Overview:

Many complex problems cannot be easily modeled by a single function in terms of the raw input of the domain. Furthermore, such functions would be difficult, if not impossible, to learn directly from the raw input. Pipeline approaches ameliorate this situation by decomposing the problem into a series of sequential stages, where the function of each stage is described both in terms of the raw input and the results from previous stages.

We have examined this approach for dependency parsing and relation extraction, emphasizing the modeling of interactions between the sequential stages. We also have looked at developing an active learning protocol for such models, where an effort is made to reduce error propagation of the pipeline throughout the learning process.