Pronoun Prediction with Latent Anaphora Resolution

This paper describes the UU-Hardmeier submissions to the WMT 2016 shared task on cross-lingual pronoun prediction. Our model is a system combination of two different approaches, one based on a neural network with latent anaphora resolution and the other one on an n-gram model with an additional dependency on the source pronoun. The combination of the two models results in an improvement over each individual system, but it appears that the contribution of the neural network is more likely due to its context modelling capacities than to the anaphora resolution subnetwork.