Method and device for cold-start author homonym disambiguation in document library based on multi-feature fusion

Through the multi-feature fusion method, multiple feature matrices in the literature library were extracted and fused, combined with the DBSCAN clustering algorithm, the problem of insufficient extraction of literature features in the existing technology was solved, and the effect of cold-starting authors' disambiguation of the same name was significantly improved. and accuracy.

CN114661902BActive Publication Date: 2024-12-03BEIJING KNOWLEDGE ATLAS TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210204836.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2022-01-10
Filing Date
2022-03-03
Publication Date
2024-12-03
Estimated Expiration
2042-03-03

AI Technical Summary

Technical Problem

The existing literature library cold-start author's disambiguation method of the same name is not rich enough to extract the literature features, resulting in poor classification effects, especially in the interactive features and structured feature distribution between documents, affecting the clustering effect.

Method used

Multi-feature fusion method is adopted, including shallow semantic information feature extraction network, deep semantic information feature extraction network, structural relationship information feature extraction network and manual interactive information feature extraction network. Through the technology such as word2vec, doc2vec, BERT and LightGBM, combined with the Heterogeneous graphs and metapaths are randomly wandering, multiple feature matrices are obtained, and the literature cluster of the only author is output through feature-weighted fusion and DBSCAN clustering algorithm.

Benefits of technology

The cold-start disambiguation effect is significantly improved, especially in the absence of fields, the accuracy rate is improved, the information characteristics of the influence of clustering results are fully considered, and the accuracy and efficiency of the authors of the same name are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114661902B_ABST
    Figure CN114661902B_ABST
Patent Text Reader

Abstract

The present invention proposes a method and device for cold-start author homonym disambiguation in a document library based on multi-feature fusion. The method constructs a cold-start author homonym disambiguation model for the document library, extracts word vector information, deep semantic information, structural information, interaction information obtained through machine learning training, and manually extracted interaction information of the document through the model, and performs feature fusion to output a document set of several clusters; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents. Through the present invention, the information features that affect the clustering results can be fully considered, and the cold-start disambiguation effect can be improved by fusing multiple different types of features, and the accuracy of disambiguation in missing fields is significantly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method, device, computer equipment and storage medium for cold-start author homonym disambiguation in a document library based on multi-feature fusion. Background Art

[0002] With the development of information technology and society, the amount of various literature data has increased dramatically, and the best way to preserve these documents is to use digital libraries. At present, digital libraries have stored a huge amount of literature, such as literature and patent data. While storing a huge amount of data, another problem has arisen. The number of authors with the same name in these documents is also increasing, and we need to find the correct authors with the same name for these documents. In literature search, the problem of authors with the same name will not only reduce the efficiency of literature retrieval, but also affect the correct attribution of scholars' research work. Therefore, it is very necessary to disambiguate authors with the same name. Disambiguation of the same name has always been a basic problem that needs to be solved urgently by online academic mining platforms such as Google Scholar, DBLP and Aminer. Solving the disambiguation of the same name can provide strong support for many specific applications of these platforms, such as scientific literature management, literature retrieval, and author search.

[0003] At present, the research on the algorithm for disambiguating authors with the same name in the document library can be divided into two types. One is cold-start disambiguation of existing documents in the database (also known as "batch disambiguation"), and the other is to incrementally disambiguate a newly arrived document record based on the existing disambiguation results, or assign it to an author that already exists in the database, or create a new author for it. Cold-start disambiguation mainly classifies all undisambiguated document sets with the same author name, and classifies them according to the category to which the author belongs. For example, there are three authors named "Zhang San" and 10 documents. Cold-start disambiguation is to divide the 10 documents into three categories, which belong to "Zhang San (1)", "Zhang San (2)", and "Zhang San (3)" respectively.

[0004] Existing cold-start disambiguation methods mainly use unsupervised learning methods to classify data according to the internal features of the data. The internal features of the data are mainly extracted by manually extracting the interaction features between the data or using the structural features between the data. Such classification methods place too much emphasis on the collaborators and institutional information of the document, and will isolate documents that do not have co-authors or common institutions with other authors. There are also a small number of cold-start disambiguation methods that use supervised learning methods. By training on labeled data sets, a similarity distance is given to any two documents, and the batch of documents is classified through the distance matrix. The clustering effect obtained in this way is too dependent on the feature distribution of the labeled data set, and some strong attribute features such as collaborators will be ignored. One of the main shortcomings of the current cold-start disambiguation algorithm of the document library is that the extraction of document features is not rich enough. By enriching the document features, the effect of the cold-start disambiguation algorithm can be improved. Summary of the invention

[0005] The present invention provides a method, device, computer equipment and storage medium for cold-start author homonym disambiguation in a document library based on multi-feature fusion, aiming to fully consider the information features that have an impact on the clustering results and improve the cold-start disambiguation effect by fusing multiple different types of features.

[0006] To this end, the first object of the present invention is to propose a cold-start author homonym disambiguation method for a document library based on multi-feature fusion, comprising:

[0007] Constructing a document library cold-start author homonym disambiguation model; wherein the document library cold-start author homonym disambiguation model comprises a feature extraction module, a feature weighted fusion module and a clustering module;

[0008] Obtaining a manually annotated homonym disambiguation dataset, performing data preprocessing, and inputting the preprocessed homonym disambiguation dataset into the document library cold-start author homonym disambiguation model for training;

[0009] The document set to be subjected to author homonym disambiguation is input into the trained cold-start author homonym disambiguation model of the document library, and a document set of several clusters is output; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents.

[0010] Among them, the manually annotated homonym disambiguation dataset uses the WhoIsWho dataset (na-v1); the preprocessing steps include:

[0011] Removing special characters from the homonym disambiguation dataset;

[0012] For documents with more than 100 authors, only the first 100 authors were retained;

[0013] The missing fields of the homonym disambiguation dataset are recorded, and the missing fields are marked.

[0014] Among them, the preprocessed data is recorded as:

[0015] D{title, abstract, author, organization, year, no_abstract, no_org, no_year}

[0016] Among them, no_abstract represents the missing abstract information in the document, no_year represents the missing publication date information in the document, and no_org represents the missing institution of the author to be disambiguated. If it is missing, it is set to 1, and if it is not missing, it is set to 0.

[0017] Among them, the feature extraction module includes a shallow semantic information feature extraction network, a deep semantic information feature extraction network, a structural relationship information feature extraction network, a training interaction information extraction network, and an artificially extracted interaction information feature extraction network;

[0018] The shallow semantic information feature extraction network is word2vec and doc2vec; word2vec is used to train the word vectors of all words in the title, abstract, and organization in the homonym disambiguation dataset, and the word vectors of the words that have appeared are obtained; the word vectors of the title, abstract, and the organization to which the disambiguation author belongs in each document are averaged as the first vector information of the corresponding document; the title and abstract words of each document are trained by doc2vec, and the second vector information of each document is directly obtained through iterative training;

[0019] The deep semantic information feature extraction network adopts the BERT pre-training model to obtain the semantic vectors of title and abstract as the third vector information of each document;

[0020] The structural relationship information feature extraction network adopts a method of constructing a heterogeneous graph and random walking based on a meta-path to construct a heterogeneous network and obtain the fourth vector information of each document.

[0021] The training interactive information extraction network calculates the cosine similarity between the first vector information, the second vector information, the third vector information and the fourth vector information as the distance between the two documents, and obtains four distance matrices E1, E2, E3 and E4.

[0022] The LightGBM model is used to train the interactive information extraction network. Any two document labels belonging to the same cluster are set to 1, otherwise they are set to 0. The training is performed as a classification task, and the ratio of positive and negative labels is close to 1:1. After the training is completed, the similarity score of any two documents is calculated as the distance between the two documents. The distance between any two documents is obtained, and the distances between all two documents are saved in a matrix format as the distance matrix E5.

[0023] Among them, the artificially extracted interactive information feature extraction network calculates the rarity of all authors, finds the weights of all collaborators and adds them up to get the first score; calculates the Jaccard similarity between any two documents after title segmentation to get the second score; adds the first score and the second score as the distance between any two documents with artificially extracted features; saves the distances between all documents in a matrix format as the distance matrix E6.

[0024] Among them, the feature weighted fusion module fuses and weights the distance matrices E1, E2, E3, E4, E5, and E6, as shown in formula (1):

[0025]

[0026] in, is the normalized value of the evaluation score, and the evaluation index is Macro Pairwise-F1; as shown in formula (2):

[0027]

[0028] Among them, the clustering module is used to cluster the fusion matrix emb; among them, the clustering method uses DBSCAN (density-based clustering algorithm) to cluster the fusion matrix.

[0029] The second object of the present invention is to propose a document library cold start author homonym disambiguation device based on multi-feature fusion, comprising:

[0030] A model building module is used to build a document library cold-start author homonym disambiguation model; wherein the document library cold-start author homonym disambiguation model includes a feature extraction module, a feature weighted fusion module and a clustering module;

[0031] A model training module is used to obtain a manually annotated homonym disambiguation dataset and perform data preprocessing, and input the preprocessed homonym disambiguation dataset into the document library cold start author homonym disambiguation model for training;

[0032] The author homonym disambiguation module is used to input the document set to be disambiguated into the trained document library cold-start author homonym disambiguation model, and output a document set of several clusters; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents.

[0033] The third object of the present invention is to provide a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method of the aforementioned technical solution is implemented.

[0034] The fourth object of the present invention is to provide a non-temporary computer-readable storage medium having a computer program stored thereon, which implements the method of the aforementioned technical solution when the computer program is executed by a processor.

[0035] Different from the prior art, the present invention provides a document library cold-start author homonym disambiguation method based on multi-feature fusion, constructs a document library cold-start author homonym disambiguation model, extracts word vector information, deep semantic information, structural information, interactive information obtained through machine learning training, and manually extracted interactive information of the document through the model, and performs feature fusion to output a document set of several clusters; wherein, the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents. Through the present invention, it is possible to fully consider the information features that affect the clustering results, improve the cold-start disambiguation effect through the fusion of multiple different types of features, and significantly improve the accuracy of disambiguation in field missing. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention and / or additional aspects and advantages will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0037] Figure 1 It is a flow chart of a method for cold-starting author homonym disambiguation in a document library based on multi-feature fusion provided by the present invention.

[0038] Figure 2 It is a schematic diagram of model processing logic in a document library cold-start author homonym disambiguation method based on multi-feature fusion provided by the present invention.

[0039] Figure 3 It is a schematic diagram of a meta-path random walk process in a cold-start author homonym disambiguation method for a document library based on multi-feature fusion provided by the present invention.

[0040] Figure 4 It is a structural schematic diagram of a document library cold-start author homonym disambiguation device based on multi-feature fusion provided by the present invention.

[0041] Figure 5It is a structural schematic diagram of a non-temporary computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0042] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0043] Figure 1 The present invention provides a method for cold-start author homonym disambiguation in a document library based on multi-feature fusion. The method comprises the following steps:

[0044] Step 101, constructing a document library cold-start author homonym disambiguation model; wherein the document library cold-start author homonym disambiguation model includes a feature extraction module, a feature weighted fusion module and a clustering module.

[0045] The technical purpose of the technical solution of the present invention is to perform author homonym disambiguation on a batch of documents with the same author in a specified format. The specific method is to construct a cold-start author homonym disambiguation model for a document library, extract feature information from the document collection and input it into the model. The model output can further classify the document collection and complete author homonym disambiguation. The model structure is as follows Figure 2 shown.

[0046] Specifically, in the cold-start author homonym disambiguation model for the document library constructed by the present invention, the feature extraction module includes a shallow semantic information feature extraction network, a deep semantic information feature extraction network, a structural relationship information feature extraction network, a trained interactive information extraction network, and an artificially extracted interactive information feature extraction network.

[0047] The shallow semantic information feature extraction network uses word2vec and doc2vec; the former trains word vectors for all words in the title, abstract, and organization of all documents to obtain the word vectors of the words that have appeared, and finally calculates the average of the word vectors of the title, abstract, and the organization to which the disambiguated author belongs (not all author organizations) in each document as the vector information of the document; the latter trains the title and abstract words of each document and directly obtains the vector information of each document through iterative training.

[0048] The deep semantic information feature extraction network obtains the semantic vectors of title and abstract through the BERT pre-training model. OAG-BERT is used here. Referring to previous literature information, no one else has used OAG-BERT for cold start disambiguation. OAG-BERT integrates structured knowledge from multiple open academic graphs (Open Academic Graph). Using OAG-BERT is more suitable for our task of extracting semantic information from literature.

[0049] The structural relationship information feature extraction network uses a method of building a heterogeneous graph and random walk based on meta-paths. The process is as follows Figure 3 As shown, the circled p r Represents a paper. The solid arrow on the left represents the CoAuthor relationship, and the dotted arrow represents the CoOrg relationship. A heterogeneous network is constructed by random walks based on the meta-path (paper→CoAuthor→paper→CoOrg→paper). The numbers above the arrows represent different weights, that is, the number of co-collaborators (or co-institutions). The path set obtained by the walk is used as text, and the Skip-Gram model in word2vec is used to train the word vector to obtain the embedding of each paper.

[0050] The LightGBM model is used to train the interactive information extraction network, and the TF-IDF method is used to manually extract the interactive information feature extraction network.

[0051] Step 102, obtaining a manually annotated homonym disambiguation dataset, performing data preprocessing, and inputting the preprocessed homonym disambiguation dataset into the document library cold-start author homonym disambiguation model for training.

[0052] The manually annotated homonym disambiguation dataset in the present invention uses the WhoIsWho dataset (na-v1); the preprocessing steps include:

[0053] Removing special characters from the homonym disambiguation dataset;

[0054] For documents with more than 100 authors, only the first 100 authors were retained;

[0055] The missing fields of the homonym disambiguation dataset are recorded, and the missing fields are marked.

[0056] D{title, abstract, author, organization, year, no_abstract, no_org, no_year}

[0057] Among them, no_abstract represents the missing abstract information in the document, no_year represents the missing publication date information in the document, and no_org represents the missing institution of the author to be disambiguated. If it is missing, it is set to 1, and if it is not missing, it is set to 0.

[0058] The shallow semantic information feature extraction network is word2vec and doc2vec; word2vec is used to train the word vectors of all words in the title, abstract, and organization in the homonym disambiguation dataset to obtain the word vectors of the words that have appeared; the word vectors of the title, abstract, and the organization to which the disambiguation author belongs in each document are averaged as the first vector information of the corresponding document;

[0059] The word vectors of the title, organization, and abstract of all documents are trained using the word2vec algorithm to obtain the embedding information of each word. Then, the embeddings of the title, abstract, and the organization to which the author of the same name belongs (not all the author organizations) of each document are averaged to obtain the embedding information of each document, which is recorded as Where i represents the i-th document. The formula is as follows:

[0060]

[0061] where w j is the word vector of the i-th word in the title, abstract, and the organization to which the author of the same name belongs in the i-th document. j The word vector is obtained by training word2vec on all the data in the literature.

[0062] The title and abstract words of each document are trained by doc2vec, and the second vector information of each document is directly obtained through iterative training; the text embedding information of each document is trained by doc2vec algorithm for the title and abstract of all documents, and the information is saved and recorded as This information is obtained through direct training of doc2vec, and i represents the i-th document.

[0063] The deep semantic information feature extraction network adopts the BERT pre-training model to obtain the semantic vectors of the title and abstract as the third vector information of each document; the present invention uses OAG-BERT to extract the semantic embedding of the title and abstract of each document. If the abstract is too long, the latter half of the information is truncated (only the first 128 words are retained). The obtained semantic vector is recorded as i represents the i-th document.

[0064] The structural relationship information feature extraction network adopts a method of constructing a heterogeneous graph and random walking based on a meta-path to construct a heterogeneous network and obtain the fourth vector information of each document.

[0065] For each name that needs to be disambiguated, the relationship between all the corresponding documents is extracted to construct a heterogeneous network. The heterogeneous network contains one type of node: document (each document represents a node) and two types of edges: CoAuthor and CoOrg. CoAuthor represents the common author between two documents, and CoOrg represents the common organization between two documents. The organization adopts fuzzy matching, such as "School of Information Science and Technology, Tsinghua University" and "Tsinghua University" are considered to be one organization. The number of co-authors and the number of common organizations represent the weights of CoAuthor and CoOrg respectively.

[0066] In the meta-path-based random walk process, we walk along the meta-path of paper→CoAuthor→paper→CoOrg→paper to generate paths consisting of document IDs. We use these paths as the input of the Skip-Gram model to complete the embedding process of the document IDs and save the information, which is recorded as i represents the i-th document.

[0067] The training interactive information extraction network calculates the cosine similarity between the first vector information, the second vector information, the third vector information and the quantity information, as the distance between the two documents, and obtains four distance matrices E1, E2, E3, and E4.

[0068] The LightGBM model is used to train the interactive information extraction network. The labels of any two documents belonging to the same cluster are set to 1, otherwise they are set to 0. The training is performed as a classification task, and the ratio of positive and negative labels is close to 1:1. After the training is completed, the similarity score of any two documents is calculated as the distance between the two documents. The distance between any two documents is obtained, and the distances between all two documents are saved in a matrix format as the distance matrix E5.

[0069] Manually extract interactive information feature extraction network calculates the rarity of all authors through TF-IDF, finds the weights of all collaborators and adds them up to get the first score; check whether the institutions of any two disambiguated authors are the same, if they are the same, assign 0.5, otherwise assign 0. Calculate the Jaccard similarity between any two documents after title segmentation to get the second score; add the first score and the second score as the distance between any two documents with manually extracted features; save the distances between all documents in a matrix format as the distance matrix E6.

[0070] The feature weighted fusion module fuses and weights the distance matrices E1, E2, E3, E4, E5, and E6, as shown in formula (1):

[0071]

[0072] in, is the normalized value of the evaluation score, and the evaluation index is Macro Pairwise-F1; as shown in formula (2):

[0073]

[0074] The clustering module is used to cluster the fusion matrix emb; wherein the clustering method uses DBSCAN (density-based clustering algorithm) to cluster the fusion matrix.

[0075] Step 103, input the document set to be subjected to author homonym disambiguation into the trained cold-start author homonym disambiguation model of the document library, and output a document set of several clusters; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents.

[0076] The present invention fully considers various characteristics between documents and the documents themselves, and the final cold start result is more effective than the previous method. Moreover, due to the rich features, it can also achieve not bad results when processing documents in some extreme cases, such as missing fields.

[0077] In order to implement the embodiment, the present invention also proposes a security protection alarm information processing device, such as Figure 4 As shown, including:

[0078] A model building module 310 is used to build a document library cold-start author homonym disambiguation model; wherein the document library cold-start author homonym disambiguation model includes a feature extraction module, a feature weighted fusion module and a clustering module;

[0079] The model training module 320 is used to obtain a manually annotated homonym disambiguation dataset, perform data preprocessing, and input the preprocessed homonym disambiguation dataset into the document library cold-start author homonym disambiguation model for training;

[0080] The author homonym disambiguation module 330 is used to input the document set to be disambiguated with the author homonym into the trained cold-start author homonym disambiguation model of the document library, and output a document set of several clusters; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents.

[0081] In order to implement the embodiment, the present invention also proposes another computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, cold-start author homonym disambiguation in the document library as in the embodiment of the present invention is implemented.

[0082] like Figure 4 As shown, the non-transitory computer-readable storage medium includes a memory 810 of instructions and an interface 830, and the instructions can be executed by a processor 820 of a cold-start author homonym disambiguation device according to a document library to complete the method. Optionally, the storage medium can be a non-transitory computer-readable storage medium, for example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0083] In order to implement the embodiment, the present invention further proposes a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the document library cold-start author homonym disambiguation is implemented as in the embodiment of the present invention.

[0084] 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 representation of the terms does 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, without contradiction.

[0085] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0086] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

[0087] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0088] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the described embodiment, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0089] A person skilled in the art may understand that all or part of the steps of implementing the method of the embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0090] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0091] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present invention have been shown and described above, it can be understood that the embodiments are exemplary and cannot be construed as limiting the present invention. A person of ordinary skill in the art may change, modify, replace and modify the embodiments within the scope of the present invention.

Claims

1. A method for cold-start author homonym disambiguation in a document library based on multi-feature fusion, characterized in that: include: Constructing a document library cold-start author homonym disambiguation model; wherein the document library cold-start author homonym disambiguation model comprises a feature extraction module, a feature weighted fusion module and a clustering module; Obtaining a manually annotated homonym disambiguation dataset, performing data preprocessing, and inputting the preprocessed homonym disambiguation dataset into the document library cold-start author homonym disambiguation model for training; Inputting the document set to be subjected to author homonym disambiguation into the trained cold-start author homonym disambiguation model of the document library, and outputting document sets of several clusters; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents; The feature extraction module includes a shallow semantic information feature extraction network, a deep semantic information feature extraction network, a structural relationship information feature extraction network, a training interaction information extraction network and an artificially extracted interaction information feature extraction network; The shallow semantic information feature extraction network is word2vec and doc2vec; word2vec is used to train the word vectors of all words in the title, abstract, and organization in the homonym disambiguation dataset, and the word vectors of the words that have appeared are obtained; the word vectors of the title, abstract, and the organization to which the disambiguation author belongs in each document are averaged as the first vector information of the corresponding document; the title and abstract words of each document are trained by doc2vec, and the second vector information of each document is directly obtained through iterative training; The deep semantic information feature extraction network adopts the BERT pre-training model to obtain the semantic vectors of title and abstract as the third vector information of each document; The structural relationship information feature extraction network adopts a method of constructing a heterogeneous graph and random walking based on a meta-path to construct a heterogeneous network and obtain the fourth vector information of each document.

2. The cold-start author homonym disambiguation method for document library based on multi-feature fusion according to claim 1 is characterized in that: The manually annotated homonym disambiguation dataset uses the WhoIsWho dataset (na-v1); The preprocessing steps include: Removing special characters from the homonym disambiguation dataset; For documents with more than 100 authors, only the first 100 authors were retained; The missing fields of the homonym disambiguation dataset are recorded, and the missing fields are marked.

3. The method for cold-start author homonym disambiguation in a document library based on multi-feature fusion according to claim 2 is characterized in that: The preprocessed data is recorded as: in, Indicates that the abstract information is missing in the document. Indicates that the publication date information is missing in the literature. Indicates the missing institution of the author to be disambiguated. If missing, set it to 1, otherwise set it to 0.

4. The cold-start author homonym disambiguation method for document library based on multi-feature fusion according to claim 1 is characterized in that: The training interactive information extraction network calculates the cosine similarity between the first vector information, the second vector information, the third vector information and the fourth vector information, as the distance between the two documents, and obtains four distance matrices E1, E2, E3, and E4; the training interactive information extraction network uses the LightGBM model, sets the labels of any two documents belonging to the same cluster to 1, otherwise to 0, and performs training as a classification task, and the ratio of positive and negative labels is close to 1:1; after the training is completed, the similarity score of any two documents is calculated as the distance between the two documents; the distance between any two documents is obtained, and the distance between all two documents is saved in a matrix form as a distance matrix E5.

5. The method for cold-start author homonym disambiguation in a document library based on multi-feature fusion according to claim 4 is characterized in that: The artificially extracted interactive information feature extraction network calculates the rarity of all authors, finds the weights of all collaborators and adds them up to obtain a first score; calculates the Jaccard similarity between any two documents after title segmentation to obtain a second score; adds the first score and the second score as the distance between any two documents of the artificially extracted features; and saves the distances between all pairs of documents in a matrix format as a distance matrix E6.

6. The method for cold-start author homonym disambiguation in a document library based on multi-feature fusion according to claim 5, characterized in that: The feature weighted fusion module fuses and weights the distance matrices E1, E2, E3, E4, E5, and E6, as shown in formula (1): (1) in, is the normalized value of the evaluation score, and the evaluation index is Macro Pairwise-F1; as shown in formula (2): (2) The clustering module is used to fusion matrix Perform clustering; the clustering method uses DBSCAN (density-based clustering algorithm) to cluster the fusion matrix.

7. A document library cold start author homonym disambiguation device based on multi-feature fusion, characterized in that: include: A model building module is used to build a document library cold-start author homonym disambiguation model; wherein the document library cold-start author homonym disambiguation model includes a feature extraction module, a feature weighted fusion module and a clustering module; A model training module is used to obtain a manually annotated homonym disambiguation dataset and perform data preprocessing, and input the preprocessed homonym disambiguation dataset into the document library cold start author homonym disambiguation model for training; The author homonym disambiguation module is used to input the document set to be disambiguated into the trained cold-start author homonym disambiguation model of the document library, and output a document set of several clusters; wherein the document author in each cluster is unique, and the number of clusters is less than or equal to the number of input documents; The feature extraction module includes a shallow semantic information feature extraction network, a deep semantic information feature extraction network, a structural relationship information feature extraction network, a training interaction information extraction network, and an artificially extracted interaction information feature extraction network; The shallow semantic information feature extraction network is word2vec and doc2vec; word2vec is used to train the word vectors of all words in the title, abstract, and organization in the homonym disambiguation dataset, and the word vectors of the words that have appeared are obtained; the word vectors of the title, abstract, and the organization to which the disambiguation author belongs in each document are averaged as the first vector information of the corresponding document; the title and abstract words of each document are trained by doc2vec, and the second vector information of each document is directly obtained through iterative training; The deep semantic information feature extraction network adopts the BERT pre-training model to obtain the semantic vectors of title and abstract as the third vector information of each document; The structural relationship information feature extraction network adopts a method of constructing a heterogeneous graph and random walking based on a meta-path to construct a heterogeneous network and obtain the fourth vector information of each document.

8. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.