Explaining Away Stylistic Coordination in Dialogues

Communication Accommodation Theory (CAT) states that people tend to adapt their communication style (voice, gestures, word choice, etc.) in response to the person with whom they interact. Originally, experiments on linguistic accommodation were confined to small scale laboratory settings with a handful of participants. The recent proliferation of online social networks sites offers an opportunity to study nuances of human communication behavior on much larger scales. Indeed, several recent studies have indicated presence of coordination in communication, e.g., [3, 1]. In particular, they show that one person’s use of a linguistic feature (e.g. prepositions) increases the probability that a response will include the same feature. Furthermore, it has been suggested that the relative strength of this effect may reflect the relative social status of participants [1]. Here we suggest an information-theoretic approach for measuring stylistic coordination in communication. Namely, given a temporally ordered sequence of utterances (verbal or electronic statements depending on the context) by two individuals, we characterize their stylistic coordination with time-shifted mutual information. We revisit some of the case studies where linguistic coordination was reported. One of our main observations is that the observed correlations in linguistic features can be explained by confounding factors rather than linguistic accommodation theory. In particular, we show that almost all of the correlation between stylistic features is explained by a simple confounding variable, namely, length of utterance. For instance, suppose a longer utterance from user Y will tend to solicit a longer response from user X. If the probability of an utterance containing a feature, e.g. prepositions or words whose second letter is “r”, depends only on length, this will create the illusion of stylistic coordination on the given feature. While our results do not completely rule out linguistic accommodation, we show that conditioning on utterance length explains almost all the correlation in stylistic features previously reported [1]. Furthermore, we provide a novel framework to account for confounders when measuring stylistic accommodation. Measuring stylistic coordination with LIWC Linguistic Inquiry Word Count, or LIWC [4], is a dictionary-based encoding scheme that has been used extensively for evaluating emotional and psychological features in various text corpora. The latest version of the LIWC dictionary contains around 4500 words and word stems. Each word or word stem belongs to one or more word categories or subcategories. Various LIWC categories include positive and negative emotion, function words, pronouns, articles, and so on. Here we focus on eight LIWC categories that have been used in previous studies [1]: articles, auxiliary verbs, conjunctions, high-frequency adverbs, impersonal pronouns, personal pronouns, prepositions, and quantifiers. Utterances are represented as eight-component binary vectors indicating the presence or absence of each linguistic marker [1]. Information-theoretic characterization of coordination Consider a sequence of exchanges between dialogue participants where user Xm responds to an utterance from user Ym. Each utterance is represented as a binary variable for a given LIWC feature, m. Thus, our dataset for this pair of users is a collection {xk, yk}k=1, where xk, yk = {0, 1} indicate the absence or presence of a particular marker, and K is the total number of exchanges. Point-wise Mutual Information (PMI) between two random variables X and Y is defined as pmi(x, y) = log p(x|y) − log p(x). Note that pmi(x = 1, y = 1) is similar to the coordination measure used in Refs. [1]. The expectation of pmi(x, y) over the joint distribution p(x, y) is the