Incremental paper author disambiguation method based on pre-trained model and decision tree

Abstract

The application discloses an incremental paper homonym author disambiguation method based on a pre-training model and a decision tree, belongs to the technical field of neural networks and author homonym disambiguation, and uses artificially defined features and XLNet extracted features in combination for feature extraction. First, artificially defined features are used to extract information of author names, institutions and other fields in a paper, and XLNet is used to extract information of a paper title, an abstract and other fields. Then, XGBoost and the extracted features are used to predict a correct author to which each paper should belong. Papers that are not allocated in this step are subjected to post-processing of agglomerative clustering, a main cluster is obtained as a new author paper set, and a prediction result uses an accuracy, a recall rate and an F1 value. The application can more fully extract information of a paper, has strong interpretability, has strong robustness, and has excellent performance in processing data with more noise points and more complete paper information.

Description

Technical Field

[0001] This invention relates to the field of neural networks and author name disambiguation, and in particular to an incremental paper author name disambiguation method based on pre-trained models and decision trees. Background Technology

[0002] With the development of information technology, the scale, storage methods, and access methods of information have undergone significant changes, leading to the emergence of various academic search engines such as CNKI, Aminer, DBLP, Microsoft Scholar, Google Scholar, and PubMed. These search engines have become the main channels for scholars to obtain various academic papers. Of course, when using search engines, people don't just use keywords; they also search by author names. However, due to the development of various disciplines and the increase in researchers in recent years, the amount of data in the databases of various search engines has also increased, resulting in a significant increase in the number of authors with the same name. To address these issues, research on author name disambiguation has emerged.

[0003] Author name disambiguation is a research area within entity disambiguation. Initially, research on author name disambiguation focused on cold-start disambiguation based on existing documents in the database, also known as batch disambiguation. This method requires disambiguating all existing documents in the database, but it is computationally intensive, time-complex, and the manually corrected data is difficult to save after the disambiguation system is updated. In reality, a large number of different documents are added to the database every day, making the timeliness of such disambiguation algorithms very poor. Furthermore, if the disambiguation algorithm is recalculated every time a new document is added to the database, the server load becomes enormous. Therefore, research on incremental disambiguation algorithms is of significant importance.

[0004] Incremental disambiguation processes each new data entry individually, building upon existing disambiguation results, rather than re-disambiguating the entire database after each new entry. The processing outcome for new data is either assignment to an existing author or creation of a new author. Compared to cold-start disambiguation, incremental disambiguation avoids frequent re-disambiguation of the entire database, thus reducing the algorithm's time complexity and server load. While incremental disambiguation has many advantages, it also has drawbacks, namely, it requires a large and accurate dataset that has already undergone disambiguation.

[0005] In conclusion, author name disambiguation is a challenging and labor-intensive data management task. Eliminating ambiguity caused by identical author names can make search engine results more accurate, which in turn affects the construction of knowledge graphs and personalized services. Therefore, researching relevant algorithms to solve the problem of identical author names has significant research value. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide an incremental method for disambiguating authors with the same name in papers based on pre-trained models and decision trees. It addresses the problem of insufficient utilization of paper information by taking two approaches: incremental disambiguation and full utilization of paper information, with the aim of making full use of information and focusing on newly added papers.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] An incremental paper same-name disambiguation method based on a pre-trained model and decision tree includes the following steps:

[0009] S1, Construct the dataset and perform data preprocessing on author names, candidate set IDs, paper IDs, and paper information;

[0010] S2: Extract the first feature without semantic information fields using manually defined rules, extract the second feature with semantic information fields using the XLNet pre-trained model, and merge the first and second features together to obtain the desired feature vector.

[0011] S3, input the feature vector obtained in S2 into XGBoost to obtain the probability that the paper to be assigned belongs to the candidate set, and determine whether to assign the paper to be assigned to the candidate set based on this.

[0012] S4. Construct an incremental disambiguation post-processing framework based on agglomerative hierarchical clustering to perform post-processing operations on the papers that were not allocated in incremental S3.

[0013] A further improvement of the technical solution of the present invention is that: the first feature includes at least the author's name and institution, and the second feature includes at least the paper title and abstract.

[0014] A further improvement to the technical solution of the present invention is that S2 specifically includes:

[0015] 2.1 Compare the author information of the paper to be assigned with all papers in the candidate set. The values ​​are represented as follows:

[0016] count a The number of authors with the same name. This value is incremented by 1 each time an author with the same name appears. The same operation is performed when an author's name appears repeatedly.

[0017] count oa The number of authors with the same name and from the same organization. This value is incremented by 1 each time an author with the same name and from the same organization appears. The same operation of incrementing the value will also occur when an author with the same name and from the same organization appears repeatedly.

[0018] r a The ratio of the number of authors with the same name to the number of papers in the candidate set, i.e.

[0019] r oa The ratio of authors with the same name and from the same organization to the number of papers in the candidate set, i.e.

[0020] count ca The number of co-authors. This value is incremented by 1 each time a co-author appears. If an author appears repeatedly, this value is not affected.

[0021] r ca The number of co-authors and the number of authors in the paper to be assigned: The ratio, i.e.

[0022] t ca Co-authors in this candidate set c k The total number of times it appears in the text;

[0023] The number of times co-authors appear and the candidate set c k Total number of authors in China t a The ratio, i.e.

[0024] After the comparison is completed, these values ​​are arranged together to form a vector, which is the extracted author name-related feature;

[0025] 2.2 Compare the author affiliation information of the paper to be assigned with all papers in the candidate set. The values ​​are represented as follows:

[0026] count org Papers to be assigned a With candidate set c k The number of all identical institutions in the country;

[0027] r org : Number of identical institutions and total number of institutions aorg The ratio between them, that is

[0028] Jaccard max After word segmentation and merging, the paper to be assigned (p)a With candidate set c k The maximum value of the Jaccard similarity coefficient among all papers in the series, i.e.

[0029] Jaccard mean After word segmentation and merging, the paper to be assigned (p) a With candidate set c k The mean of the Jaccard similarity coefficients among all papers in the paper, i.e.

[0030] Jaccard pooling After word segmentation and merging, the Jaccard similarity coefficient after processing with the Gaussian kernel function is an n-dimensional vector, where n represents the number of input center points;

[0031] r′ max After word segmentation and merging, the paper to be assigned (p) a With candidate set c k The maximum value of the ratio of the number of identical institutions among all papers to the total number of institutions, i.e.

[0032] r′ mean After word segmentation and merging, the paper to be assigned (p) a With candidate set c k The average of the ratio of the number of papers from the same institution to the total number of institutions.

[0033] r′ pooling After word segmentation and merging, the ratio of the number of identical structures to the total number of structures after processing with the Gaussian kernel function is an n-dimensional vector, where n represents the number of input center points;

[0034] After the comparison is completed, these values ​​are arranged together to form a vector two, which is the extracted author institution-related feature;

[0035] 2.3 Compare the title information of the paper to be assigned with all papers in the candidate set. The values ​​are represented as follows:

[0036] count title After the word segmentation and merging process, identical words in the unassigned word list p a Number of times it appears in the text;

[0037] count′ title After the word segmentation and merging process, identical words are in the candidate set c. k The number of times it appears in all papers;

[0038] r title′: After the word segmentation and merging process, the number of occurrences of the same word in the to-be-assigned paper p a and the total number of all words in the to-be-assigned paper p a The ratio of, that is is

[0039] r′ title′ : After the word segmentation and merging process, the number of occurrences of the same word in all papers in the candidate set c k and the total number of all words in the candidate set c k The ratio of, that is is

[0040] jaccard pooling : After the word segmentation and merging process, the Jaccard similarity coefficient after being processed by the Gaussian kernel function, which is an n-dimensional vector, where n represents the number of input center points;

[0041] cos pooling : Use XLNet to extract the title features of the paper, and then calculate the cosine similarity between the title of the to-be-assigned paper p a and the title of each paper in the candidate set c k , and then obtain cos pooling through the Gaussian kernel function processing;

[0042] After the comparison is completed, these values are arranged together to form a vector three, and the vector three is the non-semantic feature of the extracted paper title;

[0043] 2.4. Use the XLNet pre-trained model to extract semantic features for the paper fields containing semantic information;

[0044] 2.5. Combine the vector one, vector two, vector three obtained in 2.1, 2.2, 2.3 and the semantic features extracted in 2.4 to obtain the final feature vector.

[0045] A further improvement of the technical solution of the present invention lies in: S3 specifically includes:

[0046] 3.1. Input the features obtained in step 2 into the XGBoost decision tree to determine whether the to-be-assigned paper is assigned to the corresponding author or not assigned;

[0047] Assume that there are i papers in the candidate set corresponding to the current to-be-assigned paper. Then, through step 2, i feature vectors will be obtained. Input these feature vectors into XGBoost, and corresponding i scores score can be obtained, 0 < score < 1. Denote the candidate set with the highest probability as the most likely candidate set, and the highest score as score_max;

[0048] 3.2 For simpler cases, a threshold is used for judgment, generally 0.9 is selected; if score_max≥0.9, then the paper to be assigned is assigned to the candidate set corresponding to score_max, otherwise the paper to be assigned is recorded as an unassigned paper. After operating on all papers to be assigned, all unassigned papers are summarized and the set is recorded as the unassigned paper set.

[0049] 3.3 Treat each unassigned paper as a cluster, and use the method in step 2 to compare these clusters pairwise to obtain several feature vectors;

[0050] 3.4 Input the feature vector obtained in 3.3 into XGBoost or the trained MLP for probability prediction. The two clusters with the highest probability are regarded as the two closest clusters and merged together. When the number of papers in the largest cluster is greater than 5, it is considered to be the main cluster and hierarchical clustering is stopped.

[0051] 3.5. The remaining papers in the unassigned candidate set are then subjected to the incremental disambiguation operation mentioned in 3.1 and 3.2 on the main cluster obtained in 3.4. Papers that fail to be assigned are returned to the unassigned candidate set, while those that succeed are directly added to the main cluster. The final incremental main cluster is then used as a new candidate set, thus completing a full round of incremental disambiguation operation.

[0052] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0053] 1. This invention addresses the problem that most disambiguation methods fail to fully utilize the information in academic papers by proposing a feature extraction method that combines manually defined features with features extracted by XLNet. Manually defined features are extracted from the paper that do not require semantic feature extraction, such as author names and the number of institutions. XLNet, on the other hand, is used to extract features that require semantic feature extraction, such as semantic information from the paper's abstract. Using XLNet outperforms other models such as BERT, and comparisons with other incremental disambiguation methods demonstrate that this feature extraction approach helps improve the accuracy of disambiguation results.

[0054] 2. This invention addresses the issue of low accuracy caused by the use of traditional prediction models in some methods. It replaces traditional neural network models with XGBoost as the core of the matching module for prediction, and uses a decision module to determine whether to allocate the paper. The performance of using XGBoost is significantly better than that of using traditional neural network models.

[0055] 3. This invention addresses the problem that incremental disambiguation cannot allocate all papers. It proposes a cold-start disambiguation method based on agglomerative hierarchical clustering. This cold-start disambiguation method is placed after the incremental disambiguation method to post-process papers that were not handled by incremental disambiguation, thus improving the overall results of the incremental disambiguation method. Adding this cold-start disambiguation framework does indeed more effectively perform incremental disambiguation on authors with a smaller number of papers in their candidate sets. Attached Figure Description

[0056] Figure 1 This is a diagram of the incremental disambiguation framework in an embodiment of the present invention;

[0057] Figure 2 This is a flowchart of the feature extraction process in an embodiment of the present invention;

[0058] Figure 3 This is a decision graph in incremental disambiguation in an embodiment of the present invention;

[0059] Figure 4 This is a cold start disambiguation framework diagram in an embodiment of the present invention;

[0060] Figure 5 This is an overall framework diagram of the incremental paper same-name disambiguation method based on pre-trained model and decision tree in an embodiment of the present invention;

[0061] Figure 6 Analysis of the number of papers in the AMiner training set in this embodiment of the invention Figure 1 ;

[0062] Figure 7 Analysis of the number of papers in the AMiner test set in this embodiment of the invention Figure 2 ;

[0063] Figure 8 A comparison of the time consumption of each model in the embodiments of the present invention. Figure 1 ;

[0064] Figure 9 A comparison of the time consumption of each model in the embodiments of the present invention. Figure 2 ;

[0065] Figure 10 This is a comparison chart of the time consumption of each method in the embodiments of the present invention on the test set;

[0066] Figure 11 This is a comparison chart of the time consumption of each method in the embodiments of the present invention on the complete test set. Detailed Implementation

[0067] This application provides an incremental method for disambiguating authors with the same name in academic papers based on a pre-trained model and decision tree. This solves the problem in existing technologies where papers with the same name are not assigned to the correct authors, resulting in low accuracy when retrieving papers by name. The general idea is as follows:

[0068] First, this invention addresses the problem of insufficient utilization of information in academic papers by proposing a feature extraction method based on a combination of an XLNet pre-trained model and manually defined rules. This method first uses manually defined features to extract information such as author names and institutions from the papers, and then uses XLNet to extract information such as paper titles and abstracts. Finally, it uses XGBoost and the extracted features to predict the correct author to whom each paper should belong.

[0069] Secondly, this invention addresses the problem that incremental disambiguation cannot assign all papers. It proposes a cold-start disambiguation method based on agglomerative hierarchical clustering, which is placed after the incremental disambiguation method to post-process papers that failed to be assigned. This method performs agglomerative clustering on the unassigned papers, and then adds the papers to the main cluster through incremental disambiguation to obtain the main cluster as a new author.

[0070] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0071] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0072] like Figure 1-5 As shown, the incremental paper same-name disambiguation method based on pre-trained models and decision trees includes the following steps:

[0073] S1, Construct the dataset and perform data preprocessing on author names, candidate set IDs, paper IDs, paper information, etc.

[0074] S2: Extract the first feature (which does not contain semantic information, such as author name and institution) using manually defined rules; extract the second feature (which contains semantic information, such as paper title and abstract) using the XLNet pre-trained model; and combine the first and second features to obtain the desired feature vector.

[0075] 2.1 Compare the author information of the paper to be assigned with all papers in the candidate set. The values ​​are shown in Table 1 below:

[0076] Table 1

[0077]

[0078]

[0079] After the comparison is completed, these values ​​are arranged together to form a vector, which is the extracted author name-related feature.

[0080] 2.2 Compare the author affiliation information of the paper to be assigned with all papers in the candidate set. The values ​​are shown in Table 2 below:

[0081] Table 2

[0082]

[0083] After the comparison is completed, these values ​​are arranged together to form a vector, which is the extracted author institution-related feature.

[0084] 2.3 Compare the title information of the paper to be assigned with all papers in the candidate set. The values ​​are shown in Table 3 below:

[0085] Table 3

[0086]

[0087] After the comparison is completed, these values ​​are arranged together to form a vector three, which is the non-semantic feature of the extracted paper title.

[0088] 2.4. Use the XLNet pre-trained model to extract semantic features from paper fields containing semantic information (such as title, abstract, etc.);

[0089] 2.5. Combine all the vectors and semantic features obtained in 2.1, 2.2, 2.3, and 2.4 together to obtain the final feature vector.

[0090] S3, input the feature vector obtained in S2 into XGBoost to obtain the probability that the paper to be assigned belongs to the candidate set, and determine whether to assign the paper to be assigned to the candidate set based on this.

[0091] Specifically, the following steps are included:

[0092] 3.1 Input the features obtained in S2 above into the XGBoost decision tree to determine whether the paper to be assigned should be assigned to the corresponding author or not;

[0093] Assume that there are i papers in the candidate set corresponding to the current paper to be assigned. Then, i feature vectors will be obtained through the steps described in 2. Input these feature vectors into XGBoost, and i corresponding scores (0 < score < 1) can be obtained. Denote the candidate set with the highest probability as the most likely candidate set, and the highest score as score_max.

[0094] 3.2. For relatively simple cases, use a threshold for judgment. Generally, a threshold of 0.9 can be selected. If score_max >= 0.9, then assign the paper to be assigned to the candidate set corresponding to score_max; otherwise, denote the paper to be assigned as an unassigned paper. After operating on all papers to be assigned,汇总 all unassigned papers and denote this set as the unassigned paper set.

[0095] 3.3. Consider each unassigned paper as a cluster. Using the method in step 2, compare these clusters pairwise to obtain several feature vectors.

[0096] 3.4. Input the feature vectors obtained in step into XGBoost or the trained MLP for probability prediction. Consider the two clusters with the highest probability as the closest two clusters and merge them together. When the number of papers in the largest cluster is greater than 5, consider this as the main cluster and stop hierarchical clustering.

[0097] 3.5. Perform the incremental disambiguation operations mentioned in steps and

[0094] on the remaining papers in the unassigned candidate set to the main cluster obtained in step . Papers with failed assignment are returned to the unassigned candidate set, and papers with successful assignment are directly added to the main cluster. Use the finally incremented main cluster as a new candidate set. At this point, a complete round of incremental disambiguation operations is completed.

[0098] S4. Construct an incremental disambiguation post - processing framework based on agglomerative hierarchical clustering to perform post - processing operations on the papers that could not be assigned in incremental S3.

[0099] 4.1 Cold - start disambiguation feature extraction, as shown in Table 4 below:

[0100] Table 4 <s

[0101]

[0102]

[0103] For the paper author attribute, it is necessary to calculate the number c of authors with the same name ca , and the proportion r of the number of authors with the same name among all the authors of the two papers caFurthermore, due to the reduction in computational attributes, weights can be assigned to the calculated values ​​to enhance or weaken their impact on the final similarity score. This allows us to obtain the feature f representing the number of authors with the same name in a paper. ca =w ca c ca Characteristics of the proportion of authors with the same name as the paper Where w ca To influence the weight of the number of authors with the same name in a paper, To influence the weighting of the proportion of authors with the same name in a paper, the publisher attribute of the paper is subjected to the word segmentation and merging operation mentioned in the incremental disambiguation framework in step 2. Then, the Jaccard similarity coefficient between the two sets after the word segmentation and merging process is calculated, and the result is denoted as J. v For the attribute of author affiliation, the processing is largely the same as that in the incremental disambiguation framework in step 2. First, this attribute is segmented and merged. During this step, a stop word operation is added to ignore common words like "University" that do not affect the result. After this, the number of identical words, c, is calculated. co And the proportion of identical words r co Finally, the operation on the paper title attribute involves performing word segmentation and merging on this attribute, followed by calculating the number of identical words, c. ct And the proportion of identical words r ct And assign a weight to each of them to obtain f. ct =w ct c ct and Where w ct and These are the weights that affect the number of identical words in the title and the percentage of identical words, respectively.

[0104] 4.2 Cold Start Disambiguation Algorithm

[0105] We now have an unassigned set of papers P' = {p1, p2, ..., p...} N' After the above processing, we can obtain a set of vectors V = {v1, v2, ..., v}. N' The proposed cold-start disambiguation framework then utilizes this set of vectors representing unassigned papers for hierarchical clustering. First, it calculates the similarity between two papers, then uses this similarity to perform agglomerative hierarchical clustering on the input paper set. When the number of papers in the main cluster is greater than or equal to a threshold γ, an incremental assignment operation is performed. This incremental assignment operation differs from the operation in the incremental disambiguation framework, primarily in the use of paper features. After completing one hierarchical clustering and incremental assignment, it attempts to continue the loop operation until the clustering operation is complete, i.e., the loop executes until no further clustering operation can be performed. Details are shown in Table 5 below:

[0106] Table 5

[0107]

[0108]

[0109] Example

[0110] This embodiment of the incremental paper same-author disambiguation method based on pre-trained models and decision trees combines the AMiner dataset with DBLP to construct a new dataset to verify the method of this invention. The final experimental results prove the feasibility of the incremental disambiguation algorithm proposed in this invention.

[0111] The programming language used in this embodiment is Python. Python has a large number of open-source algorithm libraries, which makes it unnecessary to write all the related programs from scratch. In addition, there are commonly used deep learning frameworks such as PyTorch and TensorFlow.

[0112] Hardware: Intel Core i5-9600K processor, 16GB RAM, GeForce GTX 2080Ti graphics card; Software: Windows 10 operating system, Python 3.7.4, PyTorch-GPU programming language.

[0113] 1. Construct the dataset:

[0114] We will use the incremental disambiguation datasets provided by AMiner in recent years and the classic DBLP disambiguation dataset. For the incremental disambiguation task, the datasets provided by AMiner have already processed the overall data, including operations such as splitting the training and test sets.

[0115] (1) The AMiner dataset is compiled and made available for download by the AMiner platform. There are three versions of this dataset, which AMiner refers to as V1, V2, and V3. V1 and V2 are two manually labeled datasets with relatively accurate results, but due to the large workload of manual labeling, these two datasets are relatively small. Compared to the previous two datasets, V3 is a larger dataset, but it contains some noise. The AMiner dataset is divided into training and testing sets in a ratio of 8:2. The fields in this dataset are shown in Table 6.

[0116] Table 6 Fields of the AMiner dataset

[0117]

[0118]

[0119] (2) Unlike the AMiner dataset, the DBLP dataset does not undergo any special processing and requires manual processing. DBLP is an abbreviation for Digital Bibliography & Library Project, an author-centric integrated database system for English literature in the field of computer science. Unlike the AMiner dataset, DBLP does not split the dataset into training and test sets, but it contains many fields, as shown in Table 7.

[0120] Table 7 DBLP Dataset Fields

[0121]

[0122] First, the dataset needs to be preprocessed, which requires analyzing various aspects of the dataset, including the number of papers in each candidate set and the number of candidate sets for each author.

[0123] like Figure 6 , Figure 7 As shown, in the AMiner training set, there are 28,220 candidate sets with fewer than 5 papers, accounting for 84.54% of the total. In the test set, there are 5,333 candidate sets with fewer than 5 papers, accounting for 83.34% of the total. In this case, the candidate set coarse-ranking module in the incremental disambiguation framework may not be usable because the data limitations may lead to the module rejecting too many candidate sets, affecting the results. Although a similar situation occurs in the DBLP dataset, it is essentially the same as the problem in the AMiner dataset, so it will not be elaborated further.

[0124] Although the AMiner dataset has already divided the data into training and test sets in an 8:2 ratio, it only contains candidate sets corresponding to author names and the papers corresponding to those candidate sets. This alone is insufficient for incremental disambiguation methods. It's impossible to load all the paper data into memory for processing; therefore, the paper IDs and their various attributes are stored in a MongoDB database, and the relevant attributes are retrieved based on the paper IDs. In addition, separate training and test sets need to be constructed. For the construction of the training set, taking one author name as an example, all its candidate sets are denoted as a set C = {c1, c2, ..., c...}. m}, where m is the number of candidate sets, and all papers contained in each candidate set are denoted as a set. Where c kThis represents the total number of papers in the k-th candidate set. The specific construction method for a candidate set for a given author is shown in Table 8. It is worth noting that author names with only one candidate set are ignored, as no candidate sets with only one paper were found during data analysis, so no additional processing was performed for this issue. For the test set, the data processing is similar to that for the training set, with only two differences: one is recording which is the correct candidate set, and the other is that the value list of the result dictionary should contain the set of all papers corresponding to all candidate sets under that author name.

[0125] Table 8 shows the training set construction process for a single candidate set.

[0126]

[0127]

[0128] In addition to processing individual candidate sets, it's necessary to construct our own training and test sets. First, the complete training set is constructed by dividing DBLP into training and test sets in an 8:2 ratio based on the number of authors. Then, DBLP and the AMiner training set are combined to form a new training set, and the processing shown in Table 3 is performed. Next, DBLP and the AMiner test set are combined to form a new test set, which is then reconstructed according to the format of the AMiner test set, denoted as the complete test set. It's worth noting that for authors with the same name that might appear in DBLP and AMiner, one is randomly selected instead of merging them together to prevent introducing errors. Secondly, based on this, a new test set is created. This test set is constructed by deleting candidate sets with fewer than 5 papers from the complete test set. This test set is denoted as the filtered test set, and it is a more perfect test set compared to the complete test set.

[0129] 2. Baseline Method

[0130] This embodiment adopts the baseline methods of incremental disambiguation methods proposed in recent years, namely the nonparametric Bayesian framework proposed by CONNA and Zhang et al., and the probabilistic model proposed by Zhao et al. Since the authors of the latter two frameworks did not provide corresponding English abbreviations, this embodiment refers to these two frameworks as the Zhang nonparametric Bayesian framework and the Zhao probabilistic model, respectively. A brief introduction to the relevant models is as follows.

[0131] (1) CONNA is an incremental disambiguation framework proposed by Chen et al. CONNA trains the matching component and the decision component jointly through reinforcement learning, and the framework has been deployed on a large academic search system AMiner.

[0132] (2) Zhang’s nonparametric Bayesian framework uses DPGMM (Dirichlet Process Gaussian Mixture Model) as the core of the same-name disambiguation task and uses a SISR (Sequential Importance Sampling with Resampling) technique to reason about online classification and the discovery of new categories.

[0133] (3) The Zhao probability model uses a rich set of metadata and reduces the number of comparisons required for new articles. It also uses incremental classification to disambiguate names. In addition, the model uses fewer parameters than other algorithms.

[0134] 3. Evaluation Criteria

[0135] Since incremental disambiguation of authors with the same name is essentially a multi-classification problem, this embodiment selects the three most commonly used evaluation metrics as evaluation metrics for the incremental disambiguation results, namely precision, recall and F1 score, which are calculated as shown in formula (1), formula (2) and formula (3) respectively.

[0136]

[0137]

[0138]

[0139] In the formula, TP represents the number of correctly predicted answers, FP represents the number of results that incorrectly classify other classes as this class, and FN represents the number of results that predict the label of this class as another class label.

[0140] 4. Comparative Experiment Analysis in the Incremental Disambiguation Matching Module

[0141] For the extraction of paper titles and abstracts, this embodiment uses five models as baseline methods: BERT, GPT, GPT-2, Transformer-XL, and RoBERTa.

[0142] Table 9 shows the comparison results of the complete test set.

[0143]

[0144] As shown in Table 9, on the full test set, XLNet achieved the highest accuracy of 0.9035, recall of 0.6843, and F1 score of 0.7788. Specifically, its accuracy improved by 4.74% compared to the second-ranked GPT-2 model (0.8626), its recall improved by 1.89% compared to the second-ranked BERT model (0.6554), and its F1 score improved by 4.95% compared to the second-ranked GPT-2 model (0.7421). This clearly demonstrates the excellent performance of the XLNet model on unprocessed and somewhat messy datasets.

[0145] Table 10. Comparison Experiment Results of Screening Test Set

[0146]

[0147] As shown in Table 10, on the filtered test set, XLNet achieved the highest precision of 0.9491, and its recall of 0.9623 was only 0.0011 lower than BERT, which had the highest recall. Its final F1-score of 0.9556 was also the highest. Because the dataset was filtered, although XLNet's final F1-score was the highest, it was only 0.46% higher than BERT and GPT. This suggests that on datasets with sufficiently high overall quality, the difference in final results caused by using different models is not as significant as one might imagine; in fact, these models could be considered interchangeable.

[0148] On the two datasets mentioned above, the impact of different datasets on the experimental results is clearly visible. Although a superior dataset can narrow the gap between two different models, it generally does not change the overall superiority of one model over the other. XLNet achieves better results because of its novel permutation language model, the absence of masking, and the relative position encoding and fragment repetition mechanism introduced from Transformer-XL. These four features give XLNet essentially all the advantages of the control model in this embodiment. Regarding time consumption, this embodiment mainly calculates the entire feature extraction process, not just the individual time consumption of each model. The difference between calculating the individual time consumption of a model and the overall time consumption is theoretically a fixed value because the rest of the process, except for the model, remains unchanged. The time consumption of each model is as follows: Figure 8 As shown.

[0149] like Figure 8As shown, the time consumption of each model is not significantly different. However, the BERT model exhibits two extreme performances on the selected test set and the full test set, which is likely due to hardware limitations such as server specifications. In summary, XLNet performs adequately on the selected test set, and its time consumption on the full test set is similar to other models, resulting in a generally good overall performance.

[0150] For the scoring part in the matching module, this embodiment uses two models, DNN and Gradient Boosting Decision Tree (GBDT), as baseline methods. The indirect relationship between the relevant models is as follows.

[0151] (1) DNN is an extension of the perceptron and can be understood as a neural network with many hidden layers. This invention uses multiple DNNs with different numbers of layers for comparative experiments, with the aim of comparing the different effects of different numbers of layers on experimental results and algorithm running time.

[0152] (2) GBDT is an iterative decision tree model, which consists of multiple decision trees. The final result is obtained by summing up the results of all decision trees.

[0153] In the implementation process, in addition to modifying the scoring part in the matching module, the two test sets constructed above will also be used: a complete test set and a filtered test set. The implementation results and corresponding analysis will be provided afterward. Regarding the selection of the DNN model, this embodiment will indicate the number of hidden layers it contains after its name. For example, a DNN with 5 hidden layers will be labeled as DNN-5 in this embodiment.

[0154] For the loss function of the DNN, this embodiment chooses the ternary loss function, the Adam optimizer, the epoch is set to 100, the batch size is set to 32, the learning rate is set to 0.1, and an early stopping mechanism is used to obtain the best performing model.

[0155] For GBDT parameter settings, the maximum tree depth is set to 12, the learning rate is set to 0.1, the maximum number of iterations is set to 100, the model is selected as a regression model, and the loss function is the Mean Squared Error (MSE).

[0156] For XGBoost parameter settings, the maximum tree depth is set to 12, the learning rate is set to 0.1, the maximum number of iterations is set to 100, the model is selected as a regression model, and the loss function is the mean squared error loss function.

[0157] The experimental results are shown in Tables 11 and 12.

[0158] Table 11 Comparison Experiment Results of the Complete Test Set

[0159]

[0160] As shown in Table 11, on the complete test set, the XGBoost model achieved the highest accuracy of 0.9035, a 5.24% improvement over the second-ranked GBDT. Its recall of 0.6843 was also the highest, a 9.44% improvement over GBDT. Consequently, its F1-score was also the highest at 0.7788, a 7.90% improvement over GBDT. For these DNN models, the accuracy, recall, and F1-score all increase progressively with the number of layers in the DNN model.

[0161] Table 12 Comparison Experiment Results of Screening Test Set

[0162]

[0163] As shown in Table 12, on the selected dataset, the XGBoost model achieved the highest accuracy of 0.9491, a 1.48% improvement over the second-ranked DNN-9. Its recall of 0.9623 was also the highest, a 0.28% improvement over DNN-9. Consequently, its F1-score was also the highest at 0.9556, a 0.88% improvement over DNN-9. While the performance of several DNN models on the selected test set was not as good as XGBoost, the difference was not as significant as on the full test set.

[0164] like Figure 9 As shown, the time consumption of each model does not include the feature extraction part, only the prediction part. With the rest of the components remaining the same, the XGBoost model has an approximately 7% improvement in time consumption compared to GBDT, but it does not offer any advantage over DNN in terms of time consumption alone. Combining the results in Tables 11 and 12, DNN-9 can be used to replace XGBoost, provided the base data is of sufficient quality. It is worth noting that the DNN-9 mentioned here does not specifically refer to a DNN model with 9 hidden layers, but rather a model with similar characteristics to the DNN-9 used in the experiments—that is, a model with a relatively simple structure that yields good results.

[0165] 5. Analysis of comparative experimental results of incremental disambiguation

[0166] The main objective was to conduct comparative experimental analysis of the incremental disambiguation framework. Therefore, only the incremental disambiguation framework was used in the experiments, without combining it with the incremental post-processing disambiguation framework. Consequently, only a selected test set was used for the experiments. The comparative experimental results are shown in Table 13.

[0167] Table 13 Comparative Experiment Results of Incremental Disambiguation Framework

[0168]

[0169] As shown in Table 13, CONNA has the highest accuracy of 0.9675, while the model proposed in this invention has an accuracy of 0.9491. CONNA's accuracy is 1.94% higher than the model proposed in this invention. Zhao's proposed model has the highest recall of 0.9703, while the model proposed in this invention has an recall of 0.9623. Zhao's proposed model's recall is 0.83% higher than the model proposed in this invention. The model proposed in this invention has the highest F1-score of 0.9556, which is 0.19% higher than the second highest (0.9538). In summary, although the accuracy and recall of the model proposed in this invention are not the highest, its overall F1-score is the highest, meaning that the overall performance of the model proposed in this invention is still superior to the other three control models.

[0170] In addition to conducting experiments on the entire screening test set, this invention also selected several author names for individual processing to examine the F1-scores obtained by different methods for specific author names. The comparative experimental results are shown in Table 14.

[0171] Table 14 shows the F1-scores for different names using various methods.

[0172]

[0173] As shown in Table 14, different methods yielded similar results for different author names in the screening test set. This indicates that the differences between the four different methods on the screening test set are not significant. In this case, more attention should be paid to time consumption. The time consumption results are as follows: Figure 10 As shown.

[0174] From Table 13, Table 14 and Figure 10 The results show that although the model of this invention performs well in screening the test set, it lags far behind the other three models in terms of time consumption.

[0175] 6. Analysis of comparative experimental results of the overall framework

[0176] The main objective was to conduct comparative experimental analysis of the overall framework. Therefore, the experiments used a combined incremental and cold-start disambiguation framework to compare with other methods. Correspondingly, the complete test set was used for testing. The comparative experimental results are shown in Table 15.

[0177] Table 15. Results of the comparative experiment on the disambiguation framework.

[0178]

[0179] As shown in Table 15, CONNA has the highest accuracy of 0.9341, while the model of this invention has an accuracy of 0.9035. CONNA's accuracy is 3.39% higher than the model of this invention. The model of this invention has the highest recall of 0.6843, which is 4.73% higher than the second highest (0.6534). The model of this invention has the highest F1-score of 0.7788, which is 3.39% higher than the second highest (0.7533). In summary, although the accuracy of the model proposed in this invention is lower than that of CONNA, its recall and F1-score are the highest, indicating that the overall performance of the model proposed in this invention is better.

[0180] In addition to conducting experiments on the entire test set, this invention also selected several author names for individual processing to examine the F1-scores obtained by different methods for specific author names. The comparative experimental results are shown in Table 16.

[0181] Table 16 shows the F1-scores for different names using various methods.

[0182]

[0183] As shown in Table 16, there is a special author among the eight authors, Hongbin Liang. In the complete test set, there are many candidate sets with too few papers under this author, which will bring difficulties to incremental disambiguation. However, the performance of the model of this invention on this author is significantly better than that of the second author, with an F1-score that is 5.90% higher than that of the second author. This can, to some extent, indicate that the cold start disambiguation framework introduced by the model of this invention can make the disambiguation results better.

[0184] like Figure 11 As shown, the time consumed by the model of this invention is significantly longer than that of the other three comparative models.

[0185] From Table 15, Table 16 and Figure 11The results show that while the method of this invention has a slight advantage over the other three methods in disambiguating papers from authors with a smaller number of papers in the candidate set, its execution time is significantly longer than the other three comparative models, with the difference even exceeding the time consumption difference mentioned in the previous section. This is because the method of this invention performs cold-start disambiguation after incremental disambiguation. Although cold start can process these papers, it significantly increases the algorithm's time consumption.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An incremental method for disambiguating authors with the same name in academic papers based on a pre-trained model and a decision tree, characterized in that: It includes the following steps: S1. Construct a dataset and perform data preprocessing on author names, candidate set IDs, paper IDs, and paper information. S2. Extract the first features without semantic information fields by using the method of manually defined rules, extract the second features with semantic information fields by using the XLNet pre-trained model, and merge the first features and the second features together to obtain the required feature vectors. S3. Input the feature vectors obtained in S2 into XGBoost to obtain the probability that the paper to be assigned belongs to the candidate set, and judge whether to assign the paper to be assigned to the candidate set based on this. Specifically, it includes: 3.

1. Input the features obtained in step 2 into the XGBoost decision tree to determine whether the paper to be assigned is assigned to the corresponding author or not. Suppose there are i papers in the candidate set corresponding to the current paper to be assigned. Then, through step 2, i feature vectors will be obtained. Input these feature vectors into XGBoost, and corresponding i scores score can be obtained, where 0 < score < 1. Denote the candidate set with the highest probability as the most likely candidate set, and the highest score as score_max. 3.

2. For relatively simple cases, use a threshold for judgment. Generally, the threshold is selected as 0.

9. If score_max ≥ 0.9, then assign the paper to be assigned to the candidate set corresponding to score_max; otherwise, denote the paper to be assigned as an unassigned paper. After operating on all papers to be assigned, summarize all unassigned papers, and denote this set as the unassigned paper set. 3.

3. Regard each unassigned paper as a cluster. Use the method in step 2 to compare these clusters pairwise, and several feature vectors can be obtained. 3.

4. Input the feature vectors obtained in 3.3 into XGBoost or the trained MLP for probability prediction. Regard the two clusters with the highest probability as the closest two clusters, and merge these two clusters together. When the number of papers in the largest cluster is greater than 5, consider this as the main cluster, and stop hierarchical clustering. 3.

5. Perform the incremental disambiguation operation mentioned in 3.1 and 3.2 on the remaining papers in the unassigned candidate set to the main cluster obtained in 3.

4. Return the unassigned candidate set for those with failed assignment, and directly add those with successful assignment to the main cluster. Use the finally incremented main cluster as a new candidate set. At this time, a complete round of incremental disambiguation operation is completed. S4. Construct an incremental disambiguation post-processing framework based on agglomerative hierarchical clustering, and perform post-processing operations on the papers that cannot be assigned in the increment in S3.

2. The incremental paper same-name disambiguation method based on a pre-trained model and decision tree according to claim 1, characterized in that: The first features at least include author names and institutions, and the second features at least include paper titles and abstracts.

3. The incremental paper same-name disambiguation method based on a pre-trained model and decision tree according to claim 1, characterized in that: S2 specifically includes: 2.

1. Compare the author information between the paper to be assigned and all papers in the candidate set. The values are represented as follows: The number of authors with the same name. This value is incremented by 1 each time an author with the same name appears. The same operation is performed when an author's name appears repeatedly. The number of authors with the same name and from the same organization. This value is incremented by 1 each time an author with the same name and from the same organization appears. The same operation of incrementing the value will also occur when an author with the same name and from the same organization appears repeatedly. The ratio of the number of authors with the same name to the number of papers in the candidate set, i.e. ; The ratio of authors with the same name and from the same organization to the number of papers in the candidate set, i.e. ; The number of co-authors. This value is incremented by 1 each time a co-author appears. If an author appears repeatedly, this value is not affected. The number of co-authors and the number of authors in the paper to be assigned: The ratio, i.e. ; Co-authors in this candidate set The total number of times it appears in the text; The number of times co-authors appear in the candidate set Total number of Chinese authors The ratio, i.e. ; After completing the comparison, arrange these values together to form a vector one, and the vector one is the feature related to the author name extracted. 2.

2. Compare the author institution information between the paper to be assigned and all papers in the candidate set. The values are represented as follows: Papers to be assigned With candidate set The number of all identical institutions in the country; Number of identical institutions vs. total number of institutions The ratio between them, that is ; After word segmentation and merging, the papers to be assigned With candidate set The maximum value of the Jaccard similarity coefficient among all papers in the series, i.e. ; After word segmentation and merging, the papers to be assigned With candidate set The mean of the Jaccard similarity coefficients among all papers in the paper, i.e. ; After word segmentation and merging, the Jaccard similarity coefficient after processing with the Gaussian kernel function is one. A dimensional vector, where Indicates the number of input center points; After word segmentation and merging, the papers to be assigned With candidate set The maximum value of the ratio of the number of identical institutions among all papers to the total number of institutions, i.e. ; After word segmentation and merging, the papers to be assigned With candidate set The average of the ratio of the number of papers from the same institution to the total number of institutions. ; After word segmentation and merging, the ratio of the number of identical structures to the total number of structures after processing with a Gaussian kernel function is a... A dimensional vector, where Indicates the number of input center points; After the comparison is completed, these values ​​are arranged together to form a vector two, which is the extracted author institution-related feature; 2.3 Compare the title information of the paper to be assigned with all papers in the candidate set. The values ​​are represented as follows: After the word segmentation and merging process, identical words are in the unassigned domain. Number of times it appears in the text; After the word segmentation and merging process, identical words are in the candidate set. The number of times it appears in all papers; After the word segmentation and merging process, identical words in the paper to be assigned The number of times it appears in the paper to be assigned Total number of words in the text The ratio, i.e. ; After the word segmentation and merging process, identical words are in the candidate set. Number of occurrences in all papers and candidate set Total number of words in the text The ratio, i.e. ; After word segmentation and merging, the Jaccard similarity coefficient, processed by the Gaussian kernel function, is one. A dimensional vector, where Indicates the number of input center points; : Use XLNet to extract title features of papers, and then calculate the papers to be assigned. Titles and candidate sets The cosine similarity of the titles of each paper in the dataset was calculated, and then processed using a Gaussian kernel function to obtain the final result. ; After the comparison is completed, these values ​​are arranged together to form a vector three, which is the non-semantic feature of the extracted paper title; 2.

4. Semantic features are extracted from paper fields containing semantic information using the XLNet pre-trained model; 2.5 Combine the vectors 1, 2, and 3 obtained in 2.1, 2.2, and 2.3, as well as the semantic features extracted in 2.4, to obtain the final feature vector.

Images