Anaphora resolution is an important component of natural language processing applications like information extraction, question answering, or automatic summarization. These applications have to deal with unrestricted input which is difficult to process with symbolic anaphora resolution methods. If trained on unrestricted input, machine learning based anaphora resolution methods can robustly deal with a wide variety of input documents. Those methods are mostly implemented as binary classification realizing models of local inference. While this makes the task accessible to standard machine learning techniques, it has the drawback that knowledge about the context is lost. Based on a critical assessment of the state-of-the-art, models of global inference are introduced as a possible alternative.