Joint Entity Relation Extraction With Deep Learning

With the fast development of Internet,a large amount of free texts are produced in different forms everyday.Information extraction,which is about how to automatically extract knowledge from these free texts,is a key and important task in natural language processing.The information extraction task is to tackle this problem.In this work,we focus on the entity relation extraction task which is a subtask in information extraction.It aims to identify entities that appear in the text,and semantic relations among entities.One common framework for supervised entity relation extraction is a pipeline-based.Specifically,entity models and relation models are trained separately.An entity model is first used to identify entities in inputs,and then a relation model finds relations between these entities.The pipeline method suffers the error propagation problem.Errors from the previous task will accumulate to the next task.To alleviate this problem,many researchers study joint extraction models: extracting entity and relation in a unified model.The main difficulty of the joint extraction is how to handle the interaction between the entity model and the relation model.On the other hand,since large-scale annotation data is often difficult to obtain,distant supervision methods have been applied to the entity relation extraction.The main idea is try to align knowledge bases and large-scale text data which can automatically obtain a large number of training data.However,there are a lot of noise in the obtained dataset.It limits the performance of the distantly supervised entity relation extraction.Furthermore,in some applications,knowledge bases may be unavailable for distant supervision,which enables the distant supervision more challenging.This work explores several entity relation extraction methods from the perspective of data and joint model,and investigates the advantages and limitations of proposed methods.Specifically,the main contributions are as follows:1.In order to alleviate the problem of noisy samples in distant supervision,we propose to use a small amount of high quality heterogeneous manual labeled dataset to help distantly supervised entity relation extraction task.We design an adaptation framework based on multi-task learning,and consider some consistency constraints in the adaptation process,so as to achieve knowledge transfer.Experiments on distantly supervised dataset demonstrate the effectiveness of the proposed framework(perspective of data).2.In order to tackle the problem that there is no knowledge base for distant supervision in some domains,we propose to use linguistic rules to help distant supervision.Firstly,a training set is constructed automatically by using domain-independent linguistic rules,and then a classifier is built based on the training data.Comparing with only rule-based model,the classifier can extract relations that linguistic rules cannot cover.The proposed algorithm is fast and scalable on large-scale dataset.Experiments on Amazon online review dataset demonstrate that the proposed model is able to achieve promising performances(perspective of data).3.In order to handle the interaction between entity model and relation model,we propose a joint extraction model based on minimum risk training,which can strengthen the connection between entity model and relation model by optimizing the global loss function.Experiments on ACE05 dataset demonstrate the effectiveness of the proposed joint model(perspective of joint model).4.In order to handle the joint type inference on entities and relations,we propose a joint model based on graph convolutional network for entity relation extraction.An entityrelation bipartite graph is constructed and a graph convolution network is run on the graph to capture information between multiple entities and relations.Experiments on ACE05 dataset demonstrate the effectiveness of the proposed model(perspective of joint model).