Research On Abstractive Summarization Generation With Capsule Network

With the rapid development of deep learning,the domain of text summarization has also benefited from this and is in a period of vigorous development.However today's abstractive text summarization has developed so far,there are still problems with sentence redundancy/errors,insufficient content coverage of topics,low sentence fluency,low training efficiency and difficulty in deployment.Based on the research of abstractive summarization,this thesis combines the existing cutting-edge technology to analyze,design and implement the capsule-Attention model.On the basis of this model,we design and implement the microservice for highconcurrency composite architecture of abstractive summarization.The contribution and specific work of this thesis are as follows:(1)Aiming at the problem of insufficient content coverage and low sentence fluency,this thesis starts from the perspective of words based on the attention mechanism,and combines the attention mechanism to study the grammar and semantic structure of massive texts in a more fine-grained manner.for the semantic structure part of the model,capsule network is introduced,and a vector is formed by reorganizing the single-character text scalar,using the form of capsule transition to summarize the representation of a word.In addition,this thesis proposes a new routing function ?-Squash,which can better learn the key text information from the sparse matrix.(2)For sentence redundancy or errors and omissions,this thesis introduces.In addition,the algorithm also introduces a pointer mechanism and an overlay mechanism,which can effectively remove duplicate information and learn summary data from the source text.The algorithm combined with the attention capsule network not only considers the grammatical and semantic differences between different languages,but also generates brand-new sentences based on the meaning of the original text.In addition,combined with the corresponding synonyms and synonym scenarios,it can also perform summary output easily and solve the problem.The attention shift caused by the word scene change under the attention mechanism is solved.(3)As for the training efficiency,this thesis also designs and implements an attention capsule network based on the concept of parameter sharing,thereby reducing the training parameters of the model.Combining the vector features of the capsule mechanism to further compress the characterization space of words,reduce the waste of calculation space,and improve training efficiency.Based on the existing Chinese and English data sets,this thesis conducts corresponding comparative experiments.Combined with the ROUGE scoring results,it can be seen that the abstract content generated by this algorithm is more in line with human language description habits.The entire algorithm has relatively good performance on existing data sets.From the perspective of training efficiency,although the algorithm requires a large number of parameters for training,the corresponding total training time of this algorithm is much shorter than that of RNN algorithms,and the number of parameters is much smaller than BERT.Judging from the experimental results and considering many aspects,the capsule-attention pointer model proposed in this thesis has a certain advanced nature,and the work of this thesis has certain practical and design significance.Finally,based on the CAP theory,this thesis designs and implements the microservice composite architecture of the corresponding model.Considering the provision of synchronous and asynchronous invocation of the service interface in the concurrent scenario,the high availability and strong consistency of the service are guaranteed,and the corresponding design is completed.Idea statement and experimental process,and functional display of the corresponding results.