Efficient programmable learning to search.

摘要: We improve "learning to search" approaches structured prediction in two ways. First, we show that the search space can be defined by an arbitrary imperative program, reducing number of lines code required develop new tasks orders magnitude. Second, make magnitude faster through various algorithmic improvements. demonstrate feasibility our approach on three tasks: variants sequence labeling and entity-relation resolution. In all cases obtain accuracies at least as high alternative approaches, drastically reduced execution programming time.