that researchers do not agree on what counts as a discourse marker nor what functions markers serve in language production and comprehension. Change the list and the rate changes with it: its own instant messaging figure moves from about 1 percent to 2.5 percent on the single question of whether yeah belongs.

Telling a marker from an ordinary word is a judgement somebody makes, message by message. The 2015 paper’s rule was that the like in I like that was not counted, and in the debate experiment you know had to be dropped because the selection program could not separate the marker from the ordinary phrase. Any tally of markers in a chat log rests on where somebody drew that line.

The ratings also describe how readers took a post, not what its writer intended, and the paper’s conclusion is a claim about the class drawn across thousands of posts and 88 samples of talk and typing. It does not license a reading of one oh in one message.

The record holds the word and not the reason for it. A marker can be habit, or a pause for thought that got typed out, or precisely the signal the research describes, and the characters on the screen are identical in all three cases. Nothing in a log closes that gap.

Frequently asked questions

Frequently asked questions

A word with no content of its own can still be the part of a message that tells you how to take the rest of it.

Sources

1. Schiffrin, D. (1987). Discourse Markers. Cambridge University Press. Publisher page.
2. Fox Tree, J. E. (2010). Discourse markers across speakers and settings. Language and Linguistics Compass, 3(1), 1 to 13. Author's manuscript.
3. Fox Tree, J. E. (2015). Discourse markers in writing. Discourse Studies, 17(1), 64 to 82. Author's accepted manuscript.
4. Jucker, A. H. and Ziv, Y., eds. (1998). Discourse Markers: Descriptions and theory. John Benjamins. Publisher page.
5. Aijmer, K. (2002). English Discourse Particles: Evidence from a corpus. John Benjamins. Publisher page.
6. Portraitor privacy policy, version 1.3, last updated 30 September 2026.
7. Portraitor home page, read 1 November 2026.