Research And Implementation Of Semantic Matching Technology For Intelligent Question Answering System

Semantic matching technology has attracted much attention and has become one of the current hot topics in the application field of natural language processing technology.It has a wide range of application scenarios in the fields of question answering systems and information retrieval.At present,the most popular semantic matching model is the fine-tuning model based on BERT,but most of the semantic matching technology based on BERT model adopts a unified attention mechanism,which is not sufficient for extracting semantic information from text with complex sentences,which leads to an understanding of sentence semantics.There are deviations;at the same time,the BERT model is large in scale and the amount of calculation is really amazing.The cost of calculating a single sample can be hundreds of milliseconds,which cannot meet the requirements under strict delay constraints,and it is difficult to apply to actual production.Therefore,in view of the above described defects,this thesis proposes two different semantic matching models based on the BERT fine-tuning model.The main research contents of this article are as follows:1.Aiming at the problem that the BERT fine-tuning model adopts the unified attention mechanism and fails to fully consider the sentence structure information,this thesis proposes a multi-channel attention mechanism semantic matching model based on the Tree-LSTM structure.The model consists of a three-layer structure: BERT model,Tree-LSTM structure and multi-channel attention mechanism.Among them,the BERT model converts the input text into a feature vector to prepare for the subsequent work;the Tree-LSTM structure maps the feature vector output by the BERT model to a tree structure and distributes it on each child node of the tree;multi-way attention.The force mechanism assigns different weights to each child node in the Tree-LSTM structure,so that different components of the sentence vector get different attention.Finally,the model in this thesis is applied to the sentence semantic similarity calculation link,and the semantic similarity value between sentence pairs is calculated through the semantic feature vector representation of sentence pairs.Experiments show that the model in this thesis is more accurate than traditional semantic matching models on data sets in the financial field.2.Aiming at the situation that the BERT model is too large and consumes too much time and resources when doing semantic matching tasks,this thesis proposes a semantic matching model based on the twin network of multi-channel attention mechanism.The model is composed of three algorithms: BERT Chinese pre-training model,Siamese network and Bi-LSTM-based multi-channel attention mechanism.In this model,first obtain the vector information of the two input sentences through the BERT pre-training model,and take the pooling operation to reduce the dimension of the sentence vector;then use the Bi-LSTM-based multi-channel attention mechanism to compare the difference of the sentence vector The components give different attention to the sentence structure information to make the sentence structure information clearer;finally use the twin network to derive a meaningful and fixed-length sentence vector,and use the cosine similarity method to calculate the semantic similarity between sentence pairs,and then according to the similarity score Choose the most similar pair of sentences.Through experimental verification,the semantic matching speed of the model in this thesis is faster.3.This thesis implements an intelligent question answering system based on the two semantic matching models studied above.The system interacts with users through a visual interface to automatically respond to questions raised by users and achieve practical application purposes.