基于自动回标的地理实体关系语料库构建方法

The corpus of geographical entity relations is the basic data resource of geographical information acquisition and geographical knowledge services, and its scale directly affects the training effect of machine learning models. Fast-updated web text is constantly emerging as a new relational example, requiring the corpus to be updated in a timely manner to cover richer relational instances. Manually constructing and updating corpus are expensive. Therefore, it needs a more efficient technology of corpus construction for massive geographical entity relations. In this paper, we propose an efficient method of corpus construction for massive geographical entity relations through the automatic annotation technique. First of all, based on encyclopedia resources, referring geographical entity classification standard and semantic relation, spatial relation classification standard to establish an annotation scheme of geographical relation, which considers both the linguistic habits of natural language and the annotation normalization. Secondly, we combine the fully- matching with the approximate matching to improve the coverage rate of object entity finding. Thirdly, we define the rules of sentence scoring by using the optimal sequence diagram method, as well as quantitatively evaluate the results of mapping the seed triples to the sentences. Finally, a series of experiments based on the Chinese BaiduBaike are carried out, which is used to verify the effectiveness of the improved automatic annotation. The results show that, the average success rate of the automatic annotation is 67.83%, and the average accuracy of the annotated relations by our method is 76.36%. Comparing with the manually annotated corpus of the spatial relations, the proposed method constructed a large- scale corpus of geographical entity relations more efficiently, which provides a feasible scheme for expending geographical entity relations corpus automatic. Experimental results on self- built corpus by LSTM (Long Short Term Memory) network shows that the accuracy of geographical relation extracting from web texts is 73.2%, and the accuracy of relative corpora is 75.2%, which proofs that the corpus of geographical entity relations is available. At the same time, this method takes into account the semantic relationship and spatial relationship between geographical entities, and it can be used for open relation extraction task. Besides, the relation types are not limited, which can be applied to open relation extraction.