k-Best Max-margin Approaches for Sequence Labeling

Dejan Mančev, Branimir Todorović

Structured learning algorithms usually require inference during the training procedure. Due to their exponential size of output space, the parameter update is performed only on a relatively small collection built from the “best” structures. The k-best MIRA is an example of an online algorithm which seeks optimal parameters by making updates on k structures with the highest score at a time. Following the idea of using k-best structures during the learning process, in this paper we introduce four new k-best extensions of max-margin structured algorithms. We discuss their properties and connection, and evaluate all algorithms on two sequence labeling problems, the shallow parsing and named entity recognition. The experiments show how the proposed algorithms are affected by the changes of k in terms of the F-measure and computational time, and that the proposed algorithms can improve results in comparison to the single best case. Moreover, the restriction to the single best case produces a comparison of the existing algorithms.