Deriving Agent-Based Simulation Models from Textual Data

This paper describes a project to investigate the feasibility of building agent-based social simulation models from a combination of textual data and limited observational notes. As an example, we have developed a proof-of-concept Java-based simulation of one notional day at a scientific conference, based on the MODSIM 2005 International Congress on Modelling and Simulation. Textual processing of the conference proceedings was used to build a common-interest network between authors, which in turn was used to drive agent processes for choosing paper presentations to attend and people with which to interact. Specifically, correlations between frequencies of word pairs was used to derive similarities between papers, and hence between authors. Diary notes taken during the MODSIM 2005 International Congress were also used in developing the simulation, which was written in the Java programming language. The conference simulation shows realistic emergent phenomena, such as increasing clustering of conversing participants during the course of the day. The use of textual data does not completely eliminate the need for observational data, and more extensive observational notes would have assisted in producing a more detailed simulation of MODSIM 2005. However, we have demonstrated that the analysis of textual data can be used to simplify model development. In developing more general organisational simulations, time lags in the production of streams of text documents can be used to estimate the delays inherent in organisational processes. As a preliminary investigation of this approach, which we intend to pursue in further work, we show how a time delay can be extracted from the simple text snippets contained in two Internet news feeds.