Relation Classification Via Knowledge Base Entity Representation Learning

With the advent of data boom, it is extremely necessary for us to let the machine read data from Web. Information extraction can automatically extract structured knowledge from huge amount of data for human to query,so it becomes one of the most important issue in both natural language processing and data mining areas. Among several tasks in information extraction, relation extraction plays a key role because it can dig out relation types between named entities and then help the machine to build the structured knowledge base.Due to the variation and abstractness of natural language, traditional rule-based and pattern-based methods need expensive human annotation and feature engineering is also time-consuming. Using the error back propagation algorithm, multi-layer neural network model can learn features automatically. With an end-to-end structure, neural networks dismiss people's concern about feature engineering. This paper utilizes recurrent neural networks to solve the relation extraction task. With feature indicators, a pooling layer and a bidirectional structure, we promote the performance of relation extraction significantly. This paper also points out the deficiency of the popular SemEval-2010 Task 8 dataset and creates more difficult dataset called KBP37. On both datasets, the RNN model not only exceeds the traditional methods, but also exceeds one recent work based on convolutional neural networks.In most cases, people use relation extraction to build knowledge base,but in reverse, knowledge base can also help relation extraction task by providing prior knowledge. This paper creatively proposes a framework to let the relation extraction model absorb the information of entities in knowledge base. Firstly, we combine different forms of knowledge resources in a knowledge base and propose a joint training method to embed information of entities into a low dimensional semantic space. Then,we list the neighbor entities and entities on the path between target entity pair on the knowledge graph. After that we use an attention-based neural model to learn the feature of these entities. Finally, the feature for relation classification is composed of the entity feature from knowledge base and the sentence feature from recurrent networks. This framework finally achieves 3.5 to 6 percent on F1 value, compared with the convolutional neural network baseline.In addition, to make the method more robust to different domains of knowledge or languages, this paper also proposed a framework for entity annotation projection. This work uses parallel corpus between different languages to project entity annotations from English to another language.In conclusion, this paper mainly focuses on relation extraction task.We propose a recurrent neural network structure to extract natural language features from sentences. And we propose a joint learning method to learn the entity features from different resources in knowledge bases. Finally, we propose an attention-based model to combine the sentence feature and entity feature into one framework. This framework can not only handle relation extraction task, but also promisingly deal with other natural language tasks, such as event detection and question answering.