Question Parsing Technology Integrated With Semantic Information

With the substantial development of deep learning in the field of Artificial Intelligence,Natural Language Processing(NLP)has also made great progress.Among them,NLP tasks such as Named Entity Recognition(NER)and Question-Answering System(QA)have also started to be popularized and applied on a large scale.The research related to QA technology has also deep into the sentences semantic understanding,where question analysis technology plays an important role in the accuracy of QA systems;and NER plays a fundamental support role in tasks such as dialogue generation,relationship extraction,and information retrieval.As the initial part of a QA system,the ability to understand the question sentence correctly can affect how well the subsequent tasks are performed.Question parsing can provide important data information to the system,which is crucial to help users find the desired answers.The current question parsing system has not yet developed a unified standard,and most researchers have adopted an answer type-oriented classification system,which is easy to construct,has a more detailed classification granularity,and covers a wide range of categories.At present,the research on sentence feature extraction for question parsing has made great progress,but there are still some problems that need to be solved,such as: the low accuracy of classification for a single feature,poor classification effect,etc.These problems have led to bottlenecks in development and cannot meet the current needs of intelligent information processing applications.In previous studies,question parsing research has focused on Machine Learning based methods,where the features in the question are manually extracted or the question representation is combined from several features,and the question representation is subjective and diverse,which cannot accurately represent.The Deep Learning based methods have a strong adaptive learning capability,relatively high fault tolerance,and a strong ability to resist complex problems such as noise in large-scale datasets.The performance of question classification is better enhanced by using deep learning methods to analyze and learn the semantic information features in sentences for question parsing.Therefore,in this paper,use NER techniques combined with semantic similarity calculation methods of deep learning to study question parsing techniques and their accuracy performance,construct and compare the accuracy of question parsing techniques based on different extraction semantic methods.In this paper,the analysis is based on question semantic information as the entry point,and the three main works are as follows.(1)Construction of a BERT-Bi-LSTM-CRF based NER optimization modelA NER model based on Conditional Random Fields(CRF)is constructed and trained,and the model performance of CRF is initially verified by using a word training method.The model is trained by combining the information features of the before and after sequences,normalizing the sequence paths,and comparing and analyzing the annotation performance of the two models.The final,a NER method that incorporates multi-model strategies is implemented,and the BERT technique in deep learning is added to the traditional model as a semantic encoding layer to improve the accuracy of the NER model.(2)A question parsing method based on semantic similarity calculation is constructedThe BERT model can be used to fuse the question semantic information features forward and backward,and BERT is used as a word vector encoding layer to output the question sentence vector,solving one-way limitations using Masked Language Model(MLM)tasks and learning sentence-to-sentence relationships using the Next Sentence Prediction(NSP)task,after which the similarity calculation is done for the question sentence vector.(3)A question parsing model fused semantic information is designedIntroducing the modular fusion technology,a step-by-step and modular question parsing model is designed.This chapter investigates the performance of question parsing model fused semantic information by fusing separately designed and trained NER module and semantic similarity module to achieve high efficiency and low coupling.Fuse semantic information into the model to improve the performance of question parsing and compare it with other question parsing models based on different methods.The experimental results show that question parsing model fused semantic information has improved in performance,comparing 2 to 5 percentage points in MRR,respectively,effectively improving the overall performance of question parsing.