Short Text Relation Classification And Its Applications In Healthcare And E-commerce

With the continuous development of the Internet,the website applications are springing up,and the increasing number of users has gradually made the Internet an abundant database.Aiming at a large amount of Internet data,this study proposes a basic framework for short text relation classification.In order to reduce the data noise,this study proposes the expansion methods of important entity dictionaries,and attempts to apply the general Chinese grammatical error correction model to preprocess the web text first.Aiming at the problem of lacking labeled data,this study designs an algorithm that combines the basic rules learned with a small amount of manually labeled data and introduces external knowledge to assist in judgment,so as to construct a labeled dataset with a considerable scale.Experiments show that the quality of the constructed dataset can be guaranteed,and the model performance can be significantly improved.For relation classification task,this study proposes two methods,one is based on support vector machine models in scenarios requiring high interpretability,and the second one is a neural network based on pre-trained language models.The experimental results in different scenarios prove the effectiveness of our proposed methods.Compared with other baselines,the F1 score and accuracy are improved.In addition,this study makes extensions in real applications,in the fields of medical health and e-commerce that are closely related to people's lives.For the relation classification model in the healthcare scenario,this study designs a rank strategy for the side-effect extraction results,and obtains some symptoms missing in the associated drug package inserts,which will be further provided to the cooperating pharmaceutical companies for subsequent research.As for the relation classification model in the E-commerce scenario,we have proved the deficiency of current language models in commonsense relation judgment,indicating that classification with commonsense knowledge is still a difficulty in natural language processing research.