Research And Implementation Of Question Answering Technology Based On Knowledge Base

With the development of information technology,the knowledge accumulated by mankind is increasing exponentially.How to better represent and query knowledge and ensure that knowledge is effectively used has become a research hotspot.Traditional information retrieval methods are based on keyword matching or fuzzy matching methods for text retrieval.These methods are mainly based on character matching and cannot understand human natural language query intentions from the semantic level,and return accurate and condensed query results.We research the knowledge-based intelligent question answering technology to achieve more accurate semantic understanding and answer response.According to the difference of structure and storage methods,we divide knowledge into QA pair knowledge and knowledge graph,and combines vectorized semantic representation technology and deep models to conduct question qnswering process design and key algorithm research on two types of knowledge.For QA knowledge based question answering,the mainstream supervised text matching models have the problem of poor applicability when migrating to new domains.The visualization results show that these models are prone to semantic shift after training.In response to this problem,we propose an attention-weighted text matching model ATT-W that incorporates part-of-speech features.It uses the attention mechanism to learn the weighted feature representation of pre-trained word embeddings,and retains key information by assigning different weights to different words.The introduction of part-of-speech features further improves the matching accuracy.Model training is carried out on the financial field dataset BQ,and the field migration effect is verified on the general field dataset LCQMC.Experiments show that the generalization performance of the ATT-W model is better than traditional supervised matching models such as Siamese CNN,Siamese LSTM,and BERT,and has a 2% increase in F1 score compared to the unweighted Neural BOW model.In addition,this paper uses the vector approximate nearest neighbor matching algorithm to calculate the cosine similarity by constructing a vector index and using local matching instead of global matching to ensure fast text matching in a large-scale knowledge base.For KG knowledge based question answering,The engine first extracts the subject words,locates the knowledge map knowledge and performs question-and-answer matching.In this paper,the Bi LSTM+CRF model is used to extract the subject words in the question and locate the specific entity in the knowledge graph for preliminary recall of candidate answers.In order to match the question with triple knowledge,this paper proposes a QA matching model BERT-KGE.Based on the BERT text representation,the graph embedding model is used to learn the semantic information containing the graph structure from the large-scale knowledge graph,so as to achieve a more reasonable feature representation of the triple knowledge.Experiments based on the knowledge graph and training samples provided by the nlpcc We construct training samples based on nlpcc 2016 KGQA data and perform model training to obtain good topic word extraction and Q&A matching effects.Finally,this paper completes the implementation of related algorithms of QA Engine and KG Engine,designs and implements related functional modules such as knowledge management,QA process module,and log module,and supports question consultation and fact query.Design question answering scenarios to complete related functional tests and performance tests.The results show that the system can accurately understand user questions and return relevant answers,and ensure reasonable processing time and good concurrent processing capabilities.