Language style matching: the small words two people share

Language style matching is a measure, defined by Ireland and Pennebaker in 2010, of how closely two people use the same function words, such as articles and personal pronouns. In a 2011 study led by Ireland, couples whose instant messages matched more closely were more often still dating three months later.

The idea is easy to state and odd to sit with: the measure looks only at small grammatical words such as articles and pronouns, and not at what the conversation is about. This piece goes through what the two founding papers report, using the numbers in their own abstracts, and where a chat export fits.

The measure ignores what was said

The 2011 paper, by Ireland, Slatcher, Eastwick, Scissors, Finkel and Pennebaker in Psychological Science, opens by saying that earlier relationship research had largely ignored “similarity in how people talk with one another.” Its definition is narrow on purpose: similarity in two people’s use of function words, called language style matching, or LSM.

The examples the authors give are articles and personal pronouns. Words like these appear whatever the subject is, so a message about dinner and a message about a job offer are compared on the same footing, and the measure never looks at dinner or the job.

Nine categories, any two texts

The working definition comes from the year before. Ireland and Pennebaker (2010) in the Journal of Personality and Social Psychology describe LSM as “the relative use of 9 function word categories (e.g., articles, personal pronouns) between any 2 texts.”

Two details in that sentence matter. It is relative: it describes the two texts against each other rather than either one alone. And it applies to any two texts, which is why the same measure could be run on essays, letters, poems, speed date transcripts and instant messages.

The text analysis program LIWC, whose site is published by Pennebaker Conglomerates, now offers this as a built in function. Its current site lists a function to “calculate Language Style Matching (LSM) for comparing texts with each other.”

Forty speed dates and eighty-six couples

The 2011 paper ran two studies. In the first, greater LSM in transcripts of 40 speed dates predicted a higher likelihood of mutual romantic interest, with an odds ratio of 3.05. Of the pairs above the median for LSM, 33.3% mutually wanted future contact, against 9.1% of pairs at or below it.

In the second, the researchers measured LSM in the instant messages of 86 couples and followed up three months later. Higher LSM positively predicted that the couple was still together, with an odds ratio of 1.95. Of couples above the median, 76.7% were still dating at follow up, against 53.5% of couples at or below it. The authors conclude that LSM “appears to reflect implicit interpersonal processes central to romantic relationships.”

Read the percentages the other way round

The same figures can be read from the other side, using nothing but subtraction. If 76.7% of the higher matching couples were still dating, 23.3% of them were not. If 53.5% of the lower matching couples were still dating, more than half of the group with less matching had not split by the three month point.

The speed date figures work the same way. 33.3% of the higher matching pairs mutually wanted further contact, which leaves 66.7% of them where that interest was not mutual. The difference between the groups is real in the data the authors report. It is a difference between groups of couples, and the abstract offers no figure for picking out any single pair.

A median split puts half the sample on each side of a line. The percentages above describe those halves, not a threshold that a given conversation crosses.

Nobody could do it on purpose

The 2010 paper is where the measure gets strange. In the first of its three projects, two samples totalling 1,744 college students answered four essay questions written in very different styles, and they automatically matched the language style of the questions. Women, students of higher socioeconomic status and students with higher test grades matched more than others.

In the second project, 74 participants completed cliffhanger excerpts from popular fiction. Judges’ ratings of how similar each response was to its excerpt tracked content matching, not function word matching. And the participants were not able to intentionally increase either style or content matching.

Put together, matching showed up without being asked for, it did not respond to trying, and human readers did not pick it up. The third project tracked letters and published writing from three famous pairs: Sigmund Freud and Carl Jung, Elizabeth Barrett and Robert Browning, and Sylvia Plath and Ted Hughes. The abstract reports that matching in their poetry and letters reflected fluctuations in those relationships.

The raw material is already in an export

The second 2011 study used instant messages, and a chat export holds the same kind of material: two people’s written messages, in order. The words LSM counts are all there. What an export does not hold is the other half of that design, the follow up three months later that recorded who was still together.

Which chat export files can be analysed explains which formats carry the text in the first place, and What chat analysis cannot tell you goes through the limits that apply to any reading of a conversation.

Portraitor does not describe an LSM score among its outputs. Its home page lists text markers it tracks over time: “Emoji, punctuation, capitals and message length.” A related idea, that speakers shift their style toward each other, is covered in communication accommodation theory, measured at scale, and Analyse my WhatsApp chat with AI describes what a portrait of one person in a chat contains. Whether one of those markers carries the same meaning at both ends of the chat is a separate question, taken up in do emoji mean the same thing to the other person.

Matching is not the only thing a thread records about how someone writes. Switching between languages or registers inside one conversation is covered in code switching in text messages.

Frequently asked questions

Frequently asked questions

The measure is built on articles and pronouns, not on what anyone said.

Sources

1. Ireland ME, Slatcher RB, Eastwick PW, Scissors LE, Finkel EJ, Pennebaker JW (2011). Language style matching predicts relationship initiation and stability. Psychological Science 22(1):39-44. PubMed abstract
2. Ireland ME, Pennebaker JW (2010). Language style matching in writing: synchrony in essays, correspondence, and poetry. Journal of Personality and Social Psychology 99(3):549-71. PubMed abstract
3. LIWC, Language Style Matching function
4. Portraitor home page