Research On Joint Extraction Model Of Entities And Relations Based On Reinforcement Learning

Joint extraction of entities and relations is an important task in information extraction.Existing entities and relations extraction methods that rely on distant supervision suffer from the noisy labeling problem.In this paper,we propose a model for joint entities and relations extraction from noisy data based on reinforcement learning.The model has two modules: an instance selector module and a sequence labeling module.The instance selector chooses high-quality sentences with reinforcement learning and feeds the selected sentences into the sequence labeling model,and the sequence labeling model makes prediction and provides rewards to the instance selector.The two modules are trained jointly to optimize the instance selection and sequence labeling processes.Experiment results show that our model can deal with the noise of data effectively and obtain better performance for joint entities and relations extraction.The main contributions are shown as follows:(1)For the public data set of distant supervision,we use a label which can extract the information of entities and relations at the same time.After processing,we can transform the joint extraction of entities and relations into the problem of sequence tagging problem on the data set.That is to say,the entities and their relations can be extracted simultaneously in a model through labels containing sufficient information.(2)After transforming the joint entities and relations extraction into a sequential tagging problem,several sequential annotation models are studied,including the classical BI-LSTM-CRF model,the BI-LSTM-LSTM model of LSTM decoder,the BI-LSTM-LSTM model of optimized loss function,the BI-LSTM-LSTM model based on attention mechanism,etc.These improved encoder-decoder models can extract the entities and relationships more accurately.(3)On the basis of sequential tagging model,a sentence selector based on reinforcement learning is introduced for distant supervision noise data sets,and a batch of low noise data is selected as training set to train the joint extraction model.Finally,the joint extraction model and reinforcement learning model are trained together,namely RL-BI-LSTM-CRF model.The sentence selector model based on reinforcement learning can be used as a combination of a single module and multiple sequence annotation models.The experimental results on the public data set constructed by distant supervision show that the improved sequence tagging models and proposed joint entities and relations extraction model based on “RL-BI-LSTM-CRF” is more effective than the traditional “BI-LSTM-CRF” model.