Social Networks are Encoded in Language

Social Networks are Encoded in Language Sterling Hutchinson (schtchns@memphis.edu) Department of Psychology / Institute for Intelligent Systems, University of Memphis 365 Innovation Drive, Memphis, TN 38152 USA Vivek Datla (vvdatla@memphis.edu) Department of Computer Science / Institute for Intelligent Systems, University of Memphis 365 Innovation Drive, Memphis, TN 38152 USA Max M. Louwerse (mlouwerse@memphis.edu) Department of Psychology / Institute for Intelligent Systems, University of Memphis 365 Innovation Drive, Memphis, TN 38152 USA Abstract and exchanges. The formal, sentimental, and interactive nature of the social relationship can be determined by assessing a number of factors. For example, relationships can be predicted in part by the kinship of the individuals. In families, siblings tend to be close friends. Gender especially impacts the nature of relationships such that if a member of the dyad is a female, the relationship is more likely to be successful (Kim, McHale, & Osgood, 2006; Wright & Scanlon, 1991). Environment tends to weigh heavily in terms of whether or not two individuals are likely to build a relationship together. Proximity has also long been established as a strong predictor for relationships of all varieties, with increased proximity leading to increased likelihood of interpersonal relationships (Ebbesen, Kjos, Konecni, 1976). In addition, ties between locations (e.g., commonly trekked routes) also impact social interaction (Takhteyev, Gruzd, & Wellman, 2011). Similarly, familiarity fosters attraction between individuals (Reis, Manianci, Caprariello, Eastwick, & Finkel, 2011; Zajonc, 1968; 2001). Further, those who share interests, attitudes, and characteristics are more likely to develop friendships. In fact, any similarity between two individuals promotes the formation of a relationship between them (Bryne, 1971), with important matters (e.g., religious views, political attitudes) given more weight (Touhey, 1972). Emotions also impact relationships. When two individuals first encounter one another, a future friendship becomes more likely if the interaction is positive, whereas a friendship is not apt to blossom if the interaction is negative (Farina, Wheeler, & Mehta, 1991). Even physical features, like smell or appearance influence the relationships we form (Li, Moallem, Paller, & Gottfried, 2007). After social relations are formed, different factors help these relations to solidify. For instance, Berscheid, Snyder, and Omoto (1989) found that closeness was significantly related to satisfaction of established romantic relationships, as was self-disclosure (Sprecher & Henrick, 2004). Feeney and Noller (1992) argued that individual differences like attachment styles impact the duration of social relationships, as does equity (Hatfield, Traupmann, & Walster, 1978). In addition, when it comes to group relationships, predicting Knowledge regarding social information is thought to be derived from many different sources, such as interviews and formal relationships. Social networks can likewise be generated from such external information. Recent work has demonstrated that statistical linguistic data can explain findings thought to be explained by external factors alone, such as perceptual relations. The current study explored whether language implicitly comprises information that allows for extracting social networks, by testing the hypothesis that individuals who are socially related together are linguistically talked about together, as well as the hypothesis that individuals who are socially related more are talked about more. In the first analysis using first-order co- occurrences of names of characters in the Harry Potter novels we found that an MDS solution correlated with the actual social network of characters as rated by humans. In a second study using higher-order co-occurrences, a latent semantic analysis (LSA) space was trained on all seven Harry Potter novels. LSA cosine values for all character pairs were obtained, marking their semantic similarity. Again, an MDS analysis comparing the LSA data with the actual social relationships yielded a significant bidimensional regression. These results demonstrate that linguistic information indeed encodes social relationship information and show that implicit information within language can generate social networks. Keywords: social relations; social networks; social cognition; statistical linguistic frequencies Introduction What is the nature of social relations and how can such relations be estimated? Social media, such as Facebook, LinkedIn, and Twitter allow us to answer this question, based on individuals choosing their friends. However, when such deliberate decisions are not readily available, how can social relations be measured and social networks be plotted otherwise? Social relations can be interpreted in three non-mutually exclusive ways (Fischer, 1982). First, they can be formal in socially recognized roles, such as teacher/student, employer/employee, or father/son. Second, they can be sentimental, as when individuals feel close to others. Finally, a relation can be defined in terms of interactions