Design And Implementation Of Entity Extraction System In User Query For Mobile Terminal Devices

As the basic task of Natural Language Processing,Entity Extraction has broken through with the rising of deep learning.Named Entity Extraction has played an irreplaceable role in QA system,interactive chat and machine translation and so on.Recently,with the ascending demands for intelligent semantic intercation and AI's boosting,Entity Extraction has been emerging as a flashpoint in user query precessing.Compared to the traditional named entity recognition,it has a broader fields freedom,more strict limits on precision and recall rate and more sophisticated interactive routines.Based on the extraction results,we can complete a series of resources request and service dispatch,in order not only to meet the users' demands,but also motivate their potential desirements.So,we have implemented two systems,namely online and offline versions,which are composed of entity extraction part and related pattern matching module.The latter module is of necessity to suit the hot queries and compensate for the model's incapability.In this paper,we mainly use tensorflow for model training,fine-tuning and deployment.As to the baseline of our experiments,we used seq2 seq architecture instead of the encoder-only method and achieved better results.And then we tuned our baseline in terms of data scale,input granularity,regularization and attention mechnism.At last,we made some changes in model architecture by way of embedding ajustment,attention machinism and novel methods to yiled a better performance.Totally,the new results outperform the older one by over 10 percentage.To sum up,we tested our model on the open MSRA NER dataset and achieved state-of-art performance.Furthermore,we will continue our research in field of CNN encoder,attention mechnism and entity disambiguation.As for the deployment on mobile devices,we have done some optimizations and improvements in terms of software and hardware.More specifically,the former is mainly composed of model compression and data structure optimization.As for the latter,we mainly used dependency release and instruction-level adaptation.Overall,we have solved the problem of high memory occupation and low inference rate we may ecounter while deploying deep learning model on mobile devices.