| In recent years,the number of web pages on the Internet has grown rapidly,exposing users to the problem of information overload.The search engines are important tools to help users retrieve these web pages,so research and improvement of the summary generation of search engine can enhance both the experience and search efficiency of search engines.This thesis investigates the generative abstracting technique based on Transformer,and proposes the LS-ESS-TF model from the perspectives of keyword attention and probability distribution;and the CCDM-CN model is proposed for the problem of controlled text generation of Chinese abstracts;finally,a search engine system is constructed based on the above models.The details of this thesis are as follows.Firstly,this thesis introduces the research background and significance of the topic,followed by the introduction of the current state of research on automatic summarization,controlled text generation and search engines.Secondly,the relevant technology fundamentals are introduced,such as Transformer,text vectorization,and so on.Thirdly,awaring the problem of,when using Transformer to generate summaries,insufficient extraction ability of key information phrases of short length and the problem that Beam Search easily degrading to Greedy Search,due to the maximum value of the model output probability distribution tending to be much larger than other values.Also,the problem of reduced information extraction ability of multi-headed attention due to low-rank phenomenon at longer text input.A novel generative summarization model,i.e.,LS-ESS-TF,is proposed based on Transformer.The model proposes and uses MH-CSSA to enhance the extraction ability of short textual information,using a bi-path decoder to optimize the probability distribution of the model output,and using Talking-Head mechanism to optimize the low-rank phenomenon of multihead attention.The experimental results of the LS-ESS-TF model on the LCSTS dataset show that its performance is enhanced to some extent in ROUGE metrics compared to some models including Transformer.Fourthly,aiming at the problems of fluency and accuracy degradation of summarization in the implementation of Chinese summarization task based on the Pn P approach to control text generation,and the problem that greater semantic granularity in Chinese leading to failure of guaranteeing the control semantic integrity.A new controlled text generation model,i.e.,CCDM-CN,is proposed.In the Chinese summary task,the model uses an Ex-Trie tree to ensure the fluency of the controlled generation content and uses a control word filtering model to select reasonable control words as well as Text Rank-based calculation of control word priority and descending sorting to alleviate the accuracy degradation of the summary in the Chinese summary task.The test results on the constructed LCSTS dataset containing control words show that the CCDM-CN model can achieve a high success rate of control text generation with a lower sacrifice of accuracy and fluency.Finally,based on the above proposed generative abstraction algorithm LS-ESS-TF and controlled text generation algorithm CCDM-CN,a search engine system is designed and implemented by using My SQL database and the inverted index,Torch Serve deep learning model service framework,Django,Vue framework and Page Rank algorithm,and the operation effects of the system are demonstrated. |
