Effect of discourse on network ties: A method and business organization case study
Mehmet Gençer, Özgür AkarsuI- Motivation
Social network analysis and modeling uses several simplifications and historical development of social network models has gone through several phases to better explain dynamics of network formation. Earlier models such as preferential attachment, or assortative mixing used existing structural features themselves (such as centrality) to explain future tie formations. These models essentially treat all ties and actors as equivalent except for their network position: a structural, not "personal" feature. Later stochastic approaches such as exponential random graph models embrace more micro level differences of actors in order to understand network phenomena at a better resolution and enable the modeler to assign dyad level parameters for tie formation probabilities. These models, in effect, enable one to translate social phenomena such as homophily to network modeling, thus allowing us to treat actor differences such as traits or social class based tie formations. However, these modeling apparatus, at best, allow us to express demographic or social actor features in hypothesized models. On the other hand digital era presents us with a lot more information about actors, in the form of texts or preferences expressed in business information systems, social media, etc.
II-Method, hypotheses, and case
Our study proposes a text mining based method to quantify individual and group discourses in a multi-dimensional space. Instead of imposing externally defined thematic dimensions common in the literature, we stick to using word frequencies in order to let the data speak, combined with relevant adjustments used in text mining field.
Using the proposed method allows us to pose and test hypotheses relating the effect of discourse on presence of absence of network ties. We have posed the following questions and related hypotheses:
(i) Is presence of a tie related to similarity of discourses of the actors in the dyad?
(ii) Is centrality of an actor related to ordinariness of his/her discourse when compared to average community discourse, i.e. to how well assimilated the actor into community discourse?
(iii) Is centrality of an actor related to actor's discursive variablity, i.e. how well the actor adjusts his/her discourse to peers?
In order to test our hypotheses we have applied our method to a corpus of 151,180 texts which are written by over 5000 employees of a large business organization to comment on peers of their choosing, as part of a yearly human resources practice spanning 5 years.
III-Results and implications
Our results confirm all three hypothesis above. But more importantly our analysis demonstrates the potential of our discourse representation method in investigating the nature of social relations. Our approach enables one to pose hypothesis involving personal (micro level) variations of actors in a network.
Our approach embodies an analytical apparatus which can be used in both research and practice. Our method to quantify discourses pave way to monitoring and intervening incongruence as well as compatibility between actors and communities; thus can be used in managing organizational change and development of networks, and well being management in organizations.