Using a public good experiment with pre-play face-to-face communication (FFC) this paper investigates two channels over which FFC influences contributions by the subjects. Firstly, the contents of the FFC are investigated by categorizing specific strategic information and using simple meta-data. Secondly, a machine-learning approach to analyze facial expressions of the subjects during their communications is implemented. These approaches constitute the first of their kind, analyzing content and facial expressions in FFC aiming to predict the behavior of the subjects in a public good game. Although both approaches are conducted independently the results are consistent: verbally agreeing to fully contribute to the public good until the very end and communicating through facial clues reduces the commonly observed end-game behavior. The length of the FFC quantified in number of words is further a good measure to predict cooperation behavior towards the end of the game. The accuracy of predictions based on the best performing models lies around 7% above the trivial classifier. The obtained findings provide first insights how a priori available information can be utilized to predict free-riding behavior in public good games.