SalienceGraph: Visualizing Salience Dynamics of Written Discourse by Using Reference Probability and PLSA

Since public involvement in the decision-making process for community development needs a lot of efforts and time, support tools for speeding up the consensus building process among stakeholders are required. This paper presents a new method for finding, tracking and visualizing participants' concerns (topics) from the record of a public debate. For finding topics, we use the salience of a term, which is computed as its reference probability based on referential coherence in the Centering Theory. Our system first annotates a debate record or minute into Global Document Annotation (GDA) format automatically, and then computes the salience of each term from the GDA-annotated text sentence by sentence. Then, by using the Probalilistic Latent Semantic Analytsis (PLSA), our system reduces the dimensions of the vector of salience values of terms into a set of major latent topics. For tracking topics, we use the salience dynamics, which is computed as the temporal change of joint attention to the major latent topics with additional user-supplied terms. The resulting graph is called SalienceGraph. For visualizing SalienceGraph, we use 3D visualizer with GUI designed by "overview first, zoom and filter, then details on demand" principle. SalienceGraph provides more accurate trajectory of topics than conventional TF·IDF.