Research On Computation Method Of Semantic Similarity In Financial Field Based On Deep Learning

With the rapid development of the digital economy and the gradual maturity of artificial intelligence technology,semantic similarity calculation in the financial field has become a hot research topic in the field of natural language processing.Semantic similarity calculation aims to measure the degree of semantic similarity between texts.It is a fundamental but important task in natural language processing,and play a vital role in tasks such as information retrieval,intelligent question answering,and text summarization.As people’s demands increase,the requirements for semantic similarity calculation technology are becoming increasingly sophisticated.This paper mainly studies how to improve the accuracy of text similarity algorithm in the financial field.In most scenarios in the financial field,data appears in the form of short texts,and traditional neural network calculation methods face challenges in capturing local text features and identifying keywords accurately.In this paper,we propose two improved semantic similarity calculation methods by addressing some of the existing problems in current methods,focusing on text preprocessing,neural network structure,and attention mechanisms.The research results of the paper are as follows:(1)Caps-BiLSTM network with part-of-speech(POS)merging.Aiming at the problem of word segmentation errors when the general word segmentation method is applied in the financial field,a POS merging model is proposed to reduce the impact of segmentation errors on similarity judgment and improve the training effectiveness of the neural network model.At the same time,Capsule Network was combined with BiLSTM to extract the local features of the text,and judge the semantic similarity of the text from both global and local dimensions,which solved the problem that the traditional neural network structure could not capture the local features of the text.(2)RCNN network with head word attention mechanism.The model is improved on the RCNN network structure by using Bidirectional Gated Recurrent Unit Network(BiGRU)to learn semantic features of two short texts.After applying the head word attention mechanism and character concatenation,the model obtains the semantic enhanced representation of the two sentences.Then,1D Convolutional Neural Network(1DCNN)is used to integrate the embedded information with contextual information.Finally,the semantic expression of the sentence is completed by extracting the information with the max pooling.In the experimental verification section,the Chinese dataset used is ATEC provided by Ant Financial,while the English dataset used is the Microsoft Research Paraphrase Corpus(MSRP).Through experimental analysis,the Caps-BiLSTM network combined with the partof-speech merging method proposed in this paper is more accurate in large-scale sample datasets,while the recall rate is lower in small sample data sets..In contrast,the RCNN network with head word attention mechanism performs well in small-scale sample datasets.Therefore,in practical application scenarios,different algorithm models can be chosen for training according to the size of the dataset to achieve better results.