Graphical models are a powerful tool to model the complicated
problems in an intuitive way [10]. They have been applied to classify
semantic relations in bioscience texts [36]. Graphical models can be
categorized into directed graphical model (Bayesian networks) and
undirected graphical model. The recently emerging Conditional
Random Fields (CRF) [25] is an undirected graphical model. Compared
with the Bayesian networks, CRF directly models the conditional
probability distribution of the output given the input, so it can exploit
the rich and global features of the inputs without representing the
dependencies in the inputs. Also CRF needs to estimate fewer
parameters than the Bayesian networks, so it has excellent performance
when the training sample is small. The comparative relation
extraction involves multiple entities and long-range dependencies,
and needs to capture rich features from the inputs, so CRF is an ideal
tool to use for it.