Improved method and device for homonym disambiguation based on cross-source cross correction

Abstract

The application provides an improved same-name disambiguation method and device based on cross-source cross error correction, comprising: obtaining a cross-data-source data set; constructing an initial matching model, comparing the similarity of authors associated with a paper in an internal academic graph and authors associated with the paper in an external academic graph, dividing the cross-data-source data set to obtain a confidence set, an unconfidence set and a fuzzy set; obtaining first data in the fuzzy set, converting the hard label of the first data into a soft label, and generating a first target loss function according to the soft label; obtaining second data in the confidence set and third data in the unconfidence set, and performing cross error correction on the second and third data by using a preset cross error correction method to generate a second target loss function; training the initial matching model according to the first and second target loss functions to obtain an enhanced target matching model, and correcting a same-name disambiguation result by using the enhanced target matching model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer, network information technology. BACKGROUND

[0002] The purpose of name disambiguation is to distinguish different people with the same name, and name disambiguation is widely used in many digital libraries, such as Google Scholar, Microsoft Academic, DBLP, etc. More generally, there are many applications in knowledge graphs (such as Wikipedia, Baidu Encyclopedia, etc.), semantic search (such as Baidu, Bing) and the like. Existing work usually focuses on cold start name disambiguation [Zhang, 2018], and the goal of these works is to divide papers containing the same name into multiple disjoint clusters. On the other hand, the work focuses on incremental name disambiguation [Chen, 2020], and the goal is to assign newly published papers to existing authors online. However, existing work has paid little attention to correcting existing name disambiguation results, especially by using external knowledge to correct existing paper-author matching relationships.

[0003] To this end, the present patent proposes a new task of correcting name disambiguation results in academic graphs in a cross-data-source manner. In order to correct the existing paper-author matching relationship, the intuitive idea is to first measure the similarity between the paper and the author, and then find the inconsistent paper-author matching relationship by comparing the name disambiguation results of different data sources, and then infer the incorrect paper-author matching relationship. However, this idea faces two challenges: 1. Since the existing paper-author matching relationship may contain noise, a matching model that can achieve good matching effect and resist noise needs to be designed. However, existing matching models are difficult to meet both requirements. 2. For cross-data-source inconsistent paper-author matching relationships, the difficulty lies in determining which data source's paper-author matching relationship is incorrect. Research on the second challenge is still in the blank. SUMMARY

[0004] The present application aims to at least partially solve one of the technical problems in the related art.

[0005] To this end, the first object of the present application is to propose a name disambiguation improvement method based on cross-source cross-correction, for correcting name disambiguation results in academic graphs in a cross-data-source manner.

[0006] The second object of the present application is to propose a name disambiguation improvement device based on cross-source cross-correction.

[0007] To achieve the above object, the first aspect of the present application provides an improved method for disambiguating homonyms based on cross-source cross-correction, comprising: obtaining a cross-data-source dataset, the cross-data-source dataset comprising a plurality of triples, the triples comprising a paper, an author associated with the paper in an internal academic graph, and an author associated with the paper in an external academic graph; constructing an initial matching model, the initial matching model being used to obtain a similarity between the paper and the author; comparing the similarity of the author associated with the paper in the internal academic graph and the author associated with the paper in the external academic graph, and dividing the cross-data-source dataset according to the similarity to obtain a confident set, an unconfident set, and a fuzzy set; obtaining first data in the fuzzy set, converting a hard label of the first data into a soft label by using a linear combination between an original label and a model prediction, and generating a first target loss function according to the soft label; obtaining second data in the confident set and third data in the unconfident set, and performing cross-correction on the second data and the third data by using a preset cross-correction method to generate a second target loss function; and training the initial matching model according to the first target loss function and the second target loss function to obtain an enhanced target matching model, so as to correct a homonym disambiguation result by using the enhanced target matching model.

[0008] The improved method for disambiguating homonyms based on cross-source cross-correction provided by the embodiments of the present application comprises three stages. Stage 1: matching model design. In order to design an effective paper-author matching model, the present application proposes a matching method based on "kernel convolution", which combines the advantages of convolutional neural network and kernel aggregation method. Stage 2: data enhancement by external knowledge. By calculating the similarity of authors from different sources using external knowledge, the confident and unconfident paper-author matching relationships can be identified. Stage 3: cross-source correction. For unconfident paper-author matching relationships, the present application proposes two cross-correction methods to infer which source's paper-author relationship is more likely to be wrong. These two cross-correction methods can be used to enhance the matching model.

[0009] In addition, the improved method for disambiguating homonyms based on cross-source cross-correction according to the above embodiments of the present application can have the following additional technical features:

[0010] Further, in an embodiment of the present application, constructing an initial matching model comprises:

[0011] Constructing a similarity matrix S k , where each element represents the cosine similarity score between the vector representation of the word t i of the paper p and the word k of the paper p , is the vector representation of the word t ;

[0012] Based on the similarity matrix S k , the kernel convolution operation is defined as follows:

[0013]

[0014] where μ c and are the mean and variance of the cth RBF kernel, n is the size of the convolution filter, is the extracted n-gram similarity feature, x / x' represents the n-gram index value of paper p / p k ; where, The corresponding feature matrix is represented as where each element is C is the number of RBF kernels;

[0015] By changing the format of Z k by retaining the first dimension and flattening the other dimensions, the features Z k of all pairs of papers are concatenated to obtain h x , defined as follows:

[0016] h x = [reshape(Z 1 ) x ||...||reshape(Z r ) x ],

[0017] where h x represents the similarity between the xth n-gram of paper p and all papers of author a in ; an attention mechanism is used to learn the attention weight of the n-gram of paper p, defined as follows:

[0018]

[0019] h =∑ x α x h x ,

[0020] where w a and b a are trainable parameters, a x is the attention weight of the xth n-gram of paper p, h is the weighted hidden layer representation, containing the semantic information of paper-author matching;

[0021] In combination with artificial features as a supplement, including author, keyword, title similarity scores, hidden layer representation h and artificial features f man are weighted and combined, then passed to the fully connected layer to generate the matching score, defined as follows:

[0022]

[0023] o in = FC(δh + (1-δ)h),

[0024] where h man is the hidden layer vector of artificial features, ACT(·) is the activation function, w man and b man are trainable parameters, o in represents the output matching probability, FC(·) represents the full connection layer, and δ is a trainable parameter for balancing the importance between the hidden layer vector h and the artificial features. Further, in an embodiment of the present application, the cross-data-source dataset is divided according to the similarity, to obtain a confident set, an unconfident set and a fuzzy set, and the method comprises the following steps:

[0025] Given the cross-data-source dataset where p represents a paper, P is a set of disambiguated papers, a in / a out are authors associated with the paper p in A in / A out , then the confident set the unconfident set and the fuzzy set are defined as follows:

[0026]

[0027]

[0028]

[0029] where SIM(a in ,a out ) represents the similarity of authors a in and a out .

[0030] Further, in an embodiment of the present application, the cross error correction method based on comparison comprises:

[0031] The initial matching model is used to predict the correct probability of the two sides of the data source (p, a in ) and (p, a oyt ); wherein the prediction output is defined as:

[0032]

[0033] The objective function is the cross-entropy loss function, defined as:

[0034]

[0035] wherein, y i = 1 means that the sample belongs to the confident set, y i = 0 means that the sample belongs to the unconfident set.

[0036] Further, in an embodiment of the present application, the PSL-based cross-correction method comprises:

[0037] modeling the triple (p, a in , a out ) into a logical rule, and then judging the matching degree of (p, a in ); wherein the loss function of the PSL-based cross-correction method is defined as:

[0038]

[0039] wherein, y i = 1 means that the sample belongs to the confident set, y i = 0 means that the sample belongs to the unconfident set, N e represents the total number of samples in the confident set and the unconfident set, respectively, are the distance metrics defined in the confident set / unconfident set.

[0040] Further, in an embodiment of the present application, the initial matching model is trained according to the first and second target loss functions to obtain an enhanced target matching model, comprising:

[0041] the label self-correction defines a target function the cross-correction defines a target function the target functions and are weighted as the overall target function, the overall loss function is defined as:

[0042]

[0043] wherein, λ is a weight parameter balancing the two types of loss functions.

[0044] Further, in an embodiment of the present application, given an internal academic graph G in = {E in , R in} and an external academic graph G out = {E out , R out}, wherein E in / E out is an entity set, Rin / R out is a set of relations, and the paper entities of the internal and external academic graphs are disambiguated.

[0045] To achieve the above object, the second aspect of the present application provides an improved same-name disambiguation device based on cross-source cross-correction, characterized in that it comprises: an input module configured to obtain a cross-data-source dataset, wherein the cross-data-source dataset comprises a plurality of triples, and each triple comprises a paper, an author associated with the paper in an internal academic graph, and an author associated with the paper in an external academic graph; a design module configured to construct an initial matching model, wherein the initial matching model is configured to obtain a similarity between the paper and the author; a classification module configured to compare the similarities of the authors associated with the paper in the internal academic graph and the authors associated with the paper in the external academic graph, divide the cross-data-source dataset according to the similarities, and obtain a confident set, an unconfident set, and a fuzzy set; a label self-correction module configured to obtain first data in the fuzzy set, convert a hard label of the first data into a soft label by using a linear combination between an original label and a model prediction, and generate a first target loss function according to the soft label; a cross-correction module configured to obtain second data in the confident set and third data in the unconfident set, perform cross-correction on the second data and the third data by using a preset cross-correction method, and generate a second target loss function; and a disambiguation module configured to train the initial matching model according to the first target loss function and the second target loss function, obtain an enhanced target matching model, and correct a same-name disambiguation result by using the enhanced target matching model.

[0046] To achieve the above object, the third aspect of the present application provides a computer device, characterized in that it comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods in the improved same-name disambiguation method based on cross-source cross-correction when executing the computer program.

[0047] To achieve the above object, the fourth aspect of the present application provides a non-transitory computer readable storage medium, characterized in that it stores a computer program, and the computer program is executable on a processor to implement any of the methods in the improved same-name disambiguation method based on cross-source cross-correction. BRIEF DESCRIPTION OF DRAWINGS

[0048] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which:

[0049] Figure 1A flowchart of an improved same-name disambiguation method based on cross-source cross-correction is provided for an embodiment of the present application.

[0050] Figure 2 A flowchart of an improved same-name disambiguation device based on cross-source cross-correction is provided for an embodiment of the present application.

[0051] Figure 3 A flowchart of an improved same-name disambiguation method based on cross-source cross-correction is provided for an embodiment of the present application.

[0052] Figure 4 A cross-data-source same-name disambiguation example diagram is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0053] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0054] The improved same-name disambiguation method and device based on cross-source cross-correction of an embodiment of the present application are described below with reference to the accompanying drawings.

[0055] Figure 1 A flowchart of an improved same-name disambiguation method based on cross-source cross-correction is provided for an embodiment of the present application.

[0056] As Figure 1 shown, the improved same-name disambiguation method based on cross-source cross-correction includes the following steps:

[0057] S1: Obtain a cross-data-source dataset, the cross-data-source dataset includes a plurality of triples, and the triples include a paper, an author associated with the paper in an internal academic graph, and an author associated with the paper in an external academic graph;

[0058] wherein, given an internal academic graph G in ={E in ,R in} and another external academic graph G out ={E out ,R out}, here E in / E out is an entity set, and R in / R outis a set of relations. For both graphs, we only consider the author and paper entities. Furthermore, we assume that the paper entities in both graphs are well disambiguated, as most papers can be identified by their title and author list (i.e., they have low ambiguity). Furthermore, the relationships we primarily consider are those matching between authors and papers.

[0059] The goal of this paper is to find errors in the existing paper-author matching relationship. Here p represents the paper, P is the set of papers after disambiguation, a in / a out The paper p is in A in / A out More specifically, each paper has multiple attributes, such as title and publication location. We combine the contents of all attributes of the paper and represent each paper as p = [t1, t2, ..., t l ], here t i represents the i-th word in paper p, and l represents the maximum number of words selected for each paper. If each author is regarded as a list of papers, author a in It can be expressed as a in =[p1,p2,…p r ]. In addition, each of these papers Our goal is to learn a matching function To judge the paper-author pair (p,a in )’s correctness.

[0060] The basic process of the algorithm framework of this patent is as follows Figure 3 shown.

[0061] S2: Build an initial matching model, which is used to obtain the similarity between papers and authors;

[0062] The paper author matching model is the basic component of this application method. Because any author can be regarded as a list of papers, the core problem of paper-author matching can be transformed into paper-paper matching. k The matching between them is actually a text matching problem.

[0063] Because CNNs can achieve parameter sharing with a small number of parameters, CNN-based methods are generally considered more resistant to noise than Transformer-based methods. However, CNN convolution is not designed for similarity matching. On the other hand, the state-of-the-art CONNA model uses an expressive kernel aggregation module to extract multi-level matching features, but it extracts coarse-grained "word-paper" matching patterns.

[0064] In order to design a matching model with strong matching ability and capable of resisting noise data, the application designs a lightweight matching model CroND-base, which can simultaneously have the advantages of the current advanced CNN-based matching model and the kernel aggregation-based matching model.

[0065] Further, in an embodiment of the application, the initial matching model is constructed, comprising:

[0066] A similarity matrix S is constructed k , wherein each element represents the cosine similarity score between the word t i of the paper p and the word k of the paper p , and is the vector representation of the word ;

[0067] A similarity matrix S is constructed k , wherein each element represents the cosine similarity score between the word t i of the paper p and the word k of the paper p , and is the vector representation of the word ;

[0068] Based on the similarity matrix S k , the kernel convolution operation is defined as follows:

[0069]

[0070] wherein μ c and are the mean and variance of the cth RBF kernel, n is the size of the convolution filter, is the extracted n-gram similarity feature, and x / x' represents the n-gram index value of the paper p / p k ; wherein, The corresponding feature matrix is represented as wherein each element is C is the number of RBF kernels;

[0071] The format of Z k is changed by retaining the first dimension and flattening the other dimensions, and the features Z k of all paper pairs are connected to obtain h x , which is defined as follows:

[0072] h x = [reshape(Z 1 )x ||...||reshape(Z r ) x ],

[0073] Among them, h x Indicates the xth n-gram of paper p and author a in The similarity between all papers in p is calculated by the attention mechanism, which is defined as follows:

[0074]

[0075] h=∑ x α x h x ,

[0076] Among them, w a and b a is a trainable parameter, a x is the attention weight of the x-th n-gram of paper p, and h is the weighted hidden layer representation, which contains the semantic information of the paper author matching;

[0077] Combined with artificial features as a supplement, including similarity scores of authors, keywords, and titles, hidden layer representation h and artificial features f man are weightedly combined and then passed to a fully connected layer to generate a matching score, defined as follows:

[0078]

[0079] o in =FC(δh+(1-δ)h),

[0080] Among them, h man is the hidden vector of the artificial feature, ACT(·) is the activation function, w man and b man is a trainable parameter, o in represents the output matching probability, FC(·) represents the fully connected layer, and δ is a trainable parameter used to balance the importance between the hidden vector h and the hand-crafted features.

[0081] During training, we use existing paper author pairs (p,a in ) as a positive pair and in the internal diagram with a in Other authors of the same name obtained samples make As negative examples, we adopt cross entropy loss as the objective function.

[0082] S3: comparing the similarity of the authors associated with the paper in the internal academic graph and the authors associated with the paper in the external academic graph, dividing the cross-data-source dataset according to the similarity to obtain a confident set, an unconfident set and a fuzzy set;

[0083] Wherein, by comparing the similarity of the authors across data sources, the confident and unconfident paper-author matching relationships in the internal academic graph can be identified. The training set can be divided into three subsets, the confident set The unconfident set And the fuzzy set The confident set Among them, the similarity of authors a in and a out is high; the unconfident set Among them, the similarity of authors a in and a out is low; the fuzzy set Among them, the similarity of authors a in and a out is between the above two.

[0084] Specifically, the existing paper author pair may be problematic as positive examples, because due to the name disambiguation problem, some papers may be assigned to the incorrect author. In this patent, by using external knowledge, we can distinguish between confident and unconfident paper-author matching relationships from the existing paper-author matching relationships.

[0085] We consider the triple (p, a in , a out ), and further consider the similarity of authors a in and a out , denoted as SIM(a in , a out ). SIM(a in , a out ) is high, which means that the matching relationship of the relationship (p, a in ) is recognized by external knowledge, while SIM(a in , a out ) is low, which indicates that at least one of (p, a in ) and (p, a out ) is incorrect.

[0086] SIM(a in , a out ) can be measured in a variety of ways. In order to more accurately estimate the similarity of authors, we define some statistical features and use strict rules to discover similar and different authors. The statistical features include the overlap rate of two authors' papers / collaborators / publication locations, etc.

[0087] Further, in an embodiment of the present application, the cross-data-source dataset is divided according to the similarity to obtain a confident set, an unconfident set and a fuzzy set, and the method comprises the following steps:

[0088] Given a cross-data-source dataset Where p represents a paper, P is a set of disambiguated papers, a in / a out is an author associated with paper p, then the confident set in / A out The unconfident set and the fuzzy set are defined as follows:

[0089]

[0090]

[0091]

[0092] Where SIM(a in ,a out ) represents the similarity between authors a in and a out .

[0093] S4: Obtain the first data in the fuzzy set, convert the hard label of the first data into a soft label by using a linear combination between the original label and the model prediction, and generate a first target loss function according to the soft label;

[0094] Where, for the paper-author pairs in the fuzzy set We believe that their labels may be noisy. To solve this problem, we use a self-correction method based on label smoothing. Specifically, we convert the hard label into a soft label by using a linear combination between the original label and the model prediction.

[0095]

[0096]

[0097] Where, is the soft label, ∈ is a hyperparameter, KL represents the KL divergence, is the target function of label self-correction.

[0098] ​S5: Obtain the second data in the confident set and the third data in the unconfident set, and cross-correct the second data and the third data by using a preset cross-correcting method to generate a second target loss function; the calculation methods of the near-neighbor similarity and the structural similarity will be described in detail below.

[0099] Take two academic graphs, AMiner and Microsoft Academic Graph (MAG), as examples. In Figure 4 , Quanquan Gu, an author in UCLA, has written 3 papers in AMiner, however, the bottom paper is assigned to Quanquan Gu in Zhejiang University in another academic graph MAG. By observing the three papers on the right, it can be found that the upper two papers and the bottom paper are in different research fields, so they should belong to different authors. Our goal is to use external knowledge (such as MAG) to correct the paper-author assignment errors in the internal system (such as AMiner).

[0100] For the data in the unconfident set , it is necessary to infer which author-paper pair in the data source is incorrect. The present application proposes two cross-correction methods, including a comparison-based cross-correction and a PSL-based cross-correction.

[0101] For the data in the unconfident set , the idea is to compare the predictions of the model for (p, a in ) and (p, a out ), i.e. and where is the matching probability output by the CroND-base model. If , the (p, a in ) pair is likely to be incorrect. However, some studies have shown that neural networks tend to gradually overfit noisy labels. Therefore, we use the model from the early training for comparison, denoted as In addition, if the matching probability is lower than a certain threshold ξ, we regard (p, a in ) as a negative example.

[0102] Further, in an embodiment of the present application, the comparison-based cross-correction method includes:

[0103] Using the initial matching model to predict the probability of (p, a in ) and (p, a out ) being correct on both sides; wherein the prediction output is defined as:

[0104]

[0105] The objective function is the cross-entropy loss function defined as:

[0106]

[0107] where, y i = 1 indicates that the example belongs to the confident set, y i = 0 indicates that the example belongs to the unconfident set.

[0108] The above cross-correction method does not consider the matching similarity of (a in , a out ) by the author. Here, we model the triplets (p, a in , a out ) as logical rules, because whether any two entities match or not depends on the remaining entity in the triplet.

[0109] To this end, we introduce probabilistic soft logic (PSL) (Kimmig et al., 2012) to model the dependency of the matching probability of pairs in triples. PSL associates each rule with a probability value to represent the probability that the rule is true.

[0110]

[0111] where, represents the confidence score of rule u, and the value v u . In PSL, Lukasiewicz t-norm is used to define the basic logical operations.

[0112]

[0113]

[0114]

[0115] PSL considers a rule derivation to be satisfied when the probability value of the head is equal to or higher than the probability value of its body. For a rule γ: u1 ^ u2 → u3,

[0116]

[0117] When η γ ≤ 0, the rule γ is considered to be satisfied.

[0118] Specifically, our goal is to learn the matching probability of (p, a out ) and (a in , aout ) is deduced from the matching relationship of (p,a in ) matching relationship. For example, In this paper, we define v2=SIM(a in ,a out )and in and It has the same meaning as the comparison-based cross-error correction method. We calculate the data in

[0119]

[0120] Here we introduce As a flexible "spacing" hyperparameter, instead of taking a constant 1, The default value is 1.3.

[0121] In contrast, the untrusted set The situation is more complicated. We use Indicates whether u1 and u2 match, where Indicates true, Indicates false. We use To represent the matching probability of u1 and u2. Any triple in must satisfy one of the following rules:

[0122]

[0123]

[0124]

[0125] To jointly model these rules, we define the following distance:

[0126]

[0127] In the above formula, satisfying rules [1] and [3] requires η uf is too small, and satisfying rule [2] requires η uf Therefore, a small η uf The paper-author pairs (p,a in ) has the same goal. Here, we also use the interval parameter

[0128] Furthermore, the cross-error correction method based on PSL includes:

[0129] Triplet (p, a in , a out ) is modeled into a logical rule, and then the matching degree of (p, a in ) is determined; wherein the loss function of the cross error correction method based on PSL is defined as:

[0130]

[0131] Here, we optimize the small value of eta by the min operation, because the corresponding triplet is more likely to satisfy the corresponding rule, such as rules [1] and [3].

[0132] S6: Training the initial matching model according to the first target loss function and the second target loss function to obtain an enhanced target matching model, so as to correct the homonym disambiguation result by using the enhanced target matching model.

[0133] Wherein, the self error correction method and the cross error correction method are jointly trained, and are used for enhancing the paper-author matching model.

[0134] Further, in an embodiment of the present application, the initial matching model is trained according to the first target loss function and the second target loss function to obtain an enhanced target matching model, comprising:

[0135] The label self error correction defines a target function The cross error correction defines a target function The target functions and are weighted as the overall target function, The overall loss function is defined as:

[0136]

[0137] Wherein, lambda is a weight parameter for balancing the two types of loss functions. Refers to Or The two types of loss are jointly trained to update the parameters of the matching model and enhance the matching model.

[0138] In terms of actual data, a publicly available artificial annotation homonym disambiguation dataset WhoIsWho is used as real annotation data, MAG is used as external knowledge, and the old version of the publicly available AMiner data is used as original data containing errors. The matching model CroND-base in the present application has an AUC value of the prediction result 3.9% higher than that of the advanced CONNA model, and an MAP value 0.2% lower than that of the advanced CONNA model. The AUC value of the prediction result of the CroND-base model is 75.26%, and the MAP value is 50.23%. In the fuzzy set Using the self-correction method, the AUC is improved to 75.43%, and the MAP is improved to 50.67%. Combined with the self-correction method and the comparison-based cross-correction method, the AUC is improved to 75.80%, and the MAP is improved to 51.26%. Combined with the self-correction method and the PSL-based cross-correction method, the AUC is improved to 75.81%, and the MAP is improved to 52.10%. Therefore, the method of cross-correction adopted by the present application can improve the accuracy of name disambiguation.

[0139] The improved name disambiguation method based on cross-source cross-correction proposed in the embodiments of the present application corrects the name disambiguation results in the academic graph in a cross-data-source manner, and realizes a higher matching accuracy of paper authors than existing methods. The basic matching model CroND-base proposed in the present application has better fitting ability than the advanced CONNA model by modeling more fine-grained matching features, and has better generalization performance than the matching model based on Transformer due to the small amount of parameters of the CroND-base model. The cross-correction method proposed in the present application can identify the possible incorrect paper-author matching relationship by comparing the data across data sources, thereby avoiding manual correction and improving the matching accuracy of paper authors without using manual annotation.

[0140] In order to realize the above-mentioned embodiments, the present application further provides an improved name disambiguation device based on cross-source cross-correction.

[0141] Figure 2 A structural schematic diagram of an improved name disambiguation device based on cross-source cross-correction provided by the embodiments of the present application.

[0142] As Figure 2As shown, the same name disambiguation improvement device based on cross-source cross error correction includes an input module 10, a design module 20, a classification module 30, a label self-correction module 40, a cross error correction module 50, and a disambiguation module 60. The input module is used to obtain a cross-data-source data set, the cross-data-source data set includes a plurality of triples, and the triples include a paper, an author associated with the paper in an internal academic graph, and an author associated with the paper in an external academic graph. The design module is used to construct an initial matching model for obtaining the similarity between the paper and the author. The classification module is used to compare the similarity of the author associated with the paper in the internal academic graph and the author associated with the paper in the external academic graph, divide the cross-data-source data set according to the similarity, and obtain a confident set, an unconfident set, and a fuzzy set. The label self-correction module is used to obtain first data in the fuzzy set, convert the hard label of the first data into a soft label by using a linear combination between the original label and the model prediction, and generate a first target loss function according to the soft label. The cross error correction module is used to obtain second data in the confident set and third data in the unconfident set, and cross error correction is performed on the second data and the third data by using a preset cross error correction method to generate a second target loss function. The disambiguation module is used to train the initial matching model according to the first target loss function and the second target loss function to obtain an enhanced target matching model, and correct the same name disambiguation result by using the enhanced target matching model.

[0143] In order to realize the above-mentioned embodiments, the present application further provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to realize the method in any one of the same name disambiguation improvement method based on cross-source cross error correction.

[0144] In order to realize the above-mentioned embodiments, the present application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to realize the method in any one of the same name disambiguation improvement method based on cross-source cross error correction.

[0145] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

[0146] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0147] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. An improved method for disambiguating homonyms based on cross-source cross- correction, characterized in that, The method comprises the following steps: obtaining a cross-data-source dataset, the cross-data-source dataset comprising a plurality of triples, the triples comprising a paper, an author associated with the paper in an internal academic graph, and an author associated with the paper in an external academic graph; constructing an initial matching model for obtaining a similarity between the paper and the author; comparing a similarity of authors associated with the paper in an internal academic graph to authors associated with the paper in an external academic graph, partitioning the cross-data source dataset according to the similarity to obtain a confident set, an unconfident set, and a fuzzy set, wherein a given cross-data source dataset Where p represents the paper, P is the set of papers after disambiguation, and a in / a out The paper p is in A in / A out The associated author, then the confidence set Untrusted Set and fuzzy sets The definition is as follows: wherein SIM(a in , a out ) denotes the similarity of authors a in and a out ; obtaining first data in the fuzzy set, converting a hard label of the first data into a soft label by using a linear combination between an original label and a model prediction, and generating a first target loss function according to the soft label; obtaining second data in the confident set and third data in the unconfident set, and performing cross error correction on the second data and the third data by using a preset cross error correction method to generate a second target loss function; training the initial matching model according to the first target loss function and the second target loss function to obtain an enhanced target matching model, and correcting a same-name disambiguation result by using the enhanced target matching model.

2. The method of claim 1, wherein, The constructing an initial matching model comprises: Construct a similarity matrix S k , where each element word t representing paper p i and paper p k Word The cosine similarity score between the vector representations, It's a word Vector representation of ; Based on the similarity matrix S k The kernel convolution operation is defined as follows: where μ c and is the mean and variance of the cth RBF kernel, n is the size of the convolution filter, is the extracted n-gram similarity feature, x / x' represents the n-gram index value of the paper p / p k ; wherein, The corresponding feature matrix is represented as where each element is C is the number of RBF kernels; Change the format of Z by keeping the first dimension and flattening the others k Connect all pairs of features Z k to obtain h x defined as follows: h x = [reshape(Z 1 ) x ||...||reshape(Z r ) x ], where h x represents the similarity between the xth n-gram of paper p and all papers of author a in . learning an attention weight of an n-gram of the paper p by using an attention mechanism, and the attention weight is defined as follows: h = ∑ x α x h x , where w a and b a are trainable parameters, a x is the attention weight of the xth n-gram of paper p, and h is the weighted hidden layer representation containing semantic information of the paper-author match. In combination with the artificial features, the similarity scores of the authors, keywords, titles, the hidden layer representation h and the artificial features f man are weighted combined and then passed to a fully connected layer to generate the match score, defined as follows: o in = FC(5h + (1 - 5)h), where h man is the hidden layer vector of the artificial feature, ACT(·) is the activation function, w man and b man are trainable parameters, o in denotes the output matching probability, FC(·) denotes the fully connected layer, and δ is a trainable parameter for balancing the importance between the hidden layer vector h and the hand-crafted feature.

3. The method of claim 1, wherein the preset cross error correction method comprises a comparison-based cross error correction method, and the comparison-based cross error correction method comprises: using the initial matching model to predict the probability of the two-side data sources (p, a in ) and (p, a out ) being correct; wherein the prediction output is defined as: The target function is a cross-entropy loss function, which is defined as: wherein, y i = 1 indicates that the example belongs to the confident set, y i = 0 indicates that the example belongs to the unconfident set.

4. The method of claim 1, wherein the preset cross error correction method comprises a PSL-based cross error correction method, and the PSL-based cross error correction method comprises: Triples (p, a in , a out ) are modeled as logical rules, and then the matching degree of (p, a in ) is determined; wherein the loss function of the cross error correction method based on PSL is defined as: where y i = 1 indicates that the example belongs to the confident set, y i = 0 indicates that the example belongs to the unconfident set, N e denotes the total number of examples in the confident and unconfident sets, are the defined distance metrics in the confident / unconfident set, respectively.

5. The method of claim 1, wherein the training the initial matching model according to the first target loss function and the second target loss function to obtain an enhanced target matching model comprises: The label self-correction defines an objective function The cross-correction defines an objective function The objective function And The weighted overall objective function, The loss function of the body is defined as: wherein λ is a weight parameter for balancing the two types of loss functions.

6. The method of claim 1, wherein, comprises: Given an internal academic graph G in = {E in , R in} and an external academic graph G out = {E out , R out}, where E in / E out is a set of entities and R in / R out is a set of relations, and the paper entities of the internal and external academic graphs are disambiguated.

7. An improved device for homonym disambiguation based on cross-source cross-error correction, characterized in that: comprises: The input module is configured to obtain a cross-data-source dataset, the cross-data-source dataset comprising a plurality of triples, the triples comprising a paper, an author associated with the paper in an internal academic graph, and an author associated with the paper in an external academic graph; The design module is configured to construct an initial matching model for obtaining a similarity between the paper and the author. The classification module is configured to compare the similarity between the author associated with the paper in the internal academic graph and the author associated with the paper in the external academic graph, divide the cross-data-source dataset according to the similarity to obtain a confident set, an unconfident set and a fuzzy set, wherein the cross-data-source dataset is given where p represents a paper, P is the set of disambiguated papers, a in / a out are the authors associated with paper p, respectively, the confident set in / A out is the set of papers written by author a, the unconfident set is the set of papers written by author a, the ambiguous set is the set of papers written by author a, the ambiguous set is defined as follows: wherein SIM(a in , a out ) represents the similarity of authors a in and a out . The label self-correction module is configured to obtain first data in the fuzzy set, convert a hard label of the first data into a soft label by using a linear combination between an original label and a model prediction, and generate a first target loss function according to the soft label. The cross error correction module is configured to obtain second data in the confident set and third data in the unconfident set, and perform cross error correction on the second data and the third data by using a preset cross error correction method to generate a second target loss function. A disambiguation module is configured to train the initial matching model according to the first target loss function and the second target loss function to obtain an enhanced target matching model, so as to correct the homonym disambiguation result by using the enhanced target matching model.

8. A computer device, comprising: A computer program product comprising a memory, a processor and a computer program stored on the memory and loadable on the processor, the processor implementing the method according to any one of claims 1-6 when executing the computer program.

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

Images