SDP-JAIST: A Shallow Discourse Parsing system $@$ CoNLL 2016 Shared Task

In this paper, we present an improvement of the last year architecture for identifying shallow discourse relations in texts. In the first phase, the system will detect the connective words and both of arguments by performing the Conditional Random Fields (CRFs) learning algorithm with models that are trained based on a set of features such as words, part-of-speech (POS) and pattern based features extracted from parsing trees of sentences. The second phase will classify arguments and explicit connectives into one of thirteen types of senses by using the Sequential Minimal Optimization (SMO) and Random Forest classifiers with a set of features extracted from arguments and connective along with a set of given resources. The evaluation results of the whole system on the development, test and blind data set are 29.65%, 24.67% and 20.37% in terms of F1 scores. The results are competitive with other top baseline systems in recognition of explicit discourse relations.