Do New Romantic Couples Use More Similar Language Over Time? Evidence from Intensive Longitudinal Text Messages

Abstract

The digital text traces left by computer-mediated communication (CMC) provide a new opportunity to test theories of relational processes that were originally developed through observation of face-to-face interactions. Communication accommodation theory, for example, suggests that conversation partners’ verbal (and non-verbal) behaviors become more similar as relationships develop. Using a corpus of 1+ million text messages that 41 college-age romantic couples sent to each other during their first year of dating, this study examines how linguistic alignment of new romantic couples’ CMC changes during relationship formation. Results from nonlinear growth models indicate that three aspects of daily linguistic alignment (syntactic—language style matching, semantic—latent semantic analysis, overall—cosine similarity) all exhibit exponential growth to an asymptote as romantic relationships form. Beyond providing empirical support that communication accommodation theory also applies in romantic partners’ CMC, this study demonstrates how relational processes can be examined using digital trace data.

The digital text traces left by computer-mediated communication (CMC) provide a new opportunity to test theories of relational processes that were originally developed through observation of face-to-face (FtF) interactions. Communication accommodation theory (CAT) hypothesizes that conversation partners’ verbal and nonverbal behaviors become more similar as relationships develop (Dragojevic, Gasiorek, & Giles, 2016; Giles, 1973). The theory suggests that linguistic alignment—the convergence of linguistic choices among interlocutors (Duran, Paxton, & Fusaroli, 2019)—should increase as partners grow cognitively and emotionally closer, deepen their affiliation with each other, and gain common ground understanding in their communication. With romantic relationship initiation and maintenance now increasingly occurring through digital media (Caughlin & Sharabi, 2013; Finkel, Eastwick, Karney, Reis, & Sprecher, 2012), this study uses digital traces of CMC, specifically text messages exchanged between new couples, to examine how linguistic alignment changes as romantic relationships develop.

Communication accommodation and relationship development

A variety of theories suggest that language use within a relationship changes over time. For instance, the activation-level model, as described in Street and Giles (1982), posits that linguistic alignment is in part driven by environmental stimuli that impact and activate dyad members in similar ways (Webb, 1972). Cappella’s (1981) discrepancy-arousal model posits that linguistic alignment is driven by changes in individuals’ arousal that accompany discrepancies between the experienced and expected behaviors of their partners. More specific to language itself, CAT posits that individuals adjust their communication behaviors, particularly their language use, in order to become (or as a reflection of becoming) cognitively and emotionally closer (Dragojevic et al., 2016).

Researchers have examined linguistic alignment within a variety of CMC contexts, including Tweets (Danescu-Niculescu-Mizil, Gamon, & Dumais, 2011; Doyle, Yurovsky, & Frank, 2016) and e-mails (Srivastava et al., 2018). Prior work finds that individuals converge in their “textisms” (i.e., digital nonverbal cues), with differences in liking, power, and gender impacting the level of convergence (Adams et al., 2018). However, similar to research conducted on FtF interactions, not much is known about how linguistic alignment changes over time. We aim to start filling that gap.

Following the abundance of research investigating CAT, quantitative and qualitative, across languages, cultures, and social contexts (Zhang & Pitts, 2019), we use newly available CMC data to observe changes in language that manifest during formation of romantic relationships. These data provide a new venue for articulating and testing theoretical propositions. CAT hypothesizes that partners aim to cultivate and maintain a positive identity within the relationship and do so by adjusting their own communication behaviors (Dragojevic et al., 2016). As romantic relationships develop, partners are motivated to gain each other’s approval and decrease social distance and may thus modify their communication behaviors accordingly. For instance, partners may modify their language to more closely match their partner’s language in order to increase similarity and facilitate liking. CAT also hypothesizes that partners will adjust their communication behaviors to increase comprehension (Dragojevic et al., 2016). As romantic relationships develop, partners are motivated to increase comprehension and reduce uncertainty in their communications with each other. For example, partners may initially interpret the phrase “be there soon” differently, but eventually converge on a shared meaning (e.g., that the phrase actually means “I’ll be there within the next hour”). In sum, CAT proposes that the affective and cognitive mechanisms driving changes in communication accommodation motivate partners to (a) manage social distance/(dis)affiliation and (b) facilitate satisfying interactions and mutual understanding. In the case of romantic relationships, we expect that communication accommodation increases during relationship development as partners seek to decrease social distance (i.e., express closeness) and gain greater understanding of and with their partner.

Increasing linguistic alignment as the relationship develops

Changes in affective and cognitive processing that lead to increased understanding as a romantic relationship develops should manifest as greater convergence and accommodation in how individuals’ converse. For example, couples can coordinate their speech patterns by using similar syntax. CAT suggests that as a relationship develops, partners may match each other’s function word use—increase syntactic alignment—to communicate more efficiently, reflect a shared outlook, and signal engagement in the conversation. Indeed, a recent meta-analysis of findings from studies of romantic couples indicates that greater use of “we-talk”—the use of first-person plural pronouns instead of first-person singular pronouns—is associated with higher levels of relationship satisfaction and commitment (Karan, Rosenthal, & Robbins, 2019). As well, partners who use words that are more similar in meaning to their partner’s words—greater semantic similarity—make more intimate self-disclosures during a conversation (Babcock, Ta, & Ickes, 2014), which may reflect increased understanding of each other, greater perspective-taking, or a shared outlook. In sum, change in partners’ linguistic alignment may indicate changes in affiliation that accompany emergence (or not) of a couple identity; however, no work to date has examined how linguistic alignment changes over time as romantic relationships develop.

Stable alignment when relationships mature

Following the propositions of CAT and empirical work supporting CAT, the hypothesis is that linguistic alignment in new romantic couples’ CMC increases as the relationship progresses. However, linguistic alignment is expected to asymptote, in that the similarity in partners’ language will reach an upper limit and/or converge to an “optimal” level of similarity. Total linguistic alignment would mean that partners are solely parroting each other, saying the same thing to each other over and over. Because some level of novelty is needed to keep the conversation moving, there is a natural ceiling in linguistic alignment. As well, there is some evidence that overaccommodation—changing one’s linguistic behavior to a greater extent than expected or desired by one’s partner—can be perceived negatively (e.g., as patronizing or condescending, Giles & Smith, 1979) and thus would not facilitate relationship development. Following CAT and the notion of an optimal level of convergence, we hypothesize that linguistic alignment will, as shown in Figure 1, increase over time as romantic relationships develop and eventually plateau at an upper bounding asymptote.

Figure 1.

Exponential Growth in Linguistic Alignment as Relationships Develop.

Communication Accommodation Theory (CAT) suggests that linguistic alignment follows an exponential growth trajectory during relationship development as members of a dyad become cognitively and emotionally closer. As time progresses left-to-right along the x-axis (e.g., weeks or months since meeting), linguistic alignment increases along the y-axis up towards the optimal level of linguistic alignment indicated by the horizontal dashed line. As indicated by the spacing of the text-bubbles within the larger circles, there is little overlap of dyad members’ language at the beginning of the relationship and the extent of overlap increases toward the optimum level as the relationship progresses.

Three aspects of linguistic alignment

A variety of methods are used to measure linguistic alignment. Here, we use three commonly used metrics—syntactic alignment, semantic alignment, and overall alignment—to capture different aspects of linguistic alignment.

Syntactic alignment

Syntactic alignment is often operationalized in terms of similarity in how conversation partners use particular parts of speech or function words (e.g., pronouns, conjunctions, adverbs). One benefit of examining syntactic alignment within conversations is that syntax is not (or is at least less) dependent on the content of the conversation. Specifically, syntactic alignment focuses on how people converse rather than on the content of their conversation. Furthermore, syntactic alignment is associated with relationship qualities (e.g., relationship stability; Ireland et al., 2011) and has been used to compare conversations in different contexts, such as support conversations in married couples and conflict conversations in adolescent dating couples (e.g., Bierstetel et al., 2020).

Language style matching (LSM; Gonzales, Hancock, & Pennebaker, 2010; Ireland et al., 2011) is one metric that has been used to quantify syntactic similarity within conversations. Following the idea that relationship features are reflected in unconscious use of similar kinds of words, this metric quantifies the similarity in partners’ use of nine different types of function words: articles, auxiliary verbs, conjunctions, high-frequency adverbs, impersonal pronouns, negations, personal pronouns, prepositions, and quantifiers (Gonzales et al., 2010; Ireland et al., 2011). LSM is calculated as

$$\mathit{LSM}=\frac{1}{K}\sum _{k=1}^K\left(1\mathrm{ }-\mathrm{ }\frac{\left|{\pi }_{1k}-{\pi }_{2k}\right|}{{\pi }_{1k}+{\pi }_{2k}+0.0001}\right)\mathrm{ }$$ (1)

where π_1k and π_2k indicate the proportion of k = 1 to K types of function words used by Partner 1 and Partner 2, respectively, during a conversation (or other unit of analysis). Following this approach, the LSM score is the absolute difference in proportions for each dyad across all K = 9 function word categories listed above, normalized for total use of each category and number of categories (and a small constant to avoid division by zero). LSM scores can range from 0 to 1, with higher scores indicating greater syntactic alignment.

LSM has been used to examine syntactic alignment of the language used by a variety of types of dyads—from potential romantic partners to friendship pairs—and in a variety of different contexts—from speed dating scenarios to support conversations. In cross-sectional studies, greater LSM is related to relationship initiation and stability within romantic couples (Ireland et al., 2011), group cohesion and task performance (Gonzales et al., 2010), perceptions of support (Rains, 2016), positive emotion (albeit dependent on the conversational context; Bowen, Winczewski, & Collins, 2017), and emotional improvement as mediated by cognitive reappraisal (Cannava & Bodie, 2017). Here, we expect that LSM will increase towards an asymptote as a romantic relationship develops and matures.

Semantic alignment

Another feature on which partners’ language can align is semantics—the meaning of words. Here, we follow prior communication literature and use latent semantic analysis (LSA; Günther, Dudschig, & Kaup, 2015; Landauer, Foltz, & Laham, 1998) to examine changes in semantic alignment.

The extent of similarity in the semantic content of two documents can be quantified using LSA (Günther et al., 2015; Landauer et al., 1998). In brief, words can be considered as points in a high-dimensional semantic space where words that are similar in meaning are located closer together and words that are quite different in meaning are located farther apart. After locating the words used by each partner in the semantic space, we can assess the extent of overlap and similarity in those words—semantic similarity. Analytically, LSA is comprised of three primary steps (Landauer et al., 1998). First, the text of interest is represented as a term-document matrix. When examining semantic similarity between partners’ words during a conversation, each word used in the conversation would be represented as a row in this matrix and each document (i.e., each partner in the conversation) would be represented as a column. The frequency of each word a partner used during the conversation would be represented in the matrix in the corresponding row and column. The term-document matrix is then weighted to differentiate documents. Low-frequency words are given more weight than high-frequency words since words that occur across most or all documents will not help differentiate documents. Second, to facilitate analysis and interpretation, the number of dimensions representing the semantic space captured in the weighted term-document matrix (i.e., the number of rows in the matrix) is reduced using a singular value decomposition. This (relatively) lower-dimensional space then provides a succinct representation of where all the meaningful words are in the semantic space. Third, the words within each document (i.e., the words used by each partner in the conversation) are projected into the lower dimensional semantic space and summarized as vectors, one for each document. The relative orientation of the vectors in the lower dimensional space reflects the extent of overlap between the two documents, specifically calculated as the angle between the two vectors (i.e., cosine similarity).

LSA has been used to study initial interactions between strangers (Babcock et al., 2014; Ta, Babcock, & Ickes, 2017) with more “involving” conversations exhibiting greater semantic similarity, and interactions between patients and physicians (Vrana et al., 2018) with patients having greater trust in physicians when semantic similarity of patient and physician word use was greater. Here, LSA is used to examine semantic similarity of couples’ text messages to obtain insight into how linguistic alignment changes as romantic relationships develop.

Overall alignment

The two metrics just reviewed—LSM and LSA—quantify syntactic and semantic similarity, respectively. Although these metrics generally capture different features of language use (correlated only r = 0.35 in prior work; Babcock et al., 2014), they may miss some aspects of language entirely. Both methods make assessments about similarity based on the words included in a specific dictionary of terms (e.g., in the Linguistic Inquiry and Word Count program; Pennebaker et al., 2015) or in a semantic space defined by a specific corpus (e.g., Wikipedia). Any and all words not in the pre-established dictionary or corpus are discarded prior to assessment of similarity. Thus, we also make use of a comprehensive, overall linguistic alignment metric that encompasses both syntactic and semantic similarity and uses all words. The concern in the context of CMC is that couples may use “netspeak,” words that are not included in common dictionaries or corpuses. For example, CMC often has typographic (e.g., emoticons such as :-D, numbers for words such as “2” for “to”) and orthographic (e.g., abbreviations or acronyms such as “omw” for “on my way”, spellings that represent different changes in prosody or nonlinguistic sounds such as “heyyyyy” or “haha”) elements that are not included in most dictionaries (Herring, 2012). Thus, to accommodate the emergence of netspeak, we also assessed overall linguistic alignment.

Cosine similarity is a general measure indicating the extent of word overlap between (bags of words) documents. In the case of a conversation, all of the words said by each partner would comprise its own document, and the overlap in the words used between documents is quantified using cosine similarity. Specifically, cosine similarity is calculated as

$$\mathit{Similarity}\left({\mathit{doc}}_1,\mathrm{ }{\mathit{doc}}_2\right)=\cos \left(\theta \right)=\frac{\stackrel{⟶}{{\mathit{doc}}_1}·\stackrel{⟶}{{\mathit{doc}}_2}}{‖{\mathit{doc}}_1‖‖{\mathit{doc}}_2‖}$$ (2)

where the dot product of the term frequency vectors representing each document (in this case, each partners’ entire word space within a day) is divided by the product of the norm of each document vector. In other words, the similarity of words used across partners is obtained by multiplying and summing the term frequencies for each word across partners (i.e., documents) and then scaling to the number of total words used by the partners. The greater the similarity in frequency of partners’ use of specific words, the higher the cosine similarity.

Cosine similarity has been used to discover friendship links by examining the homophily of metadata on social media, with increased topical similarity capturing actual friendship links (Aiello et al., 2012). This example demonstrates how cosine similarity as a metric of linguistic alignment can be used to examine characteristics of computer-mediated interactions. Here, we use cosine similarity to capture changes in overall linguistic alignment that occur during relationship development.

In sum, these three metrics of linguistic alignment capture distinct aspects of communication accommodation. Specifically, LSM captures a shared style of communication between partners, LSA captures shared meaning of communication between partners, and cosine similarity captures partners’ use of a shared vocabulary.

Methods to observe changes in linguistic alignment

Linguistic alignment has been studied in dyadic and group contexts using both quantitative and qualitative approaches, with most of the quantitative studies using cross-sectional designs (e.g., laboratory-based conversations). Examining changes in linguistic alignment, however, requires observation of many conversations over time. The digital traces of CMC provide a new opportunity to study changes in linguistic alignment over time.

Most young (i.e., college age) couples use text messaging on a daily basis to keep in touch (Coyne et al., 2011; Toma & Choi, 2016), with some studies estimating that romantic partners send and receive on average between 65 and 75 messages to each other per day (Brody & Peña, 2015; Luo, 2014). Several studies have examined actual CMC exchanges in romantic couples, including e-mail exchanges between members of new couples (Sharabi & Dykstra-DeVette, 2019) and Gmail instant message exchanges between couples (Slatcher, Vazire, & Pennebaker, 2008). Only one of these studies (Slatcher et al., 2008) formally quantified the syntax used in the CMC exchanges, although these researchers did not examine how these linguistic features changed over time. The ubiquity and frequency of CMC in new romantic couples allow researchers to unobtrusively observe couples’ interactions in situ over longer periods of time and to examine how the language use changes.

The present study

We examine changes in linguistic alignment in couples’ CMC as relationships develop. Drawing from CAT (Dragojevic et al., 2016; Giles, 1973), we hypothesize that syntactic, semantic, and overall linguistic alignment will exhibit exponential growth as the relationship develops—that is, linguistic alignment will increase during earlier stages of the relationship and level off to an “optimal” asymptote. To test this hypothesis, we fit exponential growth models to daily repeated measurements of three aspects of linguistic alignment of couples’ text messages. We believe this is the first study to collect weeks/months of text messages from new romantic couples and use multiple metrics of linguistic alignment to examine changes in communication that occur during relationship formation.

Method

We tested our hypotheses using text message logs and survey responses obtained from college-age couples participating in The Texting Life of Couples (TLC) Study (approved by university Institutional Review Boards).

Participants

Participants were 41 dating couples at two large universities in the mid-Atlantic region of the United States who were recruited through course-related participant pools, university study finder webpages, flyers, and class announcements to take part in a study of new dating couples and their text messaging behaviors. Couples were eligible for participation if they (a) were adults (age > 18 years), (b) primarily spoke and used English, (c) were a member of a dating couple in which the first text messages were exchanged within (approximately) the past year, (d) were able to visit a campus-based research lab, (e) were living at or near the university (i.e., no long-distance couples), (f) used iPhones, and (g) at least one partner had the Messages application installed on their Mac laptop. After eligibility for the study was confirmed, couples were invited to a research lab on the university’s campus (between December 2018 and December 2019). The N = 82 participants (41 couples; 38 heterosexual, 3 same-sex; 44 female, 38 male) were between ages 18 and 23 years (M_Age = 20.55, SD_Age = 1.15) and self-identified (more than one category permitted) as Caucasian (n = 68), African American/Black (n = 4), Middle Eastern (n = 3), Asian American or Asian (n = 3), Hispanic or Latinx (n = 2), and Other (n = 2). Most were “seriously dating” (n = 69), while some were “casually dating” (n = 13) or “other” (n = 2). On average, couples had “been a couple” for 23.61 weeks (SD = 15.73) when they visited the laboratory.

Procedure

Interested participants visited the laboratory where they received additional details about the purpose and design of the study and how their data would be collected, encrypted, stored, managed, and analyzed. After providing consent, participants completed a survey (∼ 30 min) about themselves (e.g., personality, attachment style), their relationship (e.g., satisfaction, commitment, conflict, relational turbulence, relational uncertainty), and their media use (e.g., channels used to communicate with their partner). After the couple completed the survey, a research assistant guided the participant with a Mac laptop (macOS Mojave or later) and Messages through the steps needed to download their text message history using iMessageAnalyzer (D’souza, 2017). The participants selected the text message conversation of interest and saved the message history (images not included) as a TXT file on their own desktop. This file contained columns that indicated who sent each message, the time each text message was sent, and the content of each text message. With the research assistant, the participant replaced the names contained in the first column with “Partner 1” and “Partner 2” and transferred the file to the research lab’s encrypted hard drive. Participants were thanked and compensated with either course credit or a $10 Amazon gift card. Later all names, phone numbers, addresses, locations, and so on were removed from the data to further anonymize and protect the identities of the participants.

Privacy considerations

While digital trace data collected from mobile devices allow researchers to collect highly informative data sets that help researchers test a variety of theories, these data also raise a number of ethical considerations, including transparency, privacy, confidentiality, and beneficence (Martinez-Martin et al., 2018). Our approach to collecting this data was to be as transparent as possible with participants about (a) what information we would be collecting and what we would be doing with that information and (b) our protocols for maintaining their privacy and confidentiality. Specifically, we explained in detail that we would be collecting their text message history (no images), that a researcher would sift through their messages to remove identifiable information (e.g., names, addresses, etc.), that only a small study team would have access to their text messages that were stored on secure servers, and that our analyses would not single out any couples’ specific messages and that the text would be primarily examined in an algorithmic way (i.e., a computer would be quantifying their language). We also considered multiple perspectives of beneficence and whether and how the benefits of this study outweighed the risks. We, along with two university research protection offices, concluded that the type of data used in this study will provide valuable insights into interpersonal processes in ways previously unavailable, and that academic researchers should be developing and examining questions that benefit the general public in ways that the private companies with similar types of data may not pursue.

Measures

The text message corpus consists of 1,045,340 text messages that members of 41 couples sent to each other during 6,452 days of interaction (Brinberg et al., under review). The raw text message data were preprocessed in standard ways: (a) all non-text-based messages were removed, including images (2.3%, e.g., pictures, memes, bitmojis, games) and reactions (1.9%, e.g., “likes” or “emphasis”), (b) all text was lowercased, and (c) each day of text messages (defined as 3 a.m.–2:59 a.m.) was collected together to form separate documents for each partner for each day. Days on which only one partner sent text messages were set aside (N_days = 156, 2.4%) and a few stretches (total of 38 days across all couples; <1% of study days) where the log only contained texts from one partner were treated as missing (because linguistic alignment for nontext days is undefined) and set aside. Possible reasons for this missingness span couples spending the day together and not sending messages, traveling with no WiFi connection, a broken device, or temporary issues with participants’ cloud not storing messages. This analysis is based on 6,243 days of text messages (N = 1,001,454) provided by 41 couples.

Daily syntactic alignment: LSM

LSM scores were calculated for each couple for each day using nine function word measures obtained from the Linguistic Inquiry and Word Count program (LIWC; Pennebaker et al., 2015) using Equation 1. Prior to obtaining the necessary measures from LIWC, standard practice is to spell check words so the program recognizes all relevant words. However, given the potential for spell checkers to incorrectly “fix” words, we did not correct spelling and we assumed that misspelled words that are unidentified by the computer program were randomly distributed across couples. Additional text preprocessing steps, such as tokenization, are performed by LIWC. On the average day, the prototypical couple’s average LSM score was 0.78 (Median = 0.78, SD = 0.08, Min = 0.61, Max = 0.90) and varied across days within-couple by iSD = 0.12 (Median = 0.13, SD = 0.04, Min = 0.03, Max = 0.23).

Daily semantic alignment: LSA

LSA scores were calculated for each couple for each day as the cosine similarity between the daily documents in a latent semantic space (implemented using the LSAfun package in R, Günther et al., 2015). Specifically, the words in each daily document for each partner were mapped as vectors in the 400-dimensional latent semantic space¹ from the word2vec toolkit (https://code.google.com/p/word2vec/), which is based on a 2+ billion-word corpus (comprising the British National Corpus, the ukWaC corpus, and a 2009 Wikipedia dump) that was mapped using an 11-word continuous bag of word windows (Mikolov et al., 2013), a positive Pointwise Mutual Information weighting scheme, and singular value decomposition. The distance between the vectors was then quantified using cosine similarity. It is worth noting again here that words not part of the original corpus are ignored and do not contribute to the similarity calculation. On the average day, the prototypical couple’s average LSA score was 0.91 (Median = 0.92, SD = 0.06, Min = 0.74, Max = 0.98) and varied within-couple across days by iSD = 0.10 (Median = 0.11, SD = 0.06, Min = 0.01, Max = 0.22).

Daily overall alignment: Cosine similarity

Cosine similarity scores were calculated after some additional processing of the raw text. We first collected all the daily documents together into a corpus (segmenting text and punctuation into tokens using white space as word separators). We then stemmed words, changed all letters to lowercase, removed punctuation and common English stop words (e.g., “a”, “the”), and obtained the document frequency matrix. Overall alignment of each couple’s daily document was then calculated as the cosine similarity between the vector of words in each document. In contrast to LSA, cosine similarity defines similarity in the semantic space as the document term frequencies of our text message corpus rather than a pre-established latent semantic space obtained through analysis of a global corpus. Text preparation and calculations were done using the quanteda package in R (Benoit et al., 2018). On the average day, the prototypical couple’s average cosine similarity was 0.43 (Median = 0.39, SD = 0.15, Min = 0.20, Max = 0.92) and varied within-couple across days by iSD = 0.14 (Median = 0.14, SD = 0.03, Min = 0.07, Max = 0.25).

Relationship time

Timing of relationship formation was calculated from each partner’s response to the survey item: “When did you and your partner become a “couple”?” A couple-level relationship formation date was determined using a set of rules: (a) when both partners reported a date (n = 38), the formation date was the middle of the two reports, (b) when only one partner reported a date (n = 1), that was the formation date, and (c) when one partner reported an impossible date (i.e., a date in the future) but the month and day were close to their partner’s report (n = 2), then the year was corrected and the formation date was the middle of the two reports. The modal (n = 28) difference between partners’ relationship formation dates was 0 days (M = 21.95 days, SD = 80.69 days, Min = 0, Max = 481 days²). Relationship time was then defined for each day as the number of days until or since the relationship formation date and rescaled to a month metric (assuming 30.4167 days per month). For each couple, Day 0 was the relationship formation date.

Data analysis

Exponential growth models

We used a nonlinear growth model (Grimm, Ram, & Estabrook, 2016)—exponential function—to examine how linguistic alignment (using three metrics) changed within-couple over relationship time. Specifically, the 6,243 daily linguistic alignment scores nested within 41 couples were modeled as

$${\mathit{LinguisticAlignment}}_{it}={\beta }_{0i}+\mathrm{ }{\beta }_{1i}(1-\exp \left(-{\alpha }_i{\mathit{RelationshipTime}}_{it}\right))+\mathrm{ }{e}_{it}$$ (3)

where LinguisticAlignment_it is the observed linguistic alignment score (LSM, LSA, cosine similarity) for couple i on day t; ${\beta }_{0i}$ is the expected level of linguistic alignment at the time of relationship formation for couple i; ${\beta }_{1i}$ is the total amount of change in linguistic alignment from ${\beta }_{0i}$ to the upper asymptote for couple i; ${\alpha }_i$ is the rate of approach to the upper asymptote for couple i; and $e_{it}$ are day-specific residuals that are assumed to be normally distributed. Couple-specific intercept (${\beta }_{0i}$), amount of change (${\beta }_{1i}$), and rate of change (${\alpha }_i$) parameters were modeled as

$${\beta }_{0i}={\gamma }_{00}+u_{0i}$$ (4)

$${\beta }_{1i}={\gamma }_{10}$$ (5)

$${\alpha }_i={\gamma }_{20}$$ (6)

where ${\gamma }_{00}$ is the expected value of linguistic alignment at the time of relationship formation for the prototypical couple in the sample; ${\gamma }_{10}$ is the expected total amount of change of linguistic alignment from ${\gamma }_{00}$ to the upper asymptote; ${\gamma }_{20}$ is the expected rate at which linguistic alignment approaches the upper asymptote; and $u_{0i}$ are unexplained between-couple differences in the intercepts, which are assumed to be normally distributed. Models with additional random effects were tested, but were trimmed to facilitate model parsimony and convergence.

Growth models were fit to the data in R using the nlme library and visualized using the ggplot2 library (Pinheiro et al., 2016; R Core Team, 2018; Wickham, 2016). Incomplete data were treated as missing at random. Statistical significance was evaluated at p < .05.

Model comparison

To ensure that the exponential growth models sufficiently represented the changes in linguistic alignment during relationship development, fit of these models was compared to alternative change models, including a no growth model (flat line trajectory) and a linear growth model (straight line trajectory). Relative fits were evaluated using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), with lower values of the AIC and BIC indicating better fit.

Results

Couples contributed a span of 7 to 502 alignment-quantifiable days (i.e., both partners sent text messages that contained text) of text message exchanges (M = 197.85, Median = 208.00, SD = 114.50; note that text messages were not necessarily sent on every day within this span). On an average day, partners each sent 519.76 words (Median = 337.98, SD = 437.81, Min = 51.93, Max = 1,857.71), which varied across days by iSD = 402.93 (Median = 309.99, SD = 294.91, Min = 50.72, Max = 1,288.94). Within-couple, the linguistic alignment metrics were, on average, correlated r = 0.60 (Range = 0.09–0.80) for LSM and cosine similarity, r = 0.62 (Range = 0.09–0.87) for LSA and cosine similarity, and r = 0.71 (Range = 0.25–1.00) for LSM and LSA. Figure 2 depicts change in linguistic alignment for two exemplar couples, with each point representing the linguistic alignment score on a given day and the blue line indicating fit of the exponential growth function. From these visualizations, we can see that linguistic alignment varies from day-to-day and a general trend where linguistic alignment increases to an “optimal” plateau as the relationships develop.

Figure 2.

Exemplar Linguistic Alignment Trajectories.

This figure depicts measured linguistic alignment—language style matching (LSM), latent semantic analysis (LSA), and cosine similarity—over the course of the relationship for two couples in the study. The x-axis represents time in the relationship in months, with 0 representing “becoming a couple,” and the y-axis represents level of linguistic alignment. Each black circle indicates level of linguistic alignment on a given day, and the blue line describes changes in linguistic alignment using an exponential trajectory. The vertical red line indicates when couple members reported “becoming a couple”.

Changes in syntactic alignment—LSM—during relationship development

Results from the series of growth models examining changes in linguistic alignment as measured by LSM are shown in Table 1 and Tables S1 and S2 in the supplemental material. As expected, the exponential model provided the best (of the models fit) representation of the data (as indicated by the lowest AIC and BIC). Daily LSM of text messaging (syntactic alignment) exhibited significant change from “becoming a couple” to the upper asymptote of linguistic alignment (${\gamma }_{10}$ = 0.03, p < .001) and a strong, positive rate of approach to the asymptote of linguistic alignment (${\gamma }_{20}$ = 0.17, p < .001). There were also significant between-couple differences in levels of daily LSM at the time of relationship formation (${\sigma }_{u_0}$= 0.07). We conclude that our hypothesis was supported: Syntactic alignment as measured by LSM shows exponential growth during relationship development. Model implied exponential trajectories of daily LSM for each couple (black lines) and the prototypical couple (blue line) are shown in the left panel of Figure 3.

Table 1.

Results from the Exponential Growth Models Examining Changes in Linguistic Alignment over Relationship Development

	Syntactic similarity LSM		Semantic similarity LSA		Overall similarity cosine
Parameter	Est	SE	Est	SE	Est	SE
Fixed effects
Intercept, ${\gamma }_{00}$	0.78*	0.01	0.90*	0.01	0.41*	0.02
Amount of change, ${\gamma }_{10}$	0.03*	0.01	0.06*	0.01	0.08*	0.01
Rate of change, ${\gamma }_{20}$	0.17*	0.02	0.12*	0.02	0.14*	0.02
Random effects		CI		CI		CI
Intercept, ${\sigma }_{u_0}$	0.07	(0.06, 0.09)	0.06	(0.05, 0.08)	0.15	(0.12, 0.18)
Residual, ${\sigma }_e$	0.12	(0.12, 0.12)	0.11	(0.11, 0.12)	0.15	(0.14, 0.15)
AIC	−8073.56		−9217.76		−5733.65
BIC	−8039.87		−9184.08		−5700.24

Note: N = 6,243 days nested within 41 couples. Time is centered at month of relationship formation—that is, when couple members self-reported “becoming a couple.” LSM, language style matching; LSA, latent semantic analysis; Cosine, Cosine similarity; AIC, Akaike information criterion; BIC, Bayesian information criterion; Est, estimate; SE, standard error; CI, 95% confidence interval.

^* p < .05.

Figure 3.

Predicted Linguistic Alignment Trajectories.

Model predicted changes in linguistic alignment—language style matching (LSM), latent semantic analysis (LSA), and cosine similarity. Black lines represent couple predicted trajectories, while the blue line represents the prototypical predicted trajectory. The x-axis represents time in the relationship in months, with 0 representing “becoming a couple.” The y-axis represents linguistic alignment. The vertical red line indicates when couple members reported “becoming a couple”.

Changes in semantic alignment—LSA—during relationship development

Results from the series of growth models examining changes in linguistic alignment as measured by LSA are shown in Table 1 and Tables S1 and S2 in the supplemental material. Again, the exponential model fits the data best. As expected, daily LSA of text messaging exhibited significant change from “becoming a couple” to the upper asymptote of linguistic alignment (${\gamma }_{10}$ = 0.06, p < .001) and a strong, positive rate of approach to the asymptote of linguistic alignment (${\gamma }_{20}$ = 0.12, p < .001). There were also significant between-couple differences in levels of daily LSA at the time of relationship formation (${\sigma }_{u_0}$= 0.06). We conclude that our hypothesis was supported: Semantic alignment as measured by LSA shows exponential growth during relationship development. The middle panel of Figure 3 depicts the model predicted daily LSA trajectories for each couple (black lines) and the prototypical couple (blue line).

Changes in overall alignment—cosine similarity—during relationship development

Results from the series of growth models examining changes in linguistic alignment as measured by cosine similarity are shown in Table 1 and Tables S1 and S2 in the supplemental material. Like the other two linguistic alignment metrics, an exponential model fits the data best. As expected, daily cosine similarity of text messaging exhibited significant change from “becoming a couple” to the upper asymptote of linguistic alignment (${\gamma }_{10}$ = 0.08, p < .001) and a strong, positive rate of approach to the asymptote of linguistic alignment (${\gamma }_{20}$ = 0.14, p < .001). There were also significant between-couple differences in levels of daily cosine similarity at the time of relationship formation (${\sigma }_{u_0}$= 0.15). We conclude that our hypothesis was supported: Overall linguistic alignment as measured by cosine similarity shows exponential growth during relationship development. The right panel of Figure 3 depicts the model predicted trajectories of cosine similarity for each couple (black lines) and the prototypical couple (blue line).

Discussion

The present study examined within-couple changes in linguistic alignment in new romantic couples’ CMC to test propositions of CAT. Using a corpus of text messages obtained from 41 college-age couples who had been dating for less than 1 year, we examined how linguistic alignment—an indicator of communication accommodation—changed as these couples’ relationships formed. We collected, on average, approximately 200 days of text message exchanges from each couple—a rich corpus of actual everyday communication that allowed us to observe and describe how linguistic alignment changed over time. Results were consistent with our hypotheses: Syntactic, semantic, and overall linguistic alignment all exhibited exponential growth, increasing and eventually leveling off to an “optimal” plateau as the relationship developed.

Changes in linguistic alignment during relationship development

Evidence of exponential change in linguistic alignment was present across metrics that focused on both syntactic similarity and semantic similarity in language use. LSM examined syntactic features of language by quantifying the use of function words (e.g., adverbs, pronouns) to assess how partners wrote to each other and reveal various psychological processes, such as conversation engagement. LSA examined semantic features of language by quantifying the similarity of words using a high-dimensional semantic space to capture the amount of common-ground understanding between partners. Exponential change in both syntactic and semantic alignment provides support for the CAT propositions that increases in communication accommodation reflect the greater understanding and decreased social distance that occurs as romantic relationships progress and couples explicitly define their closeness.

It is worth noting, however, that couples had high values of syntactic and semantic similarity as measured by LSM and LSA, respectively, at the time of relationship formation. Furthermore, they exhibited relatively small amounts of change in alignment after relationship formation. One potential explanation for these high levels of syntactic and semantic alignment at relationship formation is that, as can be seen most prominently in Figure 2, most of the increase in linguistic alignment occurs before relationship formation. Imposing a causal interpretation implied by CAT (Giles & Smith, 1979), the increased signaling of affiliation and increased understanding exhibited during the initial stages of relationship development are what lead to the formal formation of a relationship.

We also examined changes in the overall similarity of couples’ language use. Similar to the syntactic and semantic alignment measures, cosine similarity exhibited an exponential trajectory during relationship development. In contrast to syntactic and semantic alignment, however, cosine similarity had a much lower initial and asymptotic value—that is, general similarity in language use started and plateaued at a much lower level. This lower level of general similarity is expected given that the metric is based on use of all words and not just on a few specific types of words (as in LSM) or just the words that also appear in a specific dictionary (as in LSA). In sum, changes in linguistic alignment—as an indicator of the extent of communication accommodation present in conversation—that manifest during relationship development as hypothesized in CAT was supported using metrics that captured syntactic, semantic, and overall alignment of language use. Thus, we infer that partners were able to cultivate and maintain a positive identity and understanding within the relationship and did so by adjusting their own communication behaviors (Dragojevic et al., 2016).

One additional contribution of our study was examining the timing of the changes in linguistic alignment. Current work on CAT is not explicit about the rate of change in (dis)similarity of communication behaviors. We found that linguistic alignment was quite high at the time of relationship formation, and that, although linguistic alignment continued to increase (albeit a relatively small amount) after relationship formation, the greatest amount of change in linguistic alignment occurred before couples formally designated their relationship. Couples may be close to reaching their “optimal” level of accommodation close to the time of relationship formation because under- or overaccommodation at this stage of the relationship may be detrimental (e.g., underaccommodation interpreted as distancing and over-accommodation interpreted as patronizing). If an optimal level of accommodation had been reached prior to relationship formation, we might conclude that accommodation is a necessary, but not necessarily sufficient, component of relationship formation. In contrast, if an optimal level of accommodation had been reached following relationship formation, we might conclude that linguistic alignment is not signaling affiliation. Thus, our finding that couples reach their “optimal” level close to relationship formation is consistent with CAT. However, we remain cautious in this interpretation because we do not know if couples’ communication patterns at the asymptote are in fact optimal. Analytically, the asymptote is simply the level at which couples’ communication patterns have stabilized. Further work is needed to understand whether a couples’ stable behaviors are indeed optimal and what the timing of this stability means in the broader context of relationship development. Future work on CAT should also examine the rate and timing of changes in linguistic alignment and what observed changes in communication accommodation behaviors imply and/or portend about couples’ relationship development. For instance, are faster rates of change in communication accommodation associated with earlier relationship formation? Does a certain threshold of communication accommodation need to be reached for partners to develop or commit to a serious relationship?

Our findings are consistent with other work on linguistic alignment and between-dyad differences in interpersonal similarity, liking, and understanding (e.g., Ireland et al., 2011). For example, prior cross-sectional research has found that higher levels of LSM during FtF conversations are associated with greater relationship stability (Ireland et al., 2011). Similarly, our findings indicate that as the relationship moves toward and beyond “becoming a couple”, LSM scores are increasing toward a relatively higher level of convergence. While not in the context of romantic relationships, researchers have examined changes in linguistic alignment in e-mail exchanges of new employees at a company to measure company enculturation. Srivastava et al. (2018) found that individuals’ linguistic alignment increased over time after initially joining a company and the increase in linguistic alignment differed depending on whether an employee eventually left the company voluntarily or involuntarily. Similarly, we might view our findings as evidence that partners “enculturate” each other when developing a relationship. Applying a similar analytic approach in the context of romantic relationships, future research might test whether initial levels, the rate of change, or the asymptote of linguistic alignment are associated with relationship stability (i.e., whether the couple stays together or breaks up) or differ by gender, race, and culture.

The examination of communication accommodation as it occurs in daily life is now possible. The digital traces of individuals’ text communications provide new opportunities for researchers to examine and probe a wide variety of communication processes. For instance, researchers have examined the association of linguistic alignment and status using Tweets (Danescu-Niculescu-Mizil et al., 2011; Doyle et al., 2016) and, contrary to their hypotheses, found that lower-status individuals (as determined by number of followers or lack of a “verified” status) generally did not change their language to match that of higher status individuals in their Twitter conversations (although it depended on the definition of status and language marker examined). Beyond examining the (a)symmetry of communication accommodation on Twitter, CMC affords many opportunities to examine communication accommodation and test specific aspects of CAT. For example, CAT suggests that the degree of accommodation within a conversation is associated with the nature of the communication episode, such that low-quality interactions should have lower levels of accommodation as partners express disagreement and lower levels of affiliation (Soliz & Giles, 2014). In this study, we aggregated all conversations into daily measures of linguistic alignment. However, it would be possible and informative to examine how communication accommodation changes during specific episodes (e.g., conflict episodes) or conversational contexts (e.g., information seeking, expressing affection). The examination of specific episodes could be both mapped onto findings emerging from studies of conversations that unfold in the laboratory or other controlled venues (e.g., a speed dating event) and provide greater ecological validity. While there may be features unique to CMC that influence linguistic alignment that are not present in FtF conversations, CMC observations provide one context in which to test communication theories in a more naturalistic setting.

Limitations and future directions

While this study provided the first in-depth description of changes in linguistic alignment in CMC during romantic relationship development, there are a few limitations and places for improvement worth considering. First, the couples were a convenience sample that was primarily heterosexual, proximally located, English-speaking (in the United States), and enrolled at a university. Furthermore, there was likely a self-selection bias related to participants’ willingness to share their private text messages with a research team. Thus, the findings from this study should not be interpreted as fully representing how communication accommodation processes manifest, particularly in CMC, across all new romantic couples. Future research should now use the paradigm to examine a more diverse set of couples—diversity with respect to age, sexual orientation, race, culture, relationship label (e.g., “friends with benefits”), occupation, location (i.e., long-distance couples; Jiang & Hancock, 2013)—and specific relational characteristics (e.g., differential status, liking). Second, complete text message history was not available for all couples. Initial interactions via CMC were not present in the text logs for all couples, and thus changes in linguistic alignment during early stages of the relationship were driven by a subset of couples. Future work should develop applications or software that obtain the entire text log or can determine why entire logs are not available. Third, we only collected one channel of these romantic couples’ communication: Text messages sent through the Messages app. Communication accommodation that occurs in other communication channels was not observed. Future studies should work toward collecting multiple channels of communication, including other CMC channels (e.g., capturing the use of phone calls, Snapchat, etc. using Screenomics; Reeves et al., 2019) and FtF interactions (e.g., using the Electronically Activated Recorder that captures conversations in daily life; Mehl, 2017). Fourth, we summarized communication accommodation at the daily time-scale—that is, we aggregated all text messages exchanged during 24-hour periods to calculate linguistic alignment. We recognize, however, that important changes in communication accommodation may also manifest at other cadences. Future work should examine how communication accommodation changes or fluctuates at other time-scales (e.g., conversation episodes, hours, weeks). Fifth, we only examined three measures of linguistic alignment, each of which has limitations. For example, LIWC does not handle misspellings, which may result in a decrease in LSM scores that randomly affects measures of couples’ syntactic alignment. Furthermore, we slightly altered some of the raw text to ensure privacy. Given that we inserted generic words (e.g., ADDRESS) to de-identify the data, the linguistic alignment scores may be slightly higher than if these metrics were calculated on the original raw text data because couple members may have both been referring to a similar general category (e.g., restaurant, friend) but to different specific instances within that category. In addition to the three metrics we used, there are a variety of other metrics that could be used to examine linguistic alignment including concept hierarchies such as Wordnet (Pedersen, Patwardhan, & Michelizzi, 2004). Sixth, our models tested changes of linguistic alignment against a “null” model of no linguistic alignment. The nature of conversation though, even randomly paired dialogue between partners from different conversations, may result in linguistic alignment values that are significantly greater than zero. We did some follow-up analyses to check that the observed growth in linguistic alignment was more systematic and larger than any natural fluctuations that might be present in these couples’ text communications. Specifically, we (a) shuffled each partners’ days of texts within a dyad, (b) re-calculated linguistic alignment for each “day” for each couple, and (c) re-fitted the exponential growth models for each of the three linguistic alignment metrics. As expected, the shuffled data did not exhibit any systematic patterns of change. Future work should further explore how permutation tests and surrogate data generation methods (Moulder et al., 2018) can be used to test/reject alternative models about how language use may change. Finally, we did not examine the direction of accommodation in language—that is, who is accommodating to whom. Future work can better parse these within-dyad dynamics (e.g., using the hierarchical alignment model; Doyle et al., 2016) to understand how communication accommodation differs based on gender, race, power structure, and other characteristics of the partners.

Conclusion

Unobtrusive measures of CMC provide new contexts to examine changes in communication processes during relationship development. Using exponential growth models and text messaging logs, we tested hypotheses derived from CAT, specifically that linguistic alignment would increase during relationship formation and stabilize as the relationship matured. As expected, syntactic, semantic, and overall linguistic alignment increased and eventually leveled off as the relationship developed. Our findings provide evidence that the propositions of CAT also apply in CMC and set the stage for further examination of relational processes using digital trace data. We look forward to work that capitalizes on longitudinal digital trace data to examine communication behaviors in situ and develop and test theory about how communication behaviors change over time.

Supporting Information

Additional Supporting Information may be found in the online version of this article. Please note: Oxford University Press is not responsible for the content or functionality of any supplementary materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.