Design And Implementation Of Product Information Question-Answering System Supporting Multiple Knowledge Sources

Question answering systems can recognize human question inputs in the form of natural language,and use highly refined and accurate natural language to quickly answer human questions.Compared with traditional information retrieval systems,users of QA systems do not need to use unnatural forms such as keywords to retrieve information,and the answers given by QA systems are not web pages or documents containing large texts,but more concise and accurate answers to specific questions,organized in natural language.Question answering systems have played great roles in many fields,but they still have their shortcomings.QA systems based on FAQ have greater limitations in the coverage of user questions;QA systems based on unstructured text have higher coverage rates,but the accuracies are difficult to improve;Although QA systems based on knowledge graph can give more accurate answers,the construction of the systems requires a lot of labor costs.It can be seen that the design of QA system based on a single knowledge source cannot make full use of the existing multiple knowledge sources in the industry,and it is difficult to balance effects and costs.As for the technology of constructing QA systems,most of the existing QA systems use rule-based or statistical methods,which in principle limits their ability to utilize multiple knowledge sources.How to use more advanced technology to improve QA systems' ability to process and utilize different knowledge sources and design a QA of better quality is a huge challenge.To make good use of knowledge from multiple different sources,it is necessary to analyze the characteristics of the knowledge from each source,and design corresponding algorithms according to their characteristics,so that the system has the ability to process multi-source knowledge with different characteristics at the same time.This is a huge project.The system that this paper will introduce uses a variety of technologies including knowledge graphs,named entity recognition,text similarity matching,and machine reading comprehension.It is based on deep learning algorithms and supplemented by other algorithms and it is an intelligent QA system that uses multi-source knowledge to answer user questions about product information.The system learns from the strengths of each knowledge source,and tries its best to answer user questions,which can better balance factors such as system effect and cost.This paper designs three QA modules for the three knowledge sources of product knowledge graph,FAQ database and product information document to answer users' questions based on the knowledge of the corresponding source.(1)Aiming at the knowledge source of the product knowledge graph,this paper designs a knowledge graph QA module,which identifies the entity in the user's question,then converts the user's question into a query command of the knowledge graph,calls the knowledge graph database and generates the answer;(2)Aiming at the knowledge source of the FAQ database,this paper designs a question similarity matching QA module,which calculates the similarity between the user's question and the FAQ in the FAQ database,and selects the most similar FAQ corresponding to the user's question.Then use the answer corresponding to the FAQ to answer user questions;(3)For the knowledge source of product information documents,this paper designs a reading comprehension QA module,which first sifts out the parts in the larger-scale unstructured documents that may be useful for answering user questions,and then use a more complex selection algorithm to select the final answer.A QA fusion module is designed in this paper to integrate the above-mentioned QA modules based on different knowledge sources.According to the characteristics of each knowledge source and the output of each module,the final answer is decided and output.The data used in this paper is strongly related to the product documentation and is inconsistent with the existing published data set,so the public data sets cannot be used to train and verify the models.In order to train and verify the multiple models designed in this paper,multiple corresponding data sets are also designed and produced in this paper.