Research On Reviewers Assignment Problem Based On Field Representation

With the explosive growth of human scientific research needs and scientific research results,a large amount of scientific literature needs to be published.Faced with such many manuscripts,that is a common challenge assigning appropriate reviewers to a manuscript from a pool of candidate reviewers in the academic community.The traditional method of finding reviewers for manuscripts cannot cope with such many manuscripts and reviewers in the complicated field,which will result in extremely high cost and error rate.Since the twenty-first century,increasing journals and conferences have adopted the method of automatically reviewer assignment.Like manually assigning reviewers for manuscripts,the key to automatically assigning reviewers is the identification of manuscript and reviewer fields,as well as precise field matching between manuscripts and reviewers.This kind of research on field identification and matching can bring other application values,such as question answerer matching,expert find,natural language understanding tasks,and so on.The research on the reviewer assignment has been based on the research results of natural language processing,and on this basis,improvements have been made to adapt to the characteristics of the reviewer assignment.Since the identification and matching of the field are very important for the reviewer's assignment,the latent semantic indexing method in natural language processing is proposed,and the research of the reviewer assignment begins to develop.The identification and matching of fields usually take the form of explicit fields,that is,direct representation of specific fields through specific information,such as methods based on topic models and language models.However,in the actual assignment,since the reviewer information consists of multiple papers,and the manuscript to be reviewed consists of only one paper,there is an imbalance with respect to the field information between the reviewer and the target manuscript.This imbalance manifests itself in two ways:first,each field of the reviewer(not necessarily the best field)corresponds to more textual information than the manuscript fields,and second,the reviewer has more fields than the target manuscript.This dissertation reduces the impact of textual information imbalance by improving the similarity calculation between reviewers and manuscripts and reduces the imbalance of field information by coupled random walk to the similarity between reviewers and manuscripts and between reviewers' papers and manuscripts.On the other hand,the current field mining method can only obtain the approximate field of the paper from a general method,and this kind of approximation field is difficult to match the specific paper field.This dissertation uses the consistency of the title and abstract fields in the paper as the supervised information to learn in the implicit field,thus avoiding the need to subjective empirical hypothesis(inductive bias)to infer the characteristics of the field.The main work of this dissertation is as follows:1)This dissertation proposes a word and semantic-based iterative model(WSIM).The method treats manuscripts and reviewers as a collection of one and multiple papers and uses topic models and language models to extract the field characteristics of manuscripts and reviewers.Firstly,in the process of extracting word information using language model,this dissertation effectively captures the importance of certain low-frequency words and reduce the weight of insignificant high-frequency words,which not only highlights the specificity of word information,but also effectively combine with semantic information.Then,the normalized damage cumulative gain(NDCG)is introduced as the similarity calculation method to solve the problem that is the imbalance of the reviewer and the manuscript on the text information.Finally,this dissertation solves the problem,the imbalance of reviewers and manuscripts in all areas,by iterating the field features between reviewers and reviewers' papers in a coupled random walk.In addition,this dissertation compares with seven methods on the real data set,and the experimental results verify the effectiveness of the method.2)This dissertation proposes a sentence pair modeling-based reviewer assignment(SPM-RA),The method uses the consistency of the title and abstract fields in the paper as supervision information,so that the neural network model that needs supervision information can be used to learn the field relationship between the papers.There is very strict logic between the title and abstract of the paper,so this dissertation assumes that the title of the paper and the abstract are consistent,so that the relation between the title and abstract in the paper can be used as knowledge,avoiding the disaster of lack of label in the reviewer assignment problem.First,the relation between the title of the paper and the abstract is trained using a convolutional neural network(CNN)and BERT(bidirectional encoder representations from transformers).Then,the similarity between the different titles and abstracts is used to obtain the similarity among the papers,and finally,the reviewers are assigned for the manuscript.The final experiment shows that the method is efficient and feasible.Experiments on real data sets show that this method is better than WSIM.