SITS: A Hierarchical Nonparametric Model using Speaker Identity for Topic Segmentation in Multiparty Conversations Supplementary Material

In this section, we describe the general Gibbs sampler for our nonparametric model. The state space of our chain is the topic indices assigned to all tokens z = {zc,t,n} and topic shifts assigned to all turns l = {lc,t}. In order to obtain zc,t,n we need to know the path assigned for token wc,t,n through the hierarchy which includes kT c,t,n, k S c,s,j and k C c,i. For ease of reference, the meaning of these symbols (and others used in this appendix) are listed in Table 1. Figure 1c shows the relationship among the latent variables in our model. As shown, once we know the three seating assignments kT c,t,n, k S c,s,j and k C c,i, zc,t,n can obtained by zc,t,n ≡ kC c,kS c,st,k T c,t,n (1)