A method and device for disambiguating names by integrating multi-source information

Abstract

The present invention discloses a method and device for disambiguating names by fusing multi-source information, the method comprising: dividing all texts into several classes; clustering each class of texts based on the institution names corresponding to the authors with the same name, the co-authors of the texts and the subject contents of the texts, so as to obtain the first clustering result of the institution names, the first clustering result of the co-authors and the first clustering result of the subject contents of the texts in the class; fusing the first clustering result of the institution names, the first clustering result of the co-authors and the first clustering result of the subject contents based on the co-occurrence information of the institution information in the cluster and the texts, so as to obtain the preliminary clustering result of the texts in the class; extracting a single cluster text from the preliminary clustering result, and fusing the single cluster text based on the similarity between the single cluster text and other texts in the class of texts, so as to obtain the name disambiguation result. The present invention can achieve a better disambiguation accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of computer data mining and analysis, and in particular to the field of information extraction, and specifically refers to a method and device for disambiguating names by integrating multi-source information. Background Art

[0002] At present, with the rapid increase in the number of scientific research documents and the needs of scientific research document management, academic author search, and academic network analysis, the problem of duplicate academic author names is becoming increasingly serious. Therefore, it is of great application value to quickly and accurately disambiguate authors with the same name.

[0003] The task of academic author name disambiguation is to return a set of paper clusters given a large number of papers with the same author name, so that the authors with the same author name in a cluster correspond to the same person in reality, and the authors with the same author name in different clusters correspond to different real people. In recent years, the methods for academic author name disambiguation are mainly divided into supervised methods and unsupervised methods:

[0004] The classification method based on supervised learning first extracts explicit or implicit correlation features between papers, then learns the classification model parameters through the extracted feature vectors and the corresponding label results, and finally disambiguates authors with the same name based on the trained classification model. Supervised methods require a sufficient amount of labeled data to be trained to obtain a better model. In addition, the learned model is sensitive to data distribution, which will affect the accuracy of disambiguation when noisy or sparse data appears.

[0005] In order to solve the problems of supervised methods, some researchers have proposed a clustering method based on unsupervised learning. This method does not require learning model parameters. It only needs to select a suitable clustering model after calculating the similar features of the attributes and relationship information between papers to obtain the final disambiguation result. However, although unsupervised methods do not require the provision of labeled data, they often need to specify the number of clusters in advance. For actual systems, the final number of clusters is often unknown. Therefore, many current name disambiguation methods based on clustering models have low accuracy.

[0006] Since the unsupervised academic author name disambiguation method has no annotation information and needs to specify the total number of clusters in advance, the final disambiguation accuracy is low.

[0007] The main problem with current unsupervised academic author name disambiguation methods is that they need to specify the number of clusters. If the specified value is inappropriate, the disambiguation accuracy will be low. In addition, unsupervised methods do not have labeled information. Therefore, how to more effectively utilize known information and solve the problem of needing to specify the number of clusters in advance is the main challenge that needs to be faced. Summary of the invention

[0008] In order to overcome the shortcomings of existing academic author name disambiguation methods, such as the need for a large amount of labeled data, the need to specify the number of clusters in advance, and low accuracy, and to effectively utilize known information and avoid the need to specify the number of clusters in advance, the present invention proposes a name disambiguation method that integrates multi-source information. The method utilizes multi-source information such as the author, institution, title, text, keywords, abstract, journal, etc. of the paper to perform phased aggregation disambiguation, and realizes unsupervised clustering without specifying the number of clusters.

[0009] The technical solution of the present invention includes:

[0010] A method for disambiguating names by fusing multi-source information, the method comprising:

[0011] Divide all texts into several categories; each category of text corresponds to an author with the same name, and each text is divided into an independent cluster;

[0012] Based on the institution names, co-authors and subject contents of the texts corresponding to the authors with the same name, each type of text is clustered to obtain the first clustering result of the institution name, the first clustering result of the co-authors and the first clustering result of the subject contents of the texts;

[0013] Based on the organization information within the cluster, clusters in the first clustering result of the organization name, the first clustering result of the co-authors, and the first clustering result of the subject content are merged respectively to obtain the second clustering result of the organization name, the second clustering result of the co-authors, and the second clustering result of the subject content of the text of this type;

[0014] According to the co-occurrence information of the text, the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content are integrated to obtain a preliminary clustering result of the text of this type;

[0015] A single cluster of texts is extracted from the preliminary clustering results, and the single cluster of texts is fused based on the similarity between the single cluster of texts and other texts in the same category to obtain a name disambiguation result.

[0016] Furthermore, clustering each type of text based on the institution names corresponding to the authors with the same name to obtain a first clustering result of the institution names of the texts in this type includes:

[0017] Get the institution name of any two texts;

[0018] The Simhash algorithm is used to map the institution name into an institution name feature vector, and the length of the institution name feature vector is calculated;

[0019] Calculate the similarity of the two texts based on the Hamming distance between the two institution name feature vectors and the length of the institution name feature vector;

[0020] When the similarity between the two texts is greater than a first similarity threshold, merging the clusters containing the two texts into one cluster;

[0021] When the similarity between the two texts is less than the first similarity threshold, the independence of the clusters to which the two texts belong is maintained.

[0022] Furthermore, clustering each type of text based on the co-authors of the text corresponding to the authors with the same name to obtain a first clustering result of the co-authors of the text in this type includes:

[0023] Obtain the names of the institutions and co-author lists of any two texts with the same author name; wherein the co-author list does not include the author with the same author name;

[0024] Count the number of co-authors and other authors with the same name in the co-author list;

[0025] When the number of co-authors is greater than a first number threshold, merging the clusters containing the two texts into one cluster;

[0026] When the number of co-authors is between a first number threshold and a second number threshold, and the similarity of the names of the institutions of the same authors is greater than the second similarity threshold, the clusters of the two texts are merged into one cluster; wherein the first number threshold is greater than the second number threshold;

[0027] When the number of the common authors is between the first number threshold and the second number threshold, and the number of other authors with the same name is greater than the third number threshold, merging the clusters of the two texts into one cluster;

[0028] When the number of the common authors is between the first number threshold and the second number threshold, and the similarity of the names of the institutions of the authors with the same name is less than the second similarity threshold, the independence of the clusters to which the two texts belong is maintained;

[0029] When the number of the common authors is between the first number threshold and the second number threshold, and the number of other authors with the same name is less than the third number threshold, the independence of the clusters to which the two texts belong is maintained;

[0030] When the number of co-authors is less than the second number threshold, the independence of the clusters to which the two texts belong is maintained.

[0031] Furthermore, each type of text is clustered based on the subject content of the text corresponding to the author with the same name, so as to obtain a first clustering result of the subject content of the text of this type, including:

[0032] Divide each text into a number of first subject domains, and set a weight coefficient for each first subject domain; the first subject domains include: a title domain, a keyword domain, a publication journal domain, and an abstract domain;

[0033] Get the text subject content of any two texts;

[0034] Based on the text subject content of the two texts, calculate the content similarity of the two texts Among them, p a Indicates the text subject content of a text, p b represents the text subject content of another text, and t represents the text subject content p a A word in the text, tf(t) represents the word t in the text topic content p b The number of occurrences in , weight(t) represents the weight coefficient of the first subject domain where word t is located, and Z is the standard factor;

[0035] Based on the content similarity core, the text topic content p is generated respectively. a and text subject content p b Sort the text content similarity from large to small;

[0036] If the content similarity core is greater than the third similarity threshold, and the text subject content p a and text subject content p b If the two texts are in the top n in descending order of similarity of the text subject content, the clusters where the two texts are located are merged into one cluster;

[0037] If the content similarity core is less than the third similarity threshold, or the text subject content p a and text subject content p b At least one of them is not in the top n in the descending order of similarity of the text subject content of each other, maintaining the independence of the clusters to which the two texts belong.

[0038] Furthermore, the cluster merging in the first clustering result of the institution name based on the information of the institution in the cluster to obtain the second clustering result of the institution name of the text of this type includes:

[0039] For each cluster in the first clustering result of the institution names, screening out institution names whose occurrence times are greater than a fourth quantity threshold as representative institutions of the cluster;

[0040] Calculate the similarity of representative organizations between clusters;

[0041] When the similarity of the representative organizations between clusters is greater than a fourth similarity threshold, the corresponding clusters are merged;

[0042] When the similarity of the representative organizations between clusters is less than a fourth similarity threshold, the independence of the corresponding clusters is maintained.

[0043] Furthermore, the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content are fused according to the co-occurrence information of the text to obtain a preliminary clustering result of the text, including:

[0044] Calculate the text co-occurrence information intersection V1 between the clusters in the second clustering result of the institution name and the second clustering result of the co-authors;

[0045] When the number of texts in the text co-occurrence information intersection V1 is greater than a fifth number threshold, merging the corresponding clusters to obtain a clustering result D of the texts of this type;

[0046] Calculate the text co-occurrence information intersection V2 between the clusters in the clustering result D of the text of this type and the second clustering result of the subject content;

[0047] When the number of texts in the text co-occurrence information intersection V1 is greater than the sixth quantity threshold, corresponding clusters are merged to obtain a preliminary clustering result for the texts of this type.

[0048] Furthermore, after the single cluster text is fused based on the similarity between the single cluster text and other texts in the same type of text, a name disambiguation result is obtained, including:

[0049] Divide each text into the institution domain, author domain, title domain, keyword domain, publication journal domain and abstract domain;

[0050] Based on the institution domain, author domain, title domain, keyword domain, publication journal domain and abstract domain, and combined with the weight coefficients set for different domains, the similarity between a single cluster text and other texts in this type of text is calculated;

[0051] Select other texts in the same category with the highest similarity, and when the similarity between the single cluster text and other texts in the same category is greater than a fifth similarity threshold, add the single cluster text to the corresponding cluster of other texts in the same category.

[0052] A device for disambiguating names by fusing multi-source information, characterized in that the device comprises:

[0053] The text classification module is used to divide all texts into several categories; each category of text corresponds to an author with the same name, and each text is divided into an independent cluster;

[0054] An initial clustering module is used to cluster each type of text based on the institution name corresponding to the author with the same name, the co-author of the text and the subject content of the text, so as to obtain the first clustering result of the institution name, the first clustering result of the co-author and the first clustering result of the subject content of the text;

[0055] The clustering fusion module is used to merge the clusters in the first clustering result of the institution name, the first clustering result of the co-authors and the first clustering result of the subject content based on the institution information within the cluster, so as to obtain the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content of this type of text; fuse the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content according to the co-occurrence information of the text to obtain the preliminary clustering result of this type of text; extract the single cluster text in the preliminary clustering result, and fuse the single cluster text based on the similarity between the single cluster text and other texts in this type of text to obtain the person name disambiguation result.

[0056] A computer device, characterized in that the computer device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for disambiguating a person's name by fusing multi-source information as described above is implemented.

[0057] A computer-readable storage medium, characterized in that computer program instructions are stored on the computer-readable storage medium, and when the computer program instructions are executed by a processor, any of the above-mentioned methods for disambiguating personal names by fusing multi-source information is implemented.

[0058] Compared with the prior art, the present invention has the following positive effects:

[0059] The present invention realizes the effective utilization of multi-source known information of papers through the proposed method of disambiguating authors with the same name by fusing multi-source information, and solves the problem that the unsupervised clustering method needs to specify the number of clusters in advance, thereby achieving better disambiguation accuracy. The verification experiment of the present invention adopts a public English academic name disambiguation data set, which contains 41,027 English papers, and the information of each paper includes: paper ID, title, author list, institution corresponding to each author, publishing journal (conference), keywords, abstract. The model needs to divide all papers into multiple clusters according to the above information, so that the papers within a cluster belong to the same person, and the papers between different clusters do not belong to the same person, so as to achieve the purpose of academic name disambiguation. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic diagram of the overall structure of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in detail below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention but not to limit the scope of the present invention.

[0062] The method for disambiguating names by fusing multi-source information of the present invention mainly includes an initial clustering stage and a clustering fusion stage. Figure 1 As shown, the initial clustering stage includes: an institution name-based clustering module, a co-author-based clustering module, and a subject content-based clustering module; the clustering fusion stage includes: an intra-cluster institution information fusion module, a paper co-occurrence information fusion module, and a single cluster paper fusion module.

[0063] 1. Initial clustering stage.

[0064] The present invention first divides all papers into multiple categories according to the author's name, ensures that all papers in each category have an author with the same name, and divides each paper into an independent cluster.

[0065] a. Based on the institution name clustering module, all paper clusters with similar institution names in the papers of the same author are merged. For authors with the same name in the same institution, they are likely to correspond to the same person in reality. Similarity calculation of institution names: First, generate the hash fingerprint corresponding to the institution name, and use the Simhash algorithm to map the institution name to a low-dimensional feature vector with a vector length of 128. The similarity p calculated based on the Hamming distance between institution a and institution b is defined as follows:

[0066]

[0067] Among them, HammingDis represents the Hamming distance between Simhash values. When the hash similarity p between institution names is greater than 0.8, two paper clusters with the same author name are merged into one cluster.

[0068] b. Based on the co-author clustering module, all paper clusters with multiple co-authors in the papers with the same author name are merged. When the author lists of two papers are highly overlapping, the authors with the same author name are likely to correspond to the same person in reality. First, calculate the intersection of the author lists in the two papers to obtain the co-author list, and calculate the name length of other authors with the same author name except the author to be disambiguated. Then merge the papers according to the following judgment conditions: when the number of co-authors is greater than 3, directly merge the two papers into one cluster; when the number of co-authors is greater than 1 and less than or equal to 3, if the similarity of the institution name corresponding to the authors with the same author name is greater than 0.8 or the name length of other authors with the same author name is greater than 2, then merge the two papers into one cluster.

[0069] c. Based on the subject content clustering module, all paper clusters with similar topics and text contents in the papers with the same author are merged. For authors with the same name in the same research field, they are likely to correspond to the same person in reality. Each paper is divided into four domains, namely the title domain, keyword domain, publication journal domain, and abstract domain, and different weight coefficients are set for different domains. According to the TF-IDF algorithm, the content similarity of any two papers with the same author is calculated, and the content similarity score calculation formula is as follows:

[0070]

[0071] where p a , p b represents two papers with the same author, and t represents paper p a A word in the paper p a The word t in the paper p b , idf(t) represents the inverse of the proportion of documents containing word t in all other papers with the same title, weight(t) represents the weight of the domain where word t is located, and Z is the normalization factor. For any paper, other papers with the same author are sorted from high to low according to the content similarity score. a , p b , if p b Located in p a The first 5 and p a Also located in p b When the two papers are in the top 5 and the similarity score of the two papers is greater than 0.8, the two papers are merged into one cluster, and so on, all papers with the same author that meet the conditions are merged.

[0072] 2. Clustering fusion stage.

[0073] In the initial clustering stage, the model used three different types of information to independently cluster all papers with the same author, and obtained three different clustering results for all papers with the same author under different clustering strategies. Since each strategy uses a single piece of information, the clustering results obtained in the initial stage are relatively scattered. In the cluster fusion stage, the above three clustering results are further merged, and the papers with the same author corresponding to the same person in reality are merged into one category as much as possible.

[0074] a. Based on the intra-cluster institution information fusion module, the results of the initial clustering stage are further merged independently, and three different clustering results with higher fusion degree are obtained. First, the institutions contained in each cluster and their frequency of occurrence in the clustering results obtained in the initial clustering stage are counted respectively. For the paper clustering clusters obtained in the initial clustering stage, the institution names with more than 3 occurrences are selected as the representative institutions of each cluster, and the similarity of the representative institutions between clusters is calculated respectively. If the similarity is greater than 0.8, the two clusters are merged into one, and so on. All clusters that meet the results are merged to obtain the further merging results of the three clustering methods in the initial clustering stage.

[0075] b. Based on the paper co-occurrence information fusion module. First, calculate the intersection of the clusters based on the institution name and the clusters based on the co-authors. If the number of papers in the intersection is greater than or equal to 3, merge the results; calculate the intersection of the result and the clusters based on the subject content. If the number of papers in the intersection is greater than or equal to 3, merge the results, and finally obtain a unified clustering result.

[0076] c. Single-cluster paper fusion module. Each paper is divided into six domains, namely, institution domain, author domain, title domain, keyword domain, publication journal domain, and abstract domain, and different weight coefficients are set for different domains. The same TF-IDF algorithm as the subject content clustering module is used to calculate the similarity between the current paper and other papers with the same author. If the highest similarity between the current paper and other papers with the same author is greater than 0.7, the current paper will be added to the cluster where the paper with the highest similarity is located, otherwise the current paper will remain as a single-cluster paper.

[0077] The present invention is described in detail below in conjunction with an embodiment.

[0078] (1) All papers are divided into 50 categories based on the author's name, where each paper in each category corresponds to an author with the same name, and each paper in the category is divided into an independent cluster.

[0079] (2) For the 1,198 papers with the same author "jie_tang", the papers with similar institution names were merged, resulting in 555 clusters corresponding to "jie_tang", including 452 unmerged independent clusters and 103 clusters with more than 1 paper. First, the hash fingerprint of the institution corresponding to the author "jie_tang" in each paper was generated, and the Simhash algorithm was used to map the institution name to a low-dimensional feature vector. The hash similarity between the institution names of papers with the same author was calculated based on the Hamming distance. For example, the similarity of the institution names corresponding to papers A and B was 0.84, which was greater than the set threshold of 0.8, so the above two papers were merged into the same cluster.

[0080] (3) For the 1,198 papers with the same author "jie_tang", the papers with multiple co-authors were merged, and 277 clusters corresponding to "jie_tang" were obtained, including 185 independent clusters that were not merged and 92 clusters with more than 1 paper. First, the intersection of the author lists in the two papers was calculated to obtain the co-author list; the pinyin recognition algorithm was used to calculate the name length of other authors with the same name except "Jie Tang", and the length of other co-authors was 3; according to the judgment condition, if the number of co-authors is 3 and the length of other co-author names is 3, then the two papers are merged into the same cluster.

[0081] (4) For the 1198 papers with the author "jie_tang" with the same name, the papers with similar topics and text contents were merged, and 368 clusters corresponding to "jie_tang" were obtained, including 200 independent clusters that were not merged and 168 clusters with more than 1 paper. First, each paper was divided into four domains, namely the title domain, keyword domain, publication journal domain, and abstract domain, and the weight coefficients of different domains were set to 2, 1.5, 1, and 1 respectively; the content similarity of any two papers was calculated according to the TF-IDF algorithm, and for any paper, other papers with the same author were sorted from high to low according to the content similarity score, and all papers with the same author with a similarity score greater than 0.8 were merged.

[0082] (5) Based on the information of the institutions within the cluster, the results obtained in steps 2, 3, and 4 of the initial clustering stage are further merged independently. First, the institutions contained in each cluster and their frequency of occurrence in the clustering results obtained in steps 2, 3, and 4 of the initial clustering stage are counted respectively; for the paper clusters obtained in steps 2, 3, and 4 of the initial clustering stage, the institution names with a frequency of occurrence greater than 3 are selected as the representative institution of each cluster; for the paper clusters obtained in step 2, the similarity of the representative institutions between clusters is calculated. If the similarity is greater than 0.8, the two clusters are merged into one. Finally, the corresponding clusters in step 2 were reduced from 555 to 517, and clustering result 5-2 was obtained; for the clustering results obtained in step 3, the same operation was taken to further merge the clusters that met the results, and the corresponding clusters were reduced from 277 to 273, and clustering result 5-3 was obtained; for the clustering results obtained in step 4, the same operation was taken to further merge the clusters that met the results, and the corresponding clusters were reduced from 368 to 361, and clustering result 5-4 was obtained.

[0083] (6) Based on the co-occurrence information of papers, the three different clustering results obtained in step 5 are fused to produce a final unified result. After fusion, there are 133 clusters corresponding to "jie_tang", including 82 unmerged independent clusters and 51 clusters with more than 1 paper. First, the intersection between the clusters of step 5-2 and 5-3 is calculated. If the number of papers in the intersection is greater than or equal to 3, the results are merged. In the fusion process, the result of 5-3 is used as the benchmark, and the clusters in 5-3 are finally reduced from 273 to 205, and the cluster result 6-1 is obtained; the intersection between the clusters of 5-4 and 6-1 is calculated. If the number of papers in the intersection is greater than or equal to 3, the results are merged. In the fusion process, the result of 6-1 is used as the benchmark, and the clusters in step 6-1 are finally reduced from 205 to 133.

[0084] (7) For the single-cluster papers that have not been merged with other papers in the results of step 6, calculate their similarity with other papers with the same author. If it is higher than 0.7, add it to other clusters. In the end, there are 67 clusters corresponding to "jie_tang", including

[0085] There are 10 unmerged independent clusters and 57 clusters with more than 1 paper. First, each paper is divided into six domains, namely, institution domain, author domain, title domain, keyword domain, publication journal domain, and abstract domain. The weight coefficients set for different domains are 2, 2, 1.5, 1, 1, 1 respectively; the TF-IDF algorithm is used to calculate the similarity between the current paper and other papers with the same author; the paper with the same author that is most similar to the current paper is selected. If the similarity between the two is greater than 0.7, the current paper is added to the cluster where the paper with the highest similarity is located, otherwise the current paper remains a single cluster paper.

[0086] In summary, the present invention realizes the effective utilization of multi-source known information of papers through the proposed method of disambiguating authors with the same name by fusing multi-source information, and solves the problem that the unsupervised clustering method needs to specify the number of clusters in advance, thereby achieving better disambiguation accuracy. The verification experiment of the present invention adopts a public English academic name disambiguation dataset, which contains 41,027 English papers. The information of each paper includes: paper ID, title, author list, institution corresponding to each author, publishing journal (conference), keywords, and abstract. The model needs to divide all papers into multiple clusters according to the above information, so that the papers within a cluster belong to the same person, and the papers between different clusters do not belong to the same person, so as to achieve the purpose of academic name disambiguation. Experimental results show that the present invention achieves better disambiguation accuracy compared with other baseline methods.

[0087] Table 1 Comparison results of disambiguation accuracy

[0088]

[0089] Although the specific contents, implementation algorithms and drawings of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the contents of the present invention and implement them accordingly, it can be understood by those skilled in the art that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. The present invention should not be limited to the contents disclosed in the best embodiment of this specification and the drawings, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A method for disambiguating names by integrating multi-source information, characterized in that: The method comprises: Divide all texts into several categories; each category of text corresponds to an author with the same name, and each text is divided into an independent cluster; Based on the institution names, co-authors and subject contents of the texts corresponding to the authors with the same name, each type of text is clustered to obtain the first clustering result of the institution name, the first clustering result of the co-authors and the first clustering result of the subject contents of the texts; Based on the intra-cluster institution information, clusters in the first clustering result of the institution name, the first clustering result of the co-authors, and the first clustering result of the subject content are merged respectively to obtain the second clustering result of the institution name, the second clustering result of the co-authors, and the second clustering result of the subject content of the text of this type; wherein, the cluster merging in the first clustering result of the institution name based on the intra-cluster institution information to obtain the second clustering result of the institution name of the text of this type includes: For each cluster in the first clustering result of the institution names, screening out institution names whose occurrence times are greater than a fourth quantity threshold as representative institutions of the cluster; Calculate the similarity of representative organizations between clusters; When the similarity of the representative organizations between clusters is greater than a fourth similarity threshold, the corresponding clusters are merged; When the similarity of the representative organizations between clusters is less than the fourth similarity threshold, the independence of the corresponding clusters is maintained; According to the co-occurrence information of the text, the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content are integrated to obtain the preliminary clustering result of the text of this type; wherein, according to the co-occurrence information of the text, the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content are integrated to obtain the preliminary clustering result of the text of this type, including: Calculate the text co-occurrence information intersection V1 between the clusters in the second clustering result of the institution name and the second clustering result of the co-authors; When the number of texts in the text co-occurrence information intersection V1 is greater than a fifth number threshold, merging the corresponding clusters to obtain a clustering result D of the texts of this type; Calculate the text co-occurrence information intersection V2 between the clusters in the clustering result D of the text of this type and the second clustering result of the subject content; When the number of texts in the text co-occurrence information intersection V2 is greater than a sixth number threshold, merging the corresponding clusters to obtain a preliminary clustering result of the texts of this type; A single cluster of texts is extracted from the preliminary clustering results, and the single cluster of texts is fused based on the similarity between the single cluster of texts and other texts in the same category to obtain a name disambiguation result.

2. The method according to claim 1, characterized in that The clustering of each type of text based on the institution names corresponding to the authors with the same name to obtain the first clustering result of the institution names of the texts in this type includes: Get the institution name of any two texts; The Simhash algorithm is used to map the institution name into an institution name feature vector, and the length of the institution name feature vector is calculated; Calculate the similarity of the two texts based on the Hamming distance between the two institution name feature vectors and the length of the institution name feature vector; When the similarity between the two texts is greater than a first similarity threshold, merging the clusters containing the two texts into one cluster; When the similarity between the two texts is less than the first similarity threshold, the independence of the clusters to which the two texts belong is maintained.

3. The method according to claim 1, characterized in that The clustering of each type of text based on the co-authors of the texts corresponding to the authors with the same name to obtain a first clustering result of the co-authors of the texts in this type includes: Obtain the names of the institutions and co-author lists of any two texts with the same author name; wherein the co-author list does not include the author with the same author name; Count the number of co-authors and other authors with the same name in the co-author list; When the number of co-authors is greater than a first number threshold, merging the clusters containing the two texts into one cluster; When the number of co-authors is between a first number threshold and a second number threshold, and the similarity of the names of the institutions of the same author is greater than the second similarity threshold, the clusters of the two texts are merged into one cluster; wherein the first number threshold is greater than the second number threshold; When the number of the common authors is between the first number threshold and the second number threshold, and the number of other authors with the same name is greater than the third number threshold, merging the clusters of the two texts into one cluster; When the number of the common authors is between the first number threshold and the second number threshold, and the similarity of the names of the institutions of the authors with the same name is less than the second similarity threshold, the independence of the clusters to which the two texts belong is maintained; When the number of the common authors is between the first number threshold and the second number threshold, and the number of other authors with the same name is less than the third number threshold, the independence of the clusters to which the two texts belong is maintained; When the number of co-authors is less than the second number threshold, the independence of the clusters to which the two texts belong is maintained.

4. The method according to claim 1, characterized in that Each type of text is clustered based on the text subject content corresponding to the author with the same name to obtain the first clustering result of the subject content of the text, including: Divide each text into a number of first subject domains, and set a weight coefficient for each first subject domain; the first subject domains include: a title domain, a keyword domain, a publication journal domain, and an abstract domain; Get the text subject content of any two texts; Based on the text subject content of the two texts, calculate the content similarity of the two texts Among them, p a Indicates the text subject content of a text, p b represents the text subject content of another text, and t represents the text subject content p a A word in the text, tf(t) represents the word t in the text topic content p b The number of occurrences in , weight(t) represents the weight coefficient of the first subject domain where word t is located, and Z is the standard factor; Based on the content similarity core, the text topic content p is generated respectively. a and text subject content p b Sort the text content similarity from large to small; If the content similarity core is greater than the third similarity threshold, and the text subject content p a and text subject content p b If the two texts are in the top n in descending order of similarity of the text subject content, the clusters where the two texts are located are merged into one cluster; If the content similarity core is less than the third similarity threshold, or the text subject content p a and text subject content p b At least one of them is not in the top n in the descending order of similarity of the text subject content of each other, maintaining the independence of the clusters to which the two texts belong.

5. The method according to claim 1, characterized in that After the single cluster text is fused based on the similarity between the single cluster text and other texts in the same type of text, a name disambiguation result is obtained, including: Divide each text into the institution domain, author domain, title domain, keyword domain, publication journal domain and abstract domain; Based on the institution domain, author domain, title domain, keyword domain, publication journal domain and abstract domain, and combined with the weight coefficients set for different domains, the similarity between a single cluster text and other texts in this category is calculated; Select other texts in the same category with the highest similarity, and when the similarity between the single cluster text and other texts in the same category is greater than a fifth similarity threshold, add the single cluster text to the corresponding cluster of other texts in the same category.

6. A device for disambiguating names by integrating multi-source information, characterized in that: The device comprises: The text classification module is used to divide all texts into several categories; each category of text corresponds to an author with the same name, and each text is divided into an independent cluster; The initial clustering module is used to cluster each type of text based on the institution names corresponding to the authors with the same name, the co-authors of the texts and the subject content of the texts, so as to obtain the first clustering result of the institution name, the first clustering result of the co-authors and the first clustering result of the subject content of the texts of this type; wherein, clusters in the first clustering result of the institution name are merged based on the information of the institutions within the clusters to obtain the second clustering result of the institution name of the texts of this type, including: For each cluster in the first clustering result of the institution names, screening out institution names whose occurrence times are greater than a fourth quantity threshold as representative institutions of the cluster; Calculate the similarity of representative organizations between clusters; When the similarity of the representative organizations between clusters is greater than a fourth similarity threshold, the corresponding clusters are merged; When the similarity of the representative organizations between clusters is less than the fourth similarity threshold, the independence of the corresponding clusters is maintained; The cluster fusion module is used to merge the clusters in the first clustering result of the institution name, the first clustering result of the co-authors and the first clustering result of the subject content respectively based on the information of the institution in the cluster, so as to obtain the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content of the text of this type; the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content are fused according to the co-occurrence information of the text to obtain the preliminary clustering result of the text of this type; the single cluster text in the preliminary clustering result is extracted, and the single cluster text is fused based on the similarity between the single cluster text and other texts in the text of this type to obtain the person name disambiguation result; wherein, the fusion of the second clustering result of the institution name, the second clustering result of the co-authors and the second clustering result of the subject content according to the co-occurrence information of the text to obtain the preliminary clustering result of the text of this type includes: Calculate the text co-occurrence information intersection V1 between the clusters in the second clustering result of the institution name and the second clustering result of the co-authors; When the number of texts in the text co-occurrence information intersection V1 is greater than a fifth number threshold, merging the corresponding clusters to obtain a clustering result D of the text of this type; Calculate the text co-occurrence information intersection V2 between the clusters in the clustering result D of the text of this type and the second clustering result of the subject content; When the number of texts in the text co-occurrence information intersection V2 is greater than the sixth quantity threshold, corresponding clusters are merged to obtain a preliminary clustering result for the texts of this type.

7. A computer device, characterized in that: The computer device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for disambiguating a person's name by fusing multi-source information as described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the method for disambiguating a person's name by fusing multi-source information according to any one of claims 1 to 5 is implemented.

Images