Literature recommendation system and literature recommendation method

Abstract

The present invention provides a document recommendation system and a document recommendation method. The system includes: a document information acquisition module, which is used to acquire document information; a semantic calculation module, which calculates the semantic similarity of the document based on the document abstract, and sorts the document according to the semantic similarity; a subject evaluation module, which classifies the document based on the subject words of the document, and sorts the document according to the subject word similarity, wherein the subject words are determined based on the document abstract and the medical subject words; an influence analysis module, which determines the influence of the document based on the influence score of the document and the influence score of the author, and sorts the document based on the influence; a knowledge graph construction module, which constructs a knowledge graph based on the above-mentioned document sorting results; and a comprehensive display module, which is used to output the results in the form of a visual graph. The application of this system for document recommendation can fully display the comprehensive information of the document, improve the accuracy of the recommendation, and make the recommended content more intuitive.

Description

Technical Field

[0001] The present invention relates to the field of biomedicine, and in particular to a document recommendation system and a document recommendation method. Background Art

[0002] While the field of life sciences has made great progress in recent years, it has also faced many problems. On the one hand, the output of academic papers has been growing exponentially, and the excessive amount of information has made it increasingly difficult for researchers to find literature, track focus, and see the relationship between literature. On the other hand, the lack of uniformity in information or knowledge description has led to a lack of relevance. How to efficiently find and use these new research results has become a major problem for scientific researchers.

[0003] The literature recommendation system aims to help researchers reduce information overload, find relevant papers through calculation and ranking, and recommend the top N papers with high relevance to the current paper. Different literature recommendation systems often use different literature recommendation methods, including content-based recommendation methods, citation-based recommendation methods, and recommendation methods based on a mixture of the two. The content-based recommendation method uses a traditional machine learning method, which obtains the tf-idf vectors of the title, journal, and medical subject terms of the paper respectively, and then splices these vectors together as the representation vector of the paper. After obtaining the vector, clustering is performed, and then the cosine similarity between the two papers is calculated with the papers of the same cluster class for sorting, and the top N papers with high similarity are taken as the recommendation results. The recommendation method based on citation relationship is mainly based on the citation relationship network between papers, and the recommendation results are obtained through the weight model or probability model of the citation relationship. The hybrid method based on these two methods combines the two recommendation methods of content and citation relationship to obtain a comprehensive literature recommendation list. There are traditional recommendation algorithms based on keyword retrieval, content and feature similarity, or recommendations based on citation relationship network.

[0004] Although these methods can give some literature as recommendation results, they are not effective, and due to incomplete description information and lack of relationships between literature, it is difficult for researchers to have a comprehensive and clear idea.

[0005] The recommended methods in the literature need further improvement. Summary of the invention

[0006] The present invention aims to solve one of the technical problems in the related art at least to a certain extent. To this end, the present invention provides a document recommendation system and a document recommendation method.

[0007] The rapid development of science and technology is accompanied by the large output of academic papers. The excessive information recorded in academic papers makes it increasingly difficult for researchers to find literature, track focus, and see the relationship between literature. Although some literature recommendation methods can provide some recommendation results for researchers to a certain extent, the effect is not good. Several commonly used recommendation methods are used as examples to illustrate. Content-based recommendation methods usually use tf-idf vectors as the representation vectors of papers, which do not contain semantic associations, resulting in the lack of semantic representativeness of the calculated similarity results, resulting in low accuracy; in the case of a large number of papers, the obtained vectors are high-dimensional and sparse, and the calculation speed is slow; in the clustering process, due to the large number of papers, the clustering process is slow and it is difficult to grasp the number of clusters. Most of the recommendation methods based on citation relationships only consider the citation relationship between papers to recommend, and occasionally combine the recommendation method with the paper-author relationship, but generally do not consider more semantic information such as paper title, theme, abstract, etc. Moreover, whether it is content-based, citation-based, or hybrid methods, only a list of literature is given in the end. Researchers can only roughly judge whether they are interested based on the titles of the literature list, and it is difficult to form a comprehensive understanding.

[0008] Therefore, whether it is a document recommendation system or a document recommendation method, they generally only provide one-way information and lack description of the overall information, which is not conducive to the overall judgment of scientific researchers. To this end, the inventors of the present invention creatively provide a document recommendation system and a document recommendation method during the research process. The method creatively combines more document-related information, uses knowledge graph technology to sort out and integrate information, and finally provides comprehensive information recommendation content and interactive visual display, so that the recommended content is clear at a glance.

[0009] Since Google proposed the concept of knowledge graph in 2012, knowledge graph technology has been well applied in assisting information fusion, association recommendation, reasoning and decision-making, etc. Due to the advantages of knowledge graph technology in information integration, information association and unified retrieval, we can more conveniently integrate multi-dimensional information related to documents, thereby implementing the present invention.

[0010] In a first aspect of the present invention, the present invention provides a document recommendation system, which includes: a document information acquisition module, which is used to acquire basic information of a document; a semantic calculation module, which calculates the semantic similarity of a document based on the abstract of the document, and sorts the document according to the degree of semantic similarity; a subject evaluation module, which includes the prediction of subject terms and the matching degree between a document and its annotated subject terms, the subject evaluation module classifies the document based on the subject terms of the document, obtains the classified document, and sorts the document according to the degree of subject term similarity, the subject terms are determined based on the abstract of the document and the medical subject terms; an influence analysis module, which determines the influence of the document based on the influence score of the document, the influence score of the author, the influence score of the document in its category after classification, and the influence score of the author of the document in its category after classification, and sorts the document based on the degree of influence; a knowledge graph construction module, which constructs a knowledge graph based on the document sorting results of the semantic calculation module, the subject evaluation module, and the influence analysis module.

[0011] The document recommendation system provided by the present invention, after obtaining the relevant data of the document by using the document information acquisition module, first represents the semantic vector based on the document abstract, and sorts the documents according to the semantic similarity through similarity calculation; on the other hand, the document is classified in combination with the abstract and medical subject words, and the document is sorted according to the similarity of the subject words; in addition, according to the document citation relationship and the relationship between the document authors, for different subject words or categories, the author and the corresponding document are evaluated for their influence in the field or category, and the documents are sorted based on the influence. Then, these sorting results are integrated into the attribute graph in the form of node-attribute-relationship with the help of the technology of knowledge graph. Finally, the final data is pushed and rendered to the front end, and the multi-dimensional comprehensive recommendation content related to the document is intuitively displayed in the form of a graph.

[0012] By using the document recommendation system provided by the present invention to recommend documents, it is not only possible to obtain highly relevant recommended documents, but also to obtain the relationship between the subject word classification and the relevance, and recommend highly influential authors and high-quality documents in related fields; and these important information can be efficiently and simultaneously displayed in a relationship map, so that scientific researchers can more intuitively obtain the overall picture of the research field related to the document. Therefore, it can be applied in document retrieval, document recommendation, document relationship analysis, research field tracking, scientific research expert tracking, etc.

[0013] According to an embodiment of the present invention, the above-mentioned document recommendation system may further include the following technical features:

[0014] According to an embodiment of the present invention, the semantic calculation module uses the BERT language representation model to generate the representation vector of the document, thereby quickly calculating and obtaining the semantic similarity of the document.

[0015] According to an embodiment of the present invention, the semantic computing module sorts the documents by the following method: (1) pre-training the BERT language representation model using a document abstract word library larger than 4000M to obtain a BERT pre-trained language model; (2) constructing a representation vector containing semantic information based on the BERT pre-trained language model and the abstract of the document; (3) obtaining the semantic similarity of the document through cosine similarity calculation based on the representation vector, and sorting the document based on the semantic similarity.

[0016] According to an embodiment of the present invention, a method for calculating the semantic similarity of documents using cosine similarity calculation is as follows:

[0017] X and Y are n-dimensional vectors, and the calculation formula of cosine similarity is as follows:

[0018]

[0019] The closer the cosine value is to 1, the closer the angle is to 0 degrees, that is, the more similar the two vectors are, and the angle is equal to 0, that is, the two vectors are equal.

[0020] The corresponding editing formula in Latex is: \cos\left(\theta\right)=\frac{\sum_{i=1}^{n}x_i\times y_i}{\sqrt{\sum_{i=1}^{n}x^2}\times\sqrt{\sum_{i=1}^{n}y^2}}.

[0021] The "4000M literature abstract word library" in the present invention refers to the corpus size of literature abstracts. The BERT pre-trained model here uses the BERT model pre-trained by PubMed literature provided by NCBI, so that it can extract semantic information in the vertical field of PubMed literature better. For details, please refer to the reference: Transfer Learning in Biomedical Natural Language Processing: An Evaluation of BERT and ELMo on Ten Benchmarking Datasets ( https: / / arxiv.org / pdf / 1906.05474.pdf ).

[0022] According to an embodiment of the present invention, the topic evaluation module determines the subject words of the document based on a bidirectional LSTM recurrent neural network structure model and a regression model, or based on a BERT model and a sigmoid layer. Preferably, the topic evaluation module determines the subject words of the document based on a BERT model and a sigmoid layer. According to an embodiment of the present invention, the topic evaluation module specifically includes: training a subject word classification model based on the document containing complete subject words as training data; predicting the subject words of the document missing subject words based on the trained subject word classification model to obtain predicted subject words; performing a relevance evaluation based on the abstract of the document missing subject words and the predicted subject words to obtain the correlation weight between the subject words and the document.

[0023] According to an embodiment of the present invention, the method for training the keyword classification model includes not freezing all parameters of the BERT network layer, not freezing the network parameters of the last two layers of BERT, and freezing at least one of the BERT network parameters. Preferably, the method for training the keyword classification model is not freezing the network parameters of the last two layers of BERT.

[0024] According to an embodiment of the present invention, after the subject word classification model is trained, for documents missing subject words, their abstract information is used as the input of the model. After calculation, the model will output the probability of each subject word. The probability value is sorted, and a threshold can be set manually to take the subject words with a probability greater than the threshold of 0.5 as the subject word subset of the document.

[0025] According to an embodiment of the present invention, the relevance between the document and the marked subject words is evaluated by extracting the keywords of each document through the abstract and title of the document by the textRank method, and then measuring the relevance between the document and the subject words by calculating the distance between the keywords and the subject words. The greater the distance, the smaller the relevance.

[0026] According to the embodiment of the present invention, the prediction of missing keywords is mainly obtained through the keyword classification model. The relevance evaluation analysis is to calculate the relevance weight between the keyword and the document, and will not filter out the keyword predicted by the classification model. When displaying in the result graph, when recommending the top k documents by keyword, the relevance weight and the document influence weight will be used for weighting and sorting, and then the top k will be taken.

[0027] The "subject words" in the present invention can be subject words in various fields. In the embodiment of the present invention, it is mainly based on medical subject words for screening and integration, which belongs to the subset of medical subject words. The subject evaluation module mainly includes two parts: one is to predict the subject words of the literature with missing topics, and the other is to evaluate the relevance between the literature and the marked subject words. The specific steps are as follows:

[0028] 1) The subject word classification model is a typical multi-label classification model. In the data preprocessing stage, the abstract and mesh information of the document are extracted. According to the hierarchical relationship of the mesh words, they are integrated and the meshes of focus are selected as subject words. The sample distribution of each subject of the original data is unbalanced, and there are too many documents for some subject words. The random sampling method is used to process it to solve the problem of unbalanced sample distribution. In the multi-label classification model construction stage, the BERT pre-training model plus a single sigmoid layer is used to construct it. Each neuron in the sigmoid layer corresponds to the probability of each subject word. The input is the abstract information of the document. The semantic information of the abstract is extracted by BERT and sent to the sigmoid layer to calculate the probability corresponding to each subject word, and then compared with the real label probability of the document. The goal is to make the calculated probability distribution infinitely close to the probability distribution of the real sample. At this time, cross entropy is used as a measurement standard. The smaller the cross entropy, the closer the two distributions are. Therefore, the learning goal of the network is to minimize the cross entropy loss function. The gradient descent method is used to perform gradient backpropagation through backpropagation to update the weight value of the network. The model is repeatedly trained in this way to minimize the loss function and eventually enable the model to have the correct classification ability. The inventors performed fine-turning downstream tasks based on BERT. During the training process, several methods were tried to train, including not freezing all parameters of the BERT network layer, not freezing the network parameters of the last two layers of BERT, and freezing the BERT network parameters. It was found that the overall performance of not freezing the network parameters of the last two layers of BERT was better.

[0029] 2) Evaluate the relevance between the article and the marked subject words: mainly through the article abstract and title, extract the keywords of each article through the textRank method, and then measure the relevance between the article and the subject words by calculating the distance between the keywords and the subject words. The greater the distance, the smaller the relevance.

[0030] According to an embodiment of the present invention, the influence score of the document is obtained by using a weight model based on the number of citations, publication year and impact factor of the document.

[0031] The formula for calculating the influence of a document is:

[0032] Score=jif*cited_cnt*date_factor

[0033] date_factor=0.9^(max(0,cur_year-year) / 5)

[0034] Among them: jif is the impact factor of the journal in which the document is published; cited_cnt is the number of citations of the document; cur_year is the current year, year is the year the document is published; date_factor is the date impact factor of the document publication time. The date impact factor decays at a rate of 10% every 5 years.

[0035] According to an embodiment of the present invention, the influence score of the classified document in its category is obtained by using a weight model based on the number of citations, publication year and impact factor of the document in its category.

[0036] According to an embodiment of the present invention, the author's influence score is obtained by calculating the h-index based on the number of papers published by the author, the number of paper citations and the author's contribution.

[0037] According to an embodiment of the present invention, the influence score of the author of the classified document in its category is obtained by calculating the h-index based on the number of papers published by the author, the number of paper citations and the author's contribution.

[0038] The contribution is mainly determined by the author's position in the author list. all is the number of all authors in the document.

[0039] if all>4:

[0040] The contribution of the first two authors and the last two authors in the author list is 0.2 respectively, and the contribution of the remaining authors is 0.2 / (all-4);

[0041] if all<=4:

[0042] The contribution of each author is: 1 / all.

[0043] The number of citations for each paper is calculated by multiplying the inverse of the paper's collaboration degree by the actual number of citations of the paper, that is, h m The author's influence is measured using the pf index. The calculation formula is:

[0044]

[0045] Among them, C f =Σr i c i , P f =Σr i , a i is the cooperation degree of the authors of the ith paper, r i The share (fraction) of honor given to the author r i =1 / a i . See specifically: http: / / blog.sciencenet.cn / blog-335532- 505453.html .

[0046] According to an embodiment of the present invention, in the influence analysis module of the present invention: 1) The influence of a document is mainly a weight model constructed by the number of citations of the document, the impact factor of the journal in which it is published, and the year of publication, which has nothing to do with the subject area in which it is located; 2) The influence of the author is mainly calculated by the number of papers published, the number of citations of the papers, and the author's contribution to the document; 3) After classification, for a single category, the top 100 influential documents and authors in the category are calculated. According to an embodiment of the present invention, the knowledge graph construction module is constructed based on documents, authors and keywords as three types of nodes. According to an embodiment of the present invention, the document recommendation system also includes a comprehensive display module, which is used to output document recommendation results, and the document recommendation results are presented in the form of a visual map.

[0047] The visual graph display can more clearly show the relationship between documents, authors, and topics, and users can interact on the graph. If they are interested in a certain node, they can click on the node to obtain related recommended information centered on the node.

[0048] According to an embodiment of the present invention, the visualization map includes at least one of the following: author, subject word and document. In the second aspect of the present invention, the present invention provides a document recommendation method, comprising: using the document recommendation system described in any embodiment of the first aspect of the present invention to recommend documents. In the existing document recommendation systems, documents are basically recommended from one dimension in a relatively simple manner, and the multi-dimensional recommendation strategy mentioned in the present invention is not adopted, and multi-dimensional relationship display is performed based on the knowledge map. Moreover, the document recommendation system of the present invention is used for document recommendation, and innovative optimization and integration are made on the basis of existing research methods, and the knowledge map technology is combined to sort out and infer the mined attributes and relationships. For example, in order to help researchers understand a document more efficiently, not only can recommendations of certain related documents be obtained, but also the relationship between its subject classification and relevance can be obtained, and high-impact authors and high-quality documents in related fields can be recommended, and these important information can be displayed in a map, so that researchers can more intuitively obtain the overall relationship of the document.

[0049] The beneficial effects achieved by the present invention are as follows: the document recommendation system and document recommendation method provided by the present invention perform relationship mining and recommendation from multiple dimensions, comprehensively display the comprehensive information of the document, improve the accuracy and availability of the recommendation, and expand the imagination space of the information. Furthermore, the present invention uses the BERT language representation model to perform semantic similarity calculations, which reduces the computational complexity and improves efficiency and accuracy. Moreover, the present invention adopts the knowledge graph method, so that the document recommendation system provided can be used for searching and inference quickly and efficiently, and is presented in the form of a visual graph, so that the results are more direct, the recommended content is rich and comprehensive, and the results are clear at a glance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the structure of a document recommendation system provided according to an embodiment of the present invention.

[0051] Figure 2 This is a result diagram of document recommendation using a document recommendation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The embodiments of the present invention are described in detail below. It should be noted that the described embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0053] like Figure 1 As shown, the present invention provides a document recommendation system, including: a document information acquisition module, a semantic calculation module, a theme evaluation module, an influence analysis module, a knowledge graph construction module and a comprehensive display module, wherein the document information acquisition module is used to acquire basic information of the document; the semantic calculation module calculates the semantic similarity of the document based on the abstract of the document acquired in the document information acquisition module, and sorts the document according to the degree of semantic similarity; the theme evaluation module classifies the document based on the subject words of the document acquired in the document information acquisition module, and sorts the document according to the degree of subject word similarity, and the subject is obtained based on the abstract and medical subject words of the document; the influence analysis module determines the influence of the document based on the influence score of the document, the influence score of the author, the influence score of the document after the subject classification of the document, and the influence score of the author of the document after the subject classification of the document, and sorts the document based on the degree of influence; the knowledge graph construction module constructs a knowledge graph based on the document sorting results of the semantic calculation module, the theme evaluation module, and the influence analysis module respectively; the comprehensive display module is used to output the document recommendation results, and the document recommendation results are presented in the form of a visual graph.

[0054] When using the document recommendation system provided by the present invention to recommend documents, it is specifically carried out according to the following modules, including: document information acquisition module, semantic calculation module, subject evaluation module, influence analysis module, knowledge graph construction module and comprehensive display module.

[0055] The document information acquisition module is used to obtain basic information of documents. For example, it mainly obtains information of scientific research papers through web crawlers, including but not limited to: title, abstract, keywords, medical subject terms (Mesh), author, author institution, citation relationship, number of citations, journals, references, etc. The document information acquisition module updates information regularly to ensure that the information is complete.

[0056] The semantic calculation module used calculates the semantic similarity of the document based on the abstract of the document, and sorts the document according to the degree of semantic similarity. The present invention adopts the BERT language representation model to generate the representation vector of the document. BERT pre-trains deep bidirectional representation by jointly adjusting the context in all layers. The learned vector has the semantic information of the context and is suitable for the construction of the most advanced model for a wide range of tasks. Specifically, on the basis of BERT, the BERT model is pre-trained using a thesis abstract word library greater than 4000M to obtain a BERT pre-trained language model, so that it has more semantic representation information in the field of scientific research papers. The abstract information of the paper is used as the input of the BERT pre-training model to obtain a low-dimensional representation vector containing rich semantic information. Based on the representation vector, the semantic similarity of the document is obtained by cosine similarity calculation, and the document is sorted based on the semantic similarity.

[0057] The calculation method of obtaining the semantic similarity of documents using cosine similarity is as follows:

[0058] X and Y are n-dimensional vectors, and the calculation formula of cosine similarity is as follows:

[0059]

[0060] The closer the cosine value is to 1, the closer the angle is to 0 degrees, that is, the more similar the two vectors are, and the angle is equal to 0, that is, the two vectors are equal.

[0061] The corresponding editing formula of Latex: \cos\left(\theta\right)=\frac{\sum_{i=1}^{n}x_i\timesy_i}{\sqrt{\sum_{i=1}^{n}x^2}\times\sqrt{\sum_{i=1}^{n}y^2}}.

[0062] The subject evaluation module determines the subject words of the document based on a bidirectional LSTM recurrent neural network structure model and a regression model, or based on a BERT model and a sigmoid layer. The subject evaluation module specifically includes: training a subject word classification model based on the document containing complete subject words as training data; predicting the subject words of the document missing subject words based on the trained subject word classification model to obtain predicted subject words; and performing a relevance evaluation based on the abstract of the document missing subject words and the predicted subject words to obtain the relevance weight between the subject words and the document.

[0063] The mentioned subject evaluation module includes the expansion and scoring of medical subject terms. Some documents have medical subject terms, while some documents lack medical subject terms.

[0064] In one embodiment of the present invention, for documents lacking medical subject terms, a model is constructed by a multi-task learning method, mainly based on the abstract information of the document as input, that is, the model includes two tasks, namely, multi-label classification, mainly using a bidirectional LSTM recurrent neural network structure, and the other task is a regression model, namely, predicting the number of labels of the paper. Then, a large number of documents with complete medical subject terms are used as training data to train the model, and after obtaining the model, subject terms are predicted for documents lacking subject terms. Then, the abstracts of the documents lacking subject terms and the predicted subject terms are combined to perform a relevance evaluation, and subject terms with high relevance evaluation are selected as the subject terms of the documents lacking subject terms.

[0065] In another embodiment of the present invention, the subject word classification model is a typical multi-label classification model. In the data preprocessing stage, the abstract and mesh information of the document are extracted. According to the hierarchical relationship of the mesh words, the meshes of focus are integrated and selected as subject words. The sample distribution of each subject of the original data is unbalanced, and the number of documents of some subject words is too large. The random sampling method is used to process it to solve the problem of unbalanced sample distribution. In the multi-label classification model construction stage, the BERT pre-training model plus a single-layer sigmoid layer is used to construct it. Each neuron in the sigmoid layer corresponds to the probability of each subject word. The input is the summary information of the document. The semantic information of the summary is extracted by BERT, and sent to the sigmoid layer to calculate the probability corresponding to each subject word, and then compared with the real label probability of the document. The goal is that the calculated probability distribution is infinitely close to the probability distribution of the real sample. At this time, cross entropy is used as a metric. The smaller the cross entropy, the closer the two distributions are. Therefore, the learning goal of the network is to minimize the loss function of the cross entropy, and the gradient descent method is used to perform gradient backpropagation through backpropagation to update the weight value of the network. The model is repeatedly trained in this way to minimize the loss function and ultimately enable the model to have the correct classification ability. The inventors performed fine-turning downstream tasks based on BERT. During the training process, the network parameters of the last two layers of BERT were not frozen.

[0066] In the topic evaluation module, the above-mentioned BERT pre-trained model plus a single-layer sigmoid layer model and the bidirectional LSTM recurrent neural network structure and regression model can meet the requirements in terms of the prediction of subject terms and the matching degree between the document and its annotated subject terms. However, on the same validation set, the F1 value of the model using the BERT pre-trained model plus a single-layer sigmoid layer is 0.02 higher than that of the model using the bidirectional LSTM recurrent neural network structure plus regression model. Evaluate the relevance of the article with the annotated subject terms: mainly through the abstract and title of the article, extract the keywords of each article through the textRank method, and then measure the relevance between the article and the subject terms by calculating the distance between the keywords and the subject terms. The greater the distance, the smaller the relevance.

[0067] After training the subject word classification model, for documents missing subject words, their abstract information is used as the input of the model. The model will output the probability of each subject word after calculation. By sorting the probability values, the threshold can be set manually, and the subject words with a probability greater than 0.5 can be taken as the subject word subset of the document.

[0068] The prediction of missing subject terms is mainly obtained through the subject term classification model. The relevance evaluation analysis is to calculate the relevance weight between the subject terms and the documents, and will not filter out the subject terms predicted by the classification model. When displaying in the result graph, when recommending the top k documents by subject terms, the relevance weight and document influence weight will be used and then sorted, and then the top k will be taken. The influence analysis module is mainly used to calculate the influence of the paper and the author. The calculation method of the document influence is mainly to build a weight model based on the number of citations, publication years, and impact factors of the journal. The author's influence can be calculated based on all the papers published by each author, the number of citations of the paper, the author's contribution, and other information. At the same time, according to the subject division of the documents in the above-mentioned subject evaluation module, the influence scores of the documents and authors in a certain field can be calculated separately. The influence of the document is determined by combining these factors, and the documents are sorted based on the level of influence. After classification, for a single category, the top 100 influential documents and authors in that category are calculated.

[0069] The formula for calculating the influence of a document is:

[0070] Score=jif*cited_cnt*date_factor

[0071] date_factor=0.9^(max(0,cur_year-year) / 5)

[0072] Among them: jif is the impact factor of the journal in which the document is published; cited_cnt is the number of citations of the document; cur_year is the current year, year is the year the document is published; date_factor is the date impact factor of the document publication time. The date impact factor decays at a rate of 10% every 5 years.

[0073] The contribution is mainly determined by the author's position in the author list. all is the number of all authors in the document.

[0074] if all>4:

[0075] The contribution of the first two authors and the last two authors in the author list is 0.2 respectively, and the contribution of the remaining authors is 0.2 / (all-4);

[0076] if all<=4:

[0077] The contribution of each author is: 1 / all.

[0078] The number of citations for each paper is calculated by multiplying the inverse of the paper's collaboration degree by the actual number of citations of the paper, that is, h mThe author's influence is measured using the pf index. The calculation formula is:

[0079]

[0080] Among them, C f =Σr i c i , P f =Σr i , a i is the cooperation degree of the authors of the ith paper, r i The share (fraction) of honor given to the author r i =1 / a i . See specifically: http: / / blog.sciencenet.cn / blog-335532- 505453.html .

[0081] The knowledge graph construction module mainly sorts out the information calculated by the above modules and constructs the corresponding relationship between them and the points and edges of the documents. For example, the structure of the attribute graph can be used to construct and store the corresponding nodes, attributes and relationships. Between the three main types of nodes, documents, authors and topics, different types of relationship edges are constructed in combination with the calculation results of the above modules, and the corresponding evaluation scores are recorded as a representation of the strength of the relationship. The higher the score, the closer the connection between the two.

[0082] After building the knowledge graph, you can use the information in the graph to make recommendations from multiple dimensions such as author, topic, and document, combined with corresponding parameter settings, to form a comprehensive display module. For example, you can design a visual graph that displays nodes and relationships. The search and calculation based on the graph are very fast and efficient. The recommended results of the search and calculation are returned to the visualization front end in real time through the API and rendered. The overall recommended content is intuitively displayed through visual elements such as the color and size of the node type, the relationship type, and the length / strength of the distance.

[0083] The document recommendation system provided by the present invention performs relationship mining and recommendation from multiple dimensions, comprehensively displays the comprehensive information of the document, improves the accuracy and availability of the recommendation, and expands the imagination space of the information. Moreover, by using the knowledge graph method, the system can search and infer quickly and efficiently, and the visualization in the form of a graph is more direct. The recommended content is rich and comprehensive, and the results are clear at a glance.

[0084] Part of the functions of the literature recommendation system described above have been put into use on the paper search page of the National Gene Bank Life Big Data Platform (CNGBdb, https: / / db.cngb.org / ). On the display page of each document, there is an area called "Literature Recommendation Graph". This graph uses a directed graph to display comprehensive information such as the field themes related to the paper, high-impact authors in the field, and highly relevant documents recommended by the present invention in multiple dimensions.

[0085] For example, taking the paper with PMID 12984 as an example, the document recommendation system provided by the present invention is used to recommend documents, and the result is as follows: Figure 2 As shown in the figure, the results displayed for this search mainly include the most influential authors of the paper, the top 5 medical subject terms closely related to the paper, and the top 5 other papers with similar content. For each author, the top five papers that are closely related to the author and have high influence are recommended. For each medical subject term, five similar documents with high relevance to the medical subject term are recommended. This achieves a multi-dimensional display of similar documents of the retrieved paper.

[0086] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0087] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A document recommendation system, characterized in that: include: A document information acquisition module, which is used to acquire basic information of the document; A semantic calculation module, wherein the semantic calculation module calculates the semantic similarity of the documents based on the abstracts of the documents and sorts the documents according to the semantic similarity; A subject evaluation module, which includes prediction of subject terms and the matching degree between a document and its annotated subject terms. The subject evaluation module classifies the document based on the subject terms of the document, obtains the classified document, and sorts the document according to the degree of similarity of the subject terms. The subject terms are determined based on the abstract and medical subject terms of the document. An influence analysis module, which determines the influence of the document based on the influence score of the document, the influence score of the author, the influence score of the document in its category after classification, and the influence score of the author of the document in its category after classification, and ranks the documents based on the influence; A knowledge graph construction module, wherein the knowledge graph construction module constructs a knowledge graph based on the literature ranking results of the semantic computing module, the subject evaluation module, and the influence analysis module, respectively. The influence score of the document is obtained based on the number of citations, publication year and impact factor of the document using a weight model. The formula for calculating the influence of a document is: Score = jif * cited_cnt * date_factor date_factor = 0.9^( max(0, cur_year - year) / 5) Among them: jif is the impact factor of the journal in which the document is published; cited_cnt is the number of citations of the document; cur_year is the current year, year is the year the document is published; date_factor is the date impact factor of the document publication time, and the date impact factor decays at a rate of 10% every 5 years. The influence score of the classified document in its category is based on the weight of the number of citations, publication year and impact factor of the document in its category, and is obtained using a weight model. The author's influence score is based on the number of papers published by the author, the number of citations of the papers and the author's contribution, calculated by calculating the h-index. The influence score of the author of the classified document in its category is calculated based on the number of papers published by the author, the number of citations of the paper and the author's contribution, and the h-index is calculated. The knowledge graph construction module is constructed based on documents, authors and keywords as three types of nodes.

2. The document recommendation system according to claim 1, characterized in that: The semantic computing module uses the BERT language representation model to generate a representation vector of the document.

3. The document recommendation system according to claim 2, characterized in that: The semantic computing module sorts the documents by the following method: (1) Pre-training the BERT language representation model using a document abstract word library larger than 4000M to obtain a BERT pre-trained language model; (2) Based on the BERT pre-trained language model and the abstract of the document, construct a representation vector containing semantic information; (3) Based on the representation vector, the semantic similarity of the documents is obtained by calculating the cosine similarity, and the documents are sorted based on the semantic similarity.

4. The document recommendation system according to claim 1, characterized in that: The topic evaluation module determines the subject words of the document based on a bidirectional LSTM recurrent neural network structure model and a regression model, or based on a BERT model and a sigmoid layer.

5. The document recommendation system according to claim 4, characterized in that: The subject evaluation module specifically includes: Training a subject term classification model based on the documents containing the complete subject terms as training data; Based on the trained subject word classification model, predicting the subject words of the document that lacks subject words, so as to obtain predicted subject words; Based on the abstract of the document with missing subject words and the predicted subject words, a relevance evaluation is performed to obtain the relevance weight between the subject words and the document.

6. The document recommendation system according to claim 5, characterized in that: The method for training the keyword classification model includes not freezing all parameters of the BERT network layer, not freezing the network parameters of the last two layers of BERT, and freezing at least one of the BERT network parameters.

7. The document recommendation system according to claim 5, characterized in that: The method for training the topic word classification model is to not freeze the network parameters of the last two layers of BERT.

8. The document recommendation system according to claim 1, characterized in that: The document recommendation system further comprises a comprehensive display module, which is used to output document recommendation results, and the document recommendation results are presented in the form of a visual map.

9. The document recommendation system according to claim 8, characterized in that: The visualization map includes at least one of the following: author, keyword and document.

10. A document recommendation method, characterized in that: include: The document recommendation system according to any one of claims 1 to 9 is used to perform document recommendation.

Images