Author name disambiguation method based on meta-path random walk network embedding and semantic representation
By constructing a heterogeneous network for papers and combining DBSCAN and xgboost classification methods, the problem of accurate assignment of papers with the same author was solved, and more accurate author name disambiguation results were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN TALENT DIGITAL TECH CO LTD
- Filing Date
- 2026-01-21
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies are insufficient to effectively assign papers by authors with the same name to their respective author profiles, resulting in inaccurate disambiguation results for authors with the same name.
We employ a method based on meta-path random walk network embedding and semantic representation. By constructing a heterogeneous network of papers, we generate path sets and train word vectors. Combined with DBSCAN clustering and XGBoost classification methods, we achieve accurate paper allocation.
It improves the accuracy of disambiguation for authors with the same name, ensures that all papers are correctly assigned to their respective author profiles, and enhances the completeness of the disambiguation results.
Smart Images
Figure CN122019772A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of information technology, specifically relating to an author name disambiguation method based on meta-path random walk network embedding and semantic representation. Background Technology
[0002] In many fields, disambiguation of names has long been considered a valuable and challenging problem, such as literature management and social network analysis. The goal of disambiguating authors with the same name in a paper is to utilize various information from the paper, such as title, authors, author affiliations, abstract, and keywords, to correctly assign the paper to the appropriate author profile using certain methods.
[0003] Currently, many researchers have proposed solutions to the problem of disambiguation of authors with the same name. These solutions mainly include rule-based matching methods, which utilize paper information or use representation learning methods to learn features from paper information. Then, clustering methods (such as hierarchical clustering, DBSCAN, etc.) are used to classify these feature vectors, so that similar papers are grouped together and dissimilar papers are classified into different categories.
[0004] However, the above-mentioned schemes usually merge pre-clustered paper sets, which makes it inconvenient to assign papers in discrete paper sets to their respective authors, thus making it difficult to obtain all papers by authors with the same name. Therefore, we need to propose an author name disambiguation method based on meta-path random walk network embedding and semantic representation to solve the above-mentioned problems. Summary of the Invention
[0005] The purpose of this invention is to provide an author name disambiguation method based on meta-path random walk network embedding and semantic representation. It uses heterogeneous network embedding technology based on meta-path random walk and semantic feature learning method based on word2vec to learn the feature vector of the paper, and combines DBSCAN clustering and xgboost classification methods to assign the paper to different authors, so as to solve the problems mentioned in the background technology.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] An author name disambiguation method based on meta-path random walk network embedding and semantic representation includes the following steps:
[0008] S1. Analyze the features of the paper and divide them into semantic features and discrete features;
[0009] S2. Construct a heterogeneous network of papers by utilizing discrete features to build relationships between papers;
[0010] S3. Use meta-path-based random walks to generate a path set composed of paper IDs;
[0011] S4. Use the skip-gram model to train the path set, treat the path set as a corpus, treat the paper IDs as words, and finally obtain the node representation vector corresponding to each paper ID.
[0012] S5. Extract K path sets, use the k path sets to train k word2vec models, and generate k sets of paper vectors;
[0013] S6. Calculate the cosine similarity matrix for each group of paper vectors, and then calculate the mean of these k similarity matrices to obtain the final paper relationship similarity matrix.
[0014] S7. Use semantic features to synthesize the same text and preprocess the same text to obtain the processed words;
[0015] S8. Generate word vectors from the processed words using the word2vec model and calculate the average value to obtain the semantic representation vector of the paper.
[0016] S9. For each name to be disambiguated, obtain the semantic representation vector of all its papers; when all words of a paper do not exist in the word2vec model, set its semantic representation vector to all 0 and save it to the outlier paper set for further processing; finally, calculate the cosine similarity between each pair of papers to obtain the semantic similarity matrix of the papers.
[0017] S10. After obtaining the paper relationship similarity matrix and the paper semantic similarity matrix, first add the paper relationship similarity matrix and the paper semantic similarity matrix together and calculate the average value to obtain the final paper similarity matrix.
[0018] S11. Based on the paper similarity matrix, input the matrix into DBSCAN to obtain the pre-clustered paper set;
[0019] S12. For the papers in the pre-clustered paper set, the xgboost algorithm is used for classification training according to the existing author names; then, the xgboost classification method is used to reassign the papers in the outlier paper set to the already clustered authors or new authors to obtain the final disambiguation result.
[0020] Preferably, in step S1, the semantic features refer to text features with semantic information, and the stylistic features are converted into text semantic vectors through a semantic representation learning model; discrete features refer to features whose textual information value is not significant, and only when an author appears in two articles does it indicate that the two papers have a common author, and therefore their similarity is high; the class features with high similarity are called discrete features, and are only used to convert them into relationships between papers.
[0021] Preferably, the features of the paper include title, abstract, author, field, organization, and keywords; the title, location, year, and keywords are defined as semantic features, and the author and publishing institution are defined as discrete features.
[0022] Preferably, in step S3, the random walk of the meta-path is performed in the heterogeneous network of the paper, where each node is selected in turn, and a random walk is performed on the edges between the nodes, with the node as the initial node. The walk path is saved as the training corpus of word2vec. The random walk based on the meta-path means that when walking randomly on the edges, the walk is not completely random, but is guided by the meta-path.
[0023] Preferably, during the random walk, the meta-path is walked in the order of p1→CoAuthor→p2→CoOrg→p3. In each walk, the next node is selected according to a certain type of edge specified by the current meta-path. A node connected to the current node through an edge of that type is randomly selected as the next node. In each long path, such meta-paths are sampled several times, that is, the last node of the previous meta-path is used as the first node of the next meta-path. Finally, the iteration continues until a certain number of iterations are reached, and then another node is selected as the starting node for the walk. Finally, several such long paths are generated, each path point is the ID of the paper, and each long path is stored row by row to generate a training corpus.
[0024] Preferably, during the process of a node randomly selecting the next node along an edge of a certain type under the guidance of the meta-path, the weight of the edge is taken into consideration; the larger the weight, the closer the relationship between the two nodes, and therefore the greater the probability that the node will jump along this relationship; each round selects each node in the graph as the starting point to start the random walk, and the number of rounds is defined as numwalks, thus generating a path set consisting of numwalks * N paths, where N is the number of nodes.
[0025] Preferably, in step S7, during the preprocessing of the same text segment, the letters are first lowercase, various non-letter symbols are removed, then extra spaces are removed, words are segmented by spaces, and stop words and words with a length of less than three are removed.
[0026] Preferably, when synthesizing the same text, the semantic features combine the paper's title, venue, organization, year, and keywords with spaces as separators to synthesize the same text.
[0027] Preferably, in step S2, when constructing the heterogeneous paper network, authors and institutions are used. For each name that needs to be disambiguated, the relationships between all papers corresponding to that name are extracted to construct a heterogeneous paper network. This heterogeneous network contains one type of node and two types of edges, where one node represents a paper and the two types of edges are CoAuthor and CoOrg, respectively.
[0028] Preferably, in step 12, the specific operation of reassigning the papers in the outlier paper set to the already clustered authors or new authors is as follows: For the papers in the pre-clustered paper set, the xgboost algorithm is used to classify and train according to the existing author names. Then, the trained model is used to classify the outlier paper set and match the outlier paper set to the existing authors or new authors. If the confidence of the classification is lower than the set threshold, the outlier paper set is considered to belong to the new authors.
[0029] The author name disambiguation method proposed in this invention, based on meta-path random walk network embedding and semantic representation, has the following advantages compared with existing technologies:
[0030] 1. This invention analyzes the features of papers and divides them into semantic features and discrete features. It uses discrete features to construct relationships between papers, building a heterogeneous paper network. A path set composed of paper IDs is generated using a meta-path-based random walk. The path set is trained using a skip-gram model, treating the path set as a corpus and the paper IDs as words. Finally, the node representation vector corresponding to each paper ID is obtained, and then the paper relationship similarity matrix is calculated. Semantic features are used to obtain the semantic representation vector of the papers, and then the paper semantic similarity matrix is calculated. After obtaining the paper relationship similarity matrix and the paper semantic similarity matrix, the paper relationship similarity matrix and the paper semantic similarity matrix are first... The similarity matrices are summed and averaged to obtain the final paper similarity matrix. Based on this matrix, the matrix is input into DBSCAN to obtain a pre-clustered paper set. For the papers in the pre-clustered paper set, the XGBoost algorithm is used for classification training according to the existing author names. Then, the XGBoost classification method is used to reassign the papers in the outlier set to the already clustered authors or new authors to obtain the final disambiguation result. After the above operations, the papers in the outlier set are also reassigned to their respective authors. The discrete paper set and the pre-clustered paper set are then integrated, thus obtaining the final result for all papers. This final result is the disambiguation result for the names to be disambiguated. Attached Figure Description
[0031] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0033] This invention provides, for example Figure 1 The author name disambiguation method shown includes the following steps: (The method is based on meta-path random walk network embedding and semantic representation.)
[0034] S1. Analyze the features of the paper and divide them into semantic features and discrete features;
[0035] The features of a paper include title, abstract, authors, field, organization, and keywords; title, location, year, and keywords are defined as semantic features, while authors and publishing institutions are defined as discrete features.
[0036] Semantic features refer to text features that possess semantic information. Stylistic features are converted into text semantic vectors through a semantic representation learning model. Discrete features are those that have little textual information value on their own. They are only considered to have a common author when an author appears in two articles, indicating that the two papers have a common author and therefore have a high similarity. Class features with high similarity are called discrete features and are only used to convert them into relationships between papers.
[0037] S2. Construct a heterogeneous network of papers by utilizing discrete features to build relationships between papers;
[0038] When constructing the heterogeneous network of papers, authors and institutions are used. For each name that needs to be disambiguated, the relationships between all papers corresponding to that name are extracted to construct a heterogeneous network of papers. This heterogeneous network contains one type of node and two types of edges, where one node represents a paper and the two types of edges are CoAuthor and CoOrg.
[0039] CoOrg represents the similarity relationship between the institutions in the names to be disambiguated in two papers, and the degree of the edge represents the number of co-occurring words between the two institutions. For example, if the institutions in two papers have the same word, and that word is not a stop word, then a CoOrg edge is constructed between them, and the weights are assigned according to the number of co-occurring words. Here, only the institution information of the names to be disambiguated is used.
[0040] S3. Use meta-path-based random walks to generate a path set composed of paper IDs;
[0041] The meta-path-based random walk in the heterogeneous network of the paper selects each node in turn, takes that node as the initial node, and performs a random walk on the edges between the nodes, saving the walk path as the training corpus of word2vec; the meta-path-based random walk means that when walking randomly on the edges, the walk is not completely random, but is guided by the meta-path.
[0042] S4. Use the skip-gram model to train the path set, treat the path set as a corpus, treat the paper IDs as words, and finally obtain the node representation vector corresponding to each paper ID.
[0043] During the random walk, the metapath is traversed in the order p1→CoAuthor→p2→CoOrg→p3. In each walk, the next node is selected based on a certain type of edge specified by the current metapath. A node connected to the current node through an edge of that type is randomly selected as the next node. In each long path, such metapaths are sampled several times, that is, the last node of the previous metapath is used as the first node of the next metapath. The iteration continues until a certain number of iterations are reached, and then another node is selected as the starting node for the walk. Finally, several such long paths are generated, each path point is the ID of the paper, and each long path is stored row by row to generate a training corpus.
[0044] During the process of a node randomly choosing the next node to hop along an edge of a certain type under the guidance of the meta-path, the edge weight is taken into account. The larger the weight, the closer the relationship between the two nodes, and therefore the greater the probability that the node will hop along this relationship; this probability is defined as being proportional to the weight. For example, if p1 is the current node and the next hop relationship is CoAuthor, then the two nodes with this relationship with p1 are p2 and p3. According to the weight of the edge between them, the probability of hopping from p1 to p2 is 1 / 3, and the probability of hopping to p3 is 2 / 3. In some cases, some relationships are missing for some papers. For example, if all the authors of a certain paper do not appear in the author list of any other paper, then the CoAuthor relationship is missing for it. When this happens, a more flexible strategy is adopted, that is, hopping according to the next relationship of the currently missing relationship in the meta-path. For the paper mentioned above, hopping is instead based on its CoOrg relationship;
[0045] In each round, a random walk is started from each node in the graph. The number of rounds is defined as numwalks, which generates a path set consisting of numwalks * N paths, where N is the number of nodes.
[0046] S5. Extract K path sets, use the k path sets to train k word2vec models, and generate k sets of paper vectors;
[0047] S6. Calculate the cosine similarity matrix for each group of paper vectors, and then calculate the mean of these k similarity matrices to obtain the final paper relationship similarity matrix.
[0048] S7. Use semantic features to synthesize the same text and preprocess the same text to obtain the processed words; when synthesizing the same text, the semantic features separate the title, venue, organization, year, and keywords of the paper with spaces to synthesize the same text; when preprocessing the same text, first lowercase the letters, remove various non-letter symbols, then remove extra spaces, segment words with spaces, and remove stop words and words with a length of less than three.
[0049] S8. Generate word vectors from the processed words using the word2vec model and calculate the average value to obtain the semantic representation vector of the paper.
[0050] S9. For each name to be disambiguated, obtain the semantic representation vector of all its papers; when all words of a paper do not exist in the word2vec model, set its semantic representation vector to all 0 and save it to the outlier paper set for further processing; finally, calculate the cosine similarity between each pair of papers to obtain the semantic similarity matrix of the papers.
[0051] S10. After obtaining the paper relationship similarity matrix and the paper semantic similarity matrix, first add the paper relationship similarity matrix and the paper semantic similarity matrix together and calculate the average value to obtain the final paper similarity matrix.
[0052] S11. Based on the paper similarity matrix, input the matrix into DBSCAN to obtain the pre-clustered paper set;
[0053] S12. For the papers in the pre-clustered paper set, the xgboost algorithm is used for classification training according to the existing author names; then, the xgboost classification method is used to reassign the papers in the outlier paper set to the already clustered authors or new authors to obtain the final disambiguation result.
[0054] The specific operation of reassigning papers in the outlier collection to authors that have already been clustered or to new authors is as follows: For papers in the pre-clustered collection, the XGBoost algorithm is used to classify and train the model according to the existing author names. Then, the trained model is used to classify the outlier collection and match it to existing or new authors. If the confidence of the classification is lower than the set threshold, the outlier collection is considered to belong to new authors.
[0055] In summary, the features of the papers are analyzed and divided into semantic features and discrete features. Relationships between papers are constructed using discrete features, resulting in a heterogeneous paper network. A path set composed of paper IDs is generated using a meta-path-based random walk, and the path set is trained using a skip-gram model. The path set is treated as a corpus, and paper IDs as words. Finally, the node representation vector corresponding to each paper ID is obtained, and the paper relationship similarity matrix is calculated. Semantic features are used to obtain the semantic representation vector of the papers, and the paper semantic similarity matrix is then calculated. After obtaining the paper relationship similarity matrix and the paper semantic similarity matrix, the paper relationship similarity matrix and the paper semantic similarity matrix are first compared... The similarity matrices are summed and averaged to obtain the final paper similarity matrix. Based on this matrix, the matrix is input into DBSCAN to obtain a pre-clustered paper set. For the papers in the pre-clustered paper set, the XGBoost algorithm is used for classification training according to the existing author names. Then, the XGBoost classification method is used to reassign the papers in the outlier set to the already clustered authors or new authors to obtain the final disambiguation result. After the above operations, the papers in the outlier set are also reassigned to their respective authors. The discrete paper set and the pre-clustered paper set are then integrated, thus obtaining the final result for all papers. This final result is the disambiguation result for the names to be disambiguated.
[0056] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for author name disambiguation based on meta-path random walk network embedding and semantic representation, characterized in that: Includes the following steps: S1. Analyze the features of the paper and divide them into semantic features and discrete features; S2. Construct a heterogeneous network of papers by utilizing discrete features to build relationships between papers; S3. Use meta-path-based random walks to generate a path set composed of paper IDs; S4. Use the skip-gram model to train the path set, treat the path set as a corpus, treat the paper IDs as words, and finally obtain the node representation vector corresponding to each paper ID. S5. Extract K path sets, use the k path sets to train k word2vec models, and generate k sets of paper vectors; S6. Calculate the cosine similarity matrix for each group of paper vectors, and then calculate the mean of these k similarity matrices to obtain the final paper relationship similarity matrix. S7. Use semantic features to synthesize the same text and preprocess the same text to obtain the processed words; S8. Generate word vectors from the processed words using the word2vec model and calculate the average value to obtain the semantic representation vector of the paper. S9. For each name to be disambiguated, obtain the semantic representation vector of all its papers; when all words of a paper do not exist in the word2vec model, set its semantic representation vector to all 0 and save it to the outlier paper set for further processing; finally, calculate the cosine similarity between each pair of papers to obtain the semantic similarity matrix of the papers. S10. After obtaining the paper relationship similarity matrix and the paper semantic similarity matrix, first add the paper relationship similarity matrix and the paper semantic similarity matrix together and calculate the average value to obtain the final paper similarity matrix. S11. Based on the paper similarity matrix, input the matrix into DBSCAN to obtain the pre-clustered paper set; S12. For the papers in the pre-clustered paper set, the xgboost algorithm is used for classification training according to the existing author names; then, the xgboost classification method is used to reassign the papers in the outlier paper set to the already clustered authors or new authors to obtain the final disambiguation result.
2. The author name disambiguation method based on meta-path random walk network embedding and semantic representation as described in claim 1, characterized in that: In step S1, the semantic features refer to text features with semantic information. The stylistic features are converted into text semantic vectors through a semantic representation learning model. Discrete features refer to features that have little textual information value on their own. Only when an author appears in two articles does it indicate that the two papers have a common author, and therefore they are highly similar. The class features with high similarity are called discrete features and are only used to convert them into relationships between papers.
3. The author name disambiguation method based on meta-path random walk network embedding and semantic representation as described in claim 2, characterized in that: The paper's features include title, abstract, author, field, organization, and keywords; title, location, year, and keywords are defined as semantic features, while author and publishing institution are defined as discrete features.
4. The author name disambiguation method based on meta-path random walk network embedding and semantic representation as described in claim 3, characterized in that: In step S3, the random walk of the meta-path is to select each node in turn in the heterogeneous network of the paper, take the node as the initial node, and perform a random walk on the edges between the nodes. The walk path is saved as the training corpus of word2vec. The random walk based on the meta-path means that when walking randomly on the edges, the walk is not completely random, but is guided by the meta-path.
5. The author name disambiguation method based on meta-path random walk network embedding and semantic representation according to claim 4, characterized in that: During the random walk, the meta-path is traversed in the order p1→CoAuthor→p2→CoOrg→p3. In each walk, the next node is selected based on a certain type of edge specified by the current meta-path. A node connected to the current node through an edge of that type is randomly selected as the next node. In each long path, such meta-paths are sampled several times, that is, the last node of the previous meta-path is used as the first node of the next meta-path. The iteration continues until a certain number of iterations are reached, and then another node is selected as the starting node for the walk. Finally, several such long paths are generated, each path point is the ID of the paper, and each long path is stored row by row to generate a training corpus.
6. The author name disambiguation method based on meta-path random walk network embedding and semantic representation according to claim 5, characterized in that: As a node walks randomly towards an edge of a certain type under the guidance of a meta-path, the weight of the edge is taken into account. The larger the weight, the closer the relationship between the two nodes, and therefore the greater the probability that the node will jump along this relationship. In each round, each node in the graph is selected as the starting point to start the random walk. The number of rounds is defined as numwalks, which generates a path set consisting of numwalks * N paths, where N is the number of nodes.
7. The author name disambiguation method based on meta-path random walk network embedding and semantic representation as described in claim 6, characterized in that: In step S7, during the preprocessing of the same text segment, the letters are first lowercase, various non-letter symbols are removed, then extra spaces are removed, words are segmented by spaces, and stop words and words with a length of less than three are removed.
8. The author name disambiguation method based on meta-path random walk network embedding and semantic representation according to claim 7, characterized in that: When synthesizing the same text, the semantic features combine the paper's title, venue, organization, year, and keywords with spaces as separators.
9. The author name disambiguation method based on meta-path random walk network embedding and semantic representation according to claim 8, characterized in that: In step S2, when constructing the heterogeneous paper network, authors and institutions are used. For each name that needs to be disambiguated, the relationships between all papers corresponding to the name that needs to be disambiguated are extracted to construct a heterogeneous paper network. This heterogeneous network contains one type of node and two types of edges. One node represents a paper, and the two types of edges are CoAuthor and CoOrg, respectively.
10. The author name disambiguation method based on meta-path random walk network embedding and semantic representation according to claim 9, characterized in that: In step 12, the specific operation of reassigning the papers in the outlier paper set to the already clustered authors or new authors is as follows: For the papers in the pre-clustered paper set, the XGBoost algorithm is used to classify and train according to the existing author names. Then, the trained model is used to classify the outlier paper set and match the outlier paper set to the existing authors or new authors. If the confidence of the classification is lower than the set threshold, the outlier paper set is considered to belong to the new authors.