| Traditional approaches to the task of enterprise information extraction usually design independent local classifiers for specific information to be extracted. Then, in order to get the overall extraction results, the isolated predictions from these independent local classifiers should be further combined to yield a comprehensive output. As a common drawback of these approaches, these independently trained classifiers are incapable of capturing inter-dependencies among multiple local predictions.To tackle the problem, we propose a joint enterprise information extraction method based on structured prediction and demonstrate the effectiveness of this method through experiments. The paper mainly includes three aspects as follows:(1) Systematically analyzes and discusses the basic methods and the current research situation of the enterprise information extraction problem(2) Develops two baseline system, one based on traditional artificial-rules method and the other based on traditional classifier method.(3) By treating the enterprise information extraction problem as a whole task, we propose a joint framework based on structured predictions which extracts all attribute information simultaneously. We design a beam search algorithm for inference and the structured perceptron is adopted for training. In addition to some local features used in the joint model, we also introduce some global features which can explicitly capture the dependencies of multiple attribute information.To verify the effectiveness of our joint model and to compare the results of different systems, a set of experiments have been performed by adjusting parameters and adding or removing the key steps of the algorithm. Experimental results show that the proposed joint approach with local features works well and adding global features further improves the performance significantly. In view of saving time and space, our approach outperforms previous baseline systems based on traditional classifier method. |
PDF Full Text Request