Research On Automatic Recognition And Application Of Moves In Scientific Articles

As the medium through which researchers analyze,study and elaborate on their own research field scientifically,scientific article is an important form of sharing scientific and technological achievements.However,in recent years,with the rapid development of science and technology,the number of scientific articles has suddenly increased.The quality gap of different articles,the massive amount of articles combined with the varying structures of articles between different journals has caused some difficulties for users,and at the same time,the information needs of article readers are becoming more diverse and precise: specific research methods,specific experimental processes and research innovation can all be the searched by a researcher,and the imbalance between literature supply and demand have become increasingly acute.The Rich HTML format of articles used by various publishers have improved the reader's reading experience by adding navigation pointing to individual chapters and similar articles recommendation information to the article web pages,but the existing navigation and recommendation module still cannot satisfy the reader's more refined information needs.Linguistic researchers use move analysis to interpret a scientific articles from the perspective of the author,and the resulting move label,which implies author rhetorical strategy and communicative purposes,is of great reference value for article users to understand and choose articles.However,the traditional move analysis process is timeconsuming and laborious,there are relatively few related studies on automatic or semiautomatic identification of moves of scientific articles,and the recognition range is also limited to the abstract or introduction of the paper.Based on this,the aim of this research is to explore the possibility of automatic identification of moves in full articles and how the identified moves can be used to provide a more refined literature information service.First,this paper makes reference to the move classification framework established in the existing studies.After making move analysis of bodies of several chemistry articles,a move classification framework consisting of seven moves is proposed,and the reliability of the framework was verified in the labeling experiments.To obtain more annotation data for subsequent move automated recognition model training while reducing the complexity of the move annotation job,this paper uses an open source dataset of scientific articles tagged with core scientific concepts,and annotates the moves on the body of 225 chemical articles in detail using their original annotation results.After an analysis of the data on the usage of moves in the articles we found that the different moves differ significantly in terms of their relative quantities,relative position,and transition characteristics.Subsequently,based on the move classification framework proposed in this paper and the obtained move data,we construct various move automatic recognition models by machine learning technology and deep learning technology and compare their performance.Result suggests that machine learning methods are not sufficiently different from deep learning methods on their recognition performance,and because of the influence of model performance and secondary classification,the use of sequence information in move recognition can instead reduce the recognition accuracy.To remedy the drawback that deep learning models ignore the structural information beyond the text content,this paper proposes a method to improve the move recognition by integrating the implicit knowledge obtained by deep learning models into machine learning models.Compare with the original model the macro average F1 value of new model improved by about 7%,indicating that the multi-features fusion method can obviously improve the performance of move automatic recognition.Finally,we explore the possible application patterns of the identified move information in making literature information services,propose a navigating interface of scientific articles based on move information,a recommendation scheme of similar articles,and a module of scientific writing guidance.Compared with existing schemes,a move-based literature information service can help readers locate to information delicately,which is more conducive to the actual use of users.The move automatic recognition method proposed in this paper is well tested in chemistry articles,and the proposed move-based literature information service is also highly feasible and will be helpful for more extensive and deep research in the future.