Do you know of any simple (form filling) dialogue state tracking systems with a graphical model based dialogue state representation? I'm looking for one to get familiar with the overall model and especially data driven learning/inference procedure.
This Theano tutorial provide a dialogue state tracking systems (slot filling) with a graphical model based (recurrent neural networks) dialogue state representation.
You probably need a decently large training set though: on the Fourth Dialog State Tracking Challenge (DSTC4) last year, we (and other teams) unsuccessfully tried some neural networks but in the end a simple classifier with decent features beats them. (more details: Franck Dernoncourt, Ji Young Lee, Trung H. Bui, and Hung H. Bui. "Robust Dialog State Tracking for Large Ontologies". International Workshop on Spoken Dialogue Systems. 2016.)
These are some examples I've found myself:
https://github.com/jeremyfix/dstc (rule-based, advanced visualization)
https://github.com/CallumMain/DNN-DST (Deep Neural Net-based)
https://github.com/UFAL-DSG/xtrack2 (Recurrent Neural Net-based)
https://github.com/UFAL-DSG/alex/blob/master/alex/components/dm/dstc_tracker.py (Bayesian Discriminative tracking)
They all are built around the Dialog State Tracking Challenge, and while most of them are not exactly PGM-based, they do track dialogue progress along with turn-wise slot filling.