A Practical Perspective on Latent Structured Prediction for Coreference Resolution

Latent structured prediction theory pro- poses powerful methods such as Latent Structural SVM (LSSVM), which can po- tentially be very appealing for coreference resolution (CR). In contrast, only small work is available, mainly targeting the la- tent structured perceptron (LSP). In this paper, we carried out a practical study comparing for the first time online learn- ing with LSSVM. We analyze the intrica- cies that may have made initial attempts to use LSSVM fail, i.e., huge training time and the much lower accuracy produced by the Kruskal’s spanning tree algorithm. In this respect, we also propose a new effec- tive feature selection approach for improv- ing system efficiency. The results show that LSP, if correctly parameterized, pro- duces the same performance as LSSVM, being at the same time much more effi- cient.