Research On Navigation Generation Of Argumentative Zoning For Acadamic Text Based On Deep Learning

In the era of big data,the publishing speed of academic texts is getting faster and faster,and the amount of academic resources has grown explosively.Therefore,more and more researchers are trying to research the semantic text of academic text from different dimensions to solve the current problem of information overload.At present,researches in fields such as information retrieval and citation analysis have not fully utilized the structural information of academic texts,and have rarely incorporated textual structure information into the semantic mining of academic texts.The theoretical system of argumentative zoning selects the dimension of “argumentation” to explore the relationship between the structural nature of academic texts and content information,which is very suitable for solving some existing problems.In order to improve the current lack of effective use of textual structure information in academic text mining research,and at the same time to better solve the problem of information overload,this paper uses the theoretical system of the argumentative zoning,combined with deep learning technology,to apply it to the academic text navigation generation task.In detail,this task is divided into two steps,including the argumentative zoning classification and identification task and the argumentative zoning navigation generation task.Aiming at the argumentative zoning classification and identification task,this paper proposes a argumentative zoning recognition model based on hierarchical attention mechanism,and compares it with the traditional SVM model and LSTM model.Finally,the model proposed in this paper has achieved the best effect,and its F1 value reached 0.90.For the argumentative zoning navigation generation task,this paper proposes different extractive and abstractive models based on the pre-training model BERT,and compares it with the traditional Text Rank algorithm,and then uses ROUGE index to evaluate.Finally,this paper finds the extractive model using Transformer reached the best results,its ROUGE-1,ROUGE-2 and ROUGE-L values reached 41.30,18.36 and 31.35,while the abstractive model using the Encoder-Decoder framework,its ROUGE-1,ROUGE-2 and ROUGE-L values of 40.70,18.07 and 28.35,also showed a certain potential.At the end of this paper,the prototype design of the academic text argumentative zoning navigation reading system is added,to better show the practical application value of this study.