Market dishonesty information knowledge graph entity identification method

Abstract

The invention relates to a market dishonesty information knowledge graph entity recognition method. The method comprises the steps of obtaining text data of market dishonesty information to be recognized; inputting the text data into a sequence splicing layer for multi-granularity processing to obtain vocabulary vectors of different structures, and splicing the vocabulary vectors of different structures to obtain a text vector; inputting the text vector into a feature extraction model based on an adaptive fractal exploration optimization algorithm, and outputting a feature extraction vector; and inputting the feature extraction vector into a conditional random field model based on strategy optimization, and outputting the entity and the category thereof in the text data. According to the method, the fine entities in the text can be identified more accurately, and the accuracy of entity identification is remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of knowledge graph entity recognition, and in particular to a method for identifying market dishonesty information knowledge graph entities. Background Art

[0002] As an important information organization and processing tool, knowledge graph has been widely used in various scenarios, including search engine optimization, intelligent question-answering systems, recommendation systems, etc. It can help machines better understand information and provide more accurate information services.

[0003] As a basic task for building knowledge graphs, entity recognition tasks still face many challenges: 1) Traditional entity recognition methods often face the problem of vocabulary loss, especially when dealing with complex text or multi-granularity data, it is difficult to effectively capture all key information; 2) Lacking effective data enhancement methods, their models perform poorly on novel or changing data sets and are susceptible to overfitting; 3) Many existing algorithms rely on traditional optimization techniques such as gradient descent. These methods are prone to fall into local optimality in non-convex optimization problems and have slow optimization speeds. Summary of the invention

[0004] Based on this, it is necessary to provide a market dishonest information knowledge graph entity recognition method based on multi-granularity features, which includes: S1: Obtain text data of market dishonesty information to be identified; S2: inputting the text data into a sequence concatenation layer for multi-granularity processing to obtain vocabulary vectors of different structures, and concatenating the vocabulary vectors of different structures to obtain a text vector; S3: inputting the text vector into a feature extraction model based on an adaptive fractal exploration optimization algorithm, and outputting a feature extraction vector; S4: Input the feature extraction vector into a conditional random field model based on strategy optimization, and output entities and their categories in the text data.

[0005] Preferably, the step of inputting the text data into a sequence concatenation layer for multi-granularity processing to obtain vocabulary vectors of different structures comprises: The sequence splicing layer includes four structures: B, M, E, and S; The B structure is used to match the text data with characters c i The first word; The M structure is used to match characters in the text data. c i The vocabulary that is in the middle; The E structure is used to match characters in the text data c i The words that are at the end; The S structure is used to represent characters c i itself; Different structures of vocabulary vector expressions include: ; ; ; ; in, Representing characters c i The vocabulary vector matched to the B structure; Representing characters c i The vocabulary vector matched to the M structure; Representing characters c i The vocabulary vector that matches the E structure; Representing characters c i The vocabulary vector matched to the S structure; c i Represents the i-th character; Represents the word whose i-th character comes before the k-th character in the text data; Indicates the word whose i-th character is between the j-th character and the k-th character in the text data; represents the vocabulary whose i-th character comes after the j-th character in the text data; L represents the dictionary; and n represents the text data sequence.

[0006] Preferably, the step of concatenating vocabulary vectors of different structures to obtain a text vector comprises: The vocabulary vectors of each structure are normalized and integrated. The calculation formula includes: ; ; in, represents normalization and integration; s represents any structure; w represents the vocabulary vector of any structure; S represents the vocabulary vector in the text data that conforms to the S structure; B represents the vocabulary vector in the text data that conforms to the B structure; M represents the vocabulary vector in the text data that conforms to the M structure; E represents the vocabulary vector in the text data that conforms to the E structure; Represents the word frequency of vocabulary vector w; represents the vocabulary embedding matrix of the vocabulary vector w; Z represents the sum of the word frequencies of the vocabulary vector w in the vocabulary vectors of the four structures; The vocabulary vectors of all structures after normalization and integration are concatenated to obtain the text vector. The calculation formula is: ; in, represents a text vector; Represents the vocabulary vector of the normalized and integrated B structure; Represents the vocabulary vector of the normalized and integrated M structure; Represents the vocabulary vector of the normalized and integrated E structure; Represents the vocabulary vector of the normalized and integrated S structure.

[0007] Preferably, the method further includes training the feature extraction model, wherein the training process includes: Step 1: Collect text data for training and preprocess it; Step 2: Input the preprocessed training text vector into the feature extraction model; Step 3: Select an initial parameter set of the feature extraction model and initialize the search granularity and search radius; the initial parameter set includes the weight and bias of the neural network; Step 4: Generate a set of self-similar exploration points based on the parameter point of the current iteration; all the exploration points are distributed around the parameter point of the current iteration according to the fractal rule based on the search granularity and the search radius; Step 5: Calculate the loss function value of all exploration points, calculate the information entropy of all exploration points based on the loss function value, and construct a search granularity adjustment function based on the information entropy; Step 6: dynamically adjusting the search granularity based on the search granularity adjustment function; Step 7: Select the exploration point with the smallest loss function value as the parameter point of the next iteration, and dynamically adjust the search radius according to the loss between the parameter point of the current iteration and the parameter point of the next iteration; Step 8: Repeat iterative steps 4-7 until the maximum number of iterations is reached to obtain a trained feature extraction model; input the training text vector into the trained feature extraction model based on the adaptive fractal exploration optimization algorithm, and output the training feature extraction vector.

[0008] Preferably, the pretreatment comprises: The text data used for training is passed through the sequence concatenation layer to obtain the text vector for training; The generative adversarial network based on topological phase evolution performs data expansion on the training text vector to obtain the training text vector.

[0009] Preferably, the training process of the generative adversarial network based on topological phase evolution includes: Step 1: Initialize the network parameters of the generator and discriminator; Step 2: Adjust the generator's generation strategy based on the real dataset and the generated dataset of the current iteration; Step 3: Use the adjusted generator to generate a batch of fake data from random noise; Step 4: Input the false data and the real data into the discriminator to obtain the discrimination result; Step 5: Calculate the discrimination loss according to the discrimination result, and update the discriminator parameters based on the discrimination loss; Step 6: Fix the updated discriminator, calculate the generation loss, and update the generator parameters based on the generation loss; Step 7: Adaptively adjust the corresponding training frequency based on the discrimination loss and the generation loss; Step 8: Repeat steps 2-7 until the maximum number of iterations is reached to obtain a trained generative adversarial network; input the training text vector into the trained generative adversarial network based on topological phase evolution to obtain the training text vector.

[0010] Preferably, the training process of the conditional random field model based on policy optimization includes: Step 1: Input the training feature extraction vector into the conditional random field model, and initialize the feature extraction layer in the conditional random field model; Step 2: Initialize the parameters of the conditional random field model, including the state transition probability and the characteristic function weight; Step 3: The feature extraction layer extracts features from the training feature extraction vector based on the feature extraction weight and feature extraction bias, and uses the Sigmoid activation function to obtain a high-level feature representation; Step 4: Construct a reward function based on the labels predicted by the conditional random field model and the real labels of the text data used for training; the real labels are entity categories; Step 5: Update the state transition probability and feature function weight based on the reward function; Step 6: Repeat iterative steps 3-5 until the maximum number of iterations is reached to obtain a trained conditional random field model; input the feature extraction vector into the trained policy-optimized conditional random field model, and output the entities and their categories in the text data.

[0011] Preferably, the maximum number of iterations is set to 1000 times.

[0012] Preferably, the feature extraction model based on the adaptive fractal exploration optimization algorithm includes a three-layer fully connected neural network.

[0013] Preferably, the sources of the text data used for training include public knowledge bases and professional websites; the text data used for training is stored in JSON format.

[0014] Beneficial effect: This method can more accurately identify subtle entities in the text and significantly improve the accuracy of entity recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 This is a flow chart of the market dishonesty information knowledge graph entity identification method of an embodiment of the present application.

[0017] Figure 2 This is a training flowchart of a feature extraction model based on an adaptive fractal exploration optimization algorithm according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0019] like Figure 1 As shown, this embodiment provides a method for identifying entities in a market dishonest information knowledge graph, the method comprising: S1: Obtain text data of market dishonesty information to be identified.

[0020] S2: Input the text data into the sequence concatenation layer for multi-granularity processing to obtain vocabulary vectors of different structures, and concatenate the vocabulary vectors of different structures to obtain the text vector.

[0021] Specifically, the text data is input into the sequence concatenation layer for multi-granularity processing to obtain vocabulary vectors of different structures, including: The sequence splicing layer includes four structures: B, M, E, and S; The B structure is used to match the text data with characters c i The first word; The M structure is used to match characters in the text data. c i The vocabulary that is in the middle; The E structure is used to match characters in the text data c i The words that are at the end; The S structure is used to represent characters c i itself; Taking the text data "Beijing ×× Technology Co., Ltd. was established on August 12, 2021, and its legal representative is Wang" as an example, the words that conform to the B structure are "Beijing ×× Technology Co., Ltd.", the words that conform to the M structure are "August 12, 2021", and the words that conform to the E structure are "Wang". By collecting the words corresponding to each character, the complete information of the words is obtained, and then the word information is represented by word frequency and splicing.

[0022] Different structures of vocabulary vector expressions include: ; ; ; ; in, Representing characters c i The vocabulary vector matched to the B structure; Representing characters c i The vocabulary vector matched to the M structure; Representing characters c i The vocabulary vector that matches the E structure; Representing characters c i The vocabulary vector matched to the S structure; c i Represents the i-th character; Represents the word whose i-th character comes before the k-th character in the text data; Indicates the word whose i-th character is between the j-th character and the k-th character in the text data; represents the vocabulary whose i-th character comes after the j-th character in the text data; L represents the dictionary; and n represents the text data sequence.

[0023] Furthermore, the step of concatenating vocabulary vectors of different structures to obtain a text vector includes: The vocabulary vectors of each structure are normalized and integrated. The calculation formula includes: ; ; in, represents normalization and integration; s represents any structure; w represents the vocabulary vector of any structure; S represents the vocabulary vector in the text data that conforms to the S structure; B represents the vocabulary vector in the text data that conforms to the B structure; M represents the vocabulary vector in the text data that conforms to the M structure; E represents the vocabulary vector in the text data that conforms to the E structure; Represents the word frequency of vocabulary vector w; represents the vocabulary embedding matrix of the vocabulary vector w; Z represents the sum of the word frequencies of the vocabulary vector w in the vocabulary vectors of the four structures; The vocabulary vectors of all structures after normalization and integration are concatenated to obtain the text vector. The calculation formula is: ; in, represents a text vector; Represents the vocabulary vector of the normalized and integrated B structure; Represents the vocabulary vector of the normalized and integrated M structure; Represents the vocabulary vector of the normalized and integrated E structure; Represents the vocabulary vector of the normalized and integrated S structure.

[0024] By using multi-granular features, the system can parse and vectorize text data more carefully, better capture entity and contextual information in the text, and reduce information loss caused by vocabulary loss.

[0025] S3: Input the text vector into a feature extraction model based on an adaptive fractal exploration optimization algorithm, and output a feature extraction vector.

[0026] like Figure 2 As shown, this embodiment provides a training process of a feature extraction model based on an adaptive fractal exploration optimization algorithm, including: Step 1: Collect text data for training and preprocess it; The sources of text data for training include public knowledge bases and professional websites; the text data for training is stored in JSON format. In this embodiment, the text data for training is collected from multiple source domains using transfer learning. The data in the source domain includes enterprise information sets and judicial documents. The information text records the enterprise name, establishment date and corporate legal person in detail, and contains a wealth of technical and scientific entities.

[0027] A piece of text data may be a text describing corporate information / market default information, which contains entities in multiple fields. For example, "Beijing ×× Technology Co., Ltd. was established on August 12, 2021, and the legal representative is Wang Moumou." The entities are "Beijing ×× Technology Co., Ltd.", "August 12, 2021", and "Wang Moumou", which belong to the three entity categories of company name, establishment date, and legal representative respectively.

[0028] Step 2: Input the preprocessed training text vector into the feature extraction model; Step 3: Select the initial parameter set of the feature extraction model and initialize the search granularity and search radius ; The initial parameter set Including the weights of the neural network and bias In this embodiment, the parameter is set to 0.1, is set to 1, the parameter set The initialization method is random initialization.

[0029] Step 4: According to the parameter point of the current iteration , generating a set of self-similar exploration points , j represents the jth exploration point; all the exploration points are distributed around the parameter point of the current iteration according to the fractal rule based on the search granularity and the search radius; The sine and cosine functions are used to generate points distributed on the spherical surface of a high-dimensional space to simulate the distribution of fractal structures in nature. In one embodiment, the exploration points are distributed in the following form: ; In the formula, is the number of exploration points generated in the current iteration, is the current search granularity, is the current search radius.

[0030] Further, and The two items actually define the position of a point on a unit circle in a high-dimensional parameter space. To make the adjustment so that the distribution of the exploration points can be richer and more flexible, and the exploration points are not only distributed on a standard circular trajectory, but can be adjusted based on the fractal dimension D, in another embodiment, this embodiment uses the fractal dimension D to adjust the angle distribution, and the calculation method can be expressed as: ; In the formula, Represents the fractal dimension, which affects the distribution density of the exploration points in space and how to adjust the distribution of these points according to fractal theory.

[0031] Step 5: Calculate the loss function value of all exploration points, calculate the information entropy of all exploration points based on the loss function value, and construct a search granularity adjustment function based on the information entropy; Specifically, a dynamic information entropy feedback mechanism is used to optimize the search process of neural network parameters. By calculating the information entropy of the current exploration point set, the search strategy is dynamically adjusted to respond to the complexity changes in the parameter space, the search granularity and direction are dynamically adjusted, and the global search ability and adaptability of the algorithm are enhanced, so as to explore the parameter space more effectively. Suppose the set of all exploration points in the current iteration is ,in is the number of iterations, is the index of the exploration point. Further, for each exploration point, calculate its loss function value , and normalize the loss function values ​​so that their sum is 1. The normalization process can be expressed as: ; In the formula, Represents an exploration point Normalized weights of; loss function is the cross entropy loss.

[0032] The information entropy of all exploration points in the current iteration is calculated based on the normalized weights. The calculation formula is: ; in, represents information entropy; The expression of the search granularity adjustment function is: ; in, Represents the search granularity adjustment function; is a preset information entropy threshold, which is set to 0.8 in this embodiment; Step 6: dynamically adjusting the search granularity based on the search granularity adjustment function; Specifically, if the loss of multiple exploration points is low, the search granularity is reduced to refine the search; if the loss is high, the search granularity is increased to speed up the search. The adaptive dynamic adjustment of the search granularity ensures that the search granularity is refined when a better solution is found, and is increased when it is difficult to find a better solution to speed up the exploration. At the Discovery Point Dynamically adjust the search granularity can be expressed as: ; ; In the formula, is the reduction factor, is the expansion factor, The parameter is The loss value when is an indicator based on historical search efficiency. In this embodiment, the reduction factor is set to 0.9, the expansion factor is set to 1.1; K represents the number of historical iterations; represents the im-th iteration, and the parameter is The loss value at time It means the i-m+1th iteration, the parameter is The loss value at that time.

[0033] Step 7: Select the exploration point with the smallest loss function value as the parameter point of the next iteration, and dynamically adjust the search radius according to the loss between the parameter point of the current iteration and the parameter point of the next iteration; the calculation formula is: ; ; ; in, is the optimal parameter exploration point for the current iteration, is the transformation function; is a point based on the current and the transformation parameter set Dynamic adjustment function, used to control the local enlargement or reduction of the parameter space; is a regulating factor, It is the inverse of the loss gradient covariance matrix of the current exploration point set, which is used to measure the heterogeneity of the loss function changes in different directions in the parameter space.

[0034] Dynamically adjust functions By dynamically adjusting the representation of the parameter space, the algorithm's search efficiency and accuracy in different areas are improved. By dynamically transforming the parameter space, the algorithm can more effectively adapt to the local characteristics of the loss function. The dynamic nonlinear transformation can be regarded as the use of an adaptive mirror in the parameter space. The adaptive mirror dynamically adjusts its shape according to the local characteristics of the loss function, thereby achieving local magnification or reduction of the search space and accelerating the algorithm to find the optimal solution. Furthermore, the transformation function middle represents a point in the original parameter space, represents a set of transformation parameters, and the transformation function The purpose of is to map the original parameter space into a new space so that search can be performed in the new space.

[0035] Furthermore, a search radius adjustment function is constructed based on the information entropy, and its expression is: ; in, represents the search radius adjustment function; Furthermore, according to the effect of the current iteration, the search radius is dynamically adjusted to optimize the breadth and depth of the search to more effectively explore the parameter space. The search radius is dynamically adjusted according to the degree of improvement of the loss function, and the adjustment method can be expressed as: ; In the formula, , is a threshold value set to determine whether the loss function has been significantly improved. is set to 1.05, the reduction factor is set to 0.95, Set to 0.01.

[0036] This embodiment also provides another search radius adjustment method, including: In order to make the adjustment of the search radius more detailed, so that the search radius can be adaptively adjusted according to the actual improvement of the loss function, in another embodiment, a nonlinear adjustment factor is defined. Consider the improvement of loss function With threshold The relationship can be expressed as: ; The search radius is adjusted by the nonlinear adjustment factor, which can be expressed as: ; in, represents the updated search radius; represents the search radius of the last iteration; Indicates the scale factor, which is preset manually.

[0037] Step 8: Repeat iterative steps 4-7 until the maximum number of iterations is reached to obtain a trained feature extraction model; input the training text vector into the trained feature extraction model based on the adaptive fractal exploration optimization algorithm, and output the training feature extraction vector.

[0038] The expanded data is input into the feature extraction model to train the feature extraction model. This embodiment uses a 3-layer fully connected neural network for feature extraction. Some existing solutions use neural networks for feature extraction. In some neural network structures, traditional neural network parameter optimization algorithms, such as gradient descent and its variants, mainly rely on calculating the gradient of the loss function with respect to each parameter to guide the parameter update direction. However, such methods are prone to fall into local optimal solutions when faced with complex, non-convex loss surfaces.

[0039] This embodiment adopts a neural network feature extraction model based on an adaptive fractal exploration optimization algorithm, which is inspired by the self-similarity and ubiquitous characteristics of fractal structures in nature, as well as the complexity and self-organization in fractal theory. Fractal structures are widely present in nature, such as the shape of snowflakes, the outline of mountains, etc., which show similarities at different scales. In this embodiment, this self-similarity principle is introduced into the parameter optimization process of the neural network, and the parameter space of the neural network is explored by simulating the generation process of fractal structures in nature, thereby achieving efficient optimization of the weights and biases of the neural network.

[0040] The adaptive fractal exploration optimization algorithm proposed in this embodiment does not rely on gradient information, but uses a self-similar exploration mode to search in the parameter space, so that the algorithm has a higher probability of jumping out of the local optimal solution and exploring a wider parameter space. In addition, the adaptive fractal exploration optimization algorithm adopts an adaptive mechanism to dynamically adjust the granularity and direction of the search according to the current search state, so as to adapt to different task scenarios more flexibly, and can effectively improve the parameter optimization efficiency of the neural network.

[0041] The adaptive fractal exploration optimization algorithm does not rely on gradient information. It explores the parameter space of neural networks by simulating fractal structures in nature, improves the efficiency of the parameter optimization process and the global search capability, and helps avoid falling into local optimality.

[0042] Further, the preprocessing includes: The text data used for training is passed through the sequence concatenation layer to obtain the text vector for training; The generative adversarial network based on topological phase evolution performs data expansion on the training text vector to obtain the training text vector.

[0043] It is understandable that the collection, acquisition, annotation and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. This embodiment adopts a generative adversarial network algorithm based on topological phase evolution. The generative adversarial network algorithm based on topological phase evolution includes two components: a generator (G) and a discriminator (D). The input of the generator is a vectorized feature, which is responsible for generating data that is as close to the real data distribution as possible. The task of the discriminator is to distinguish between the generated data and the real data. In order to enhance the stability of the model and the quality of the generated data, the present invention adopts a stability loss function in the traditional generative adversarial network framework. The stability loss function takes into account the topological structure differences between the generated data and the real data, thereby guiding the generator to better capture the distribution characteristics of the real data.

[0044] Furthermore, the training process of the topological phase evolution-based generative adversarial network includes: Step 1: Randomly initialize the network parameters of the generator and discriminator; Step 2: Adjust the generator's generation strategy based on the real dataset and the generated dataset of the current iteration; Adjusting the generation strategy of the generator includes: adjusting the generated data through the topological phase evolution strategy to reduce the difference in topological structure with the real data set, setting is the generated data at iteration t, For real data sets, topological dissimilarity The adjustment method can be expressed as: ; ; in, is the data batch size; is a mapping function that maps data to a topological space; represents the L2 norm; is the Sigmoid activation function, and are the weight and bias of the mapping function respectively; Represents the i-th real data; Represents the i-th generated data at iteration t.

[0045] Step 3: Use the adjusted generator to generate a batch of fake data from random noise; specifically: Generator Receive noise vector As input, the generated data of the output can be expressed as:

[0046] In the formula, Represents the parameters of the generator at time t.

[0047] Step 4: Input the false data and the real data into the discriminator to obtain the discrimination result; Step 5: Calculate the discrimination loss according to the discrimination result, and update the discriminator parameters based on the discrimination loss; the discrimination loss calculation formula is: ; in, represents the discrimination loss; represents the i-th noise vector; represents the discriminator; Represents a generator; The discriminator parameter update formula is: ; in, represents the updated discriminator parameters; represents the discriminator parameters before updating; represents the discriminator learning rate; Represents the gradient of the discriminator parameters before and after the update with respect to the loss function.

[0048] Step 6: Fix the updated discriminator, calculate the generation loss, and update the generator parameters based on the generation loss; In this embodiment, a stability loss function is used to guide the generator to generate data that is closer to the real data distribution, and the topological structure difference is used as feedback to prompt the generator to generate higher quality data; the generation loss calculation formula is: ; ; in, represents the generation loss; The weight coefficient representing the topological dissimilarity; is a parameter that adjusts the steepness of the curve. is the topological dissimilarity threshold. When the topological dissimilarity exceeds When the weight coefficient Increase, strengthen the penalty for topological dissimilarity. In this embodiment, Set to 4.

[0049] The generator parameter update formula is: ; in, Represents the updated generator parameters; Represents the generator parameters before updating; represents the generator learning rate; Represents the gradient of the generator parameters before and after the update with respect to the loss function.

[0050] Step 7: Adaptively adjust the corresponding training frequency based on the discriminant loss and the generator loss; if the discriminator performs much better than the generator, increase the training frequency of the generator, otherwise reduce the training frequency of the generator. Specifically, the way to increase the learning rate of the generator can be expressed as: ; ; In the formula, is the updated generator learning rate, is the adjustment factor; is the tanh activation function, and is the first adjustment factor hyperparameter and the second adjustment factor hyperparameter, which are used to control the sensitivity and direction of the learning rate adjustment. The function ensures that the adjustment factor varies within a reasonable range. In this embodiment, and Set them to 0.4 and 8 respectively.

[0051] Similarly, if the generator loss is low, decrease the generator learning rate and increase the discriminator learning rate.

[0052] Step 8: Repeat steps 2-7 until the maximum number of iterations is reached to obtain a trained generative adversarial network; input the training text vector into the trained generative adversarial network based on topological phase evolution to obtain the training text vector.

[0053] The concept of topological phase evolution is adopted in generative adversarial networks, and the data quality and stability of the generative model are improved by considering the topological structure differences between generated data and real data.

[0054] S4: Input the feature extraction vector into a conditional random field model based on strategy optimization, and output entities and their categories in the text data.

[0055] The feature-extracted data is input into the conditional random field model for classifier training. This embodiment adopts a conditional random field classification algorithm based on policy optimization, which is a probabilistic graph model for labeling and segmenting sequence data. This embodiment improves the traditional conditional random field model and adopts a feature extraction layer based on deep learning to automatically learn complex feature representations from data instead of relying on manual features.

[0056] This embodiment provides a training process of a conditional random field model based on policy optimization, including: Step 1: Input the training feature extraction vector into the conditional random field model, and initialize the feature extraction layer in the conditional random field model; Step 2: Randomly initialize the parameters of the conditional random field model, including the state transition probability and feature function weights ; Given a sequence and state sequence , the conditional probability of the conditional random field model can be expressed as: ; ; in, Indicates location The characteristic function at is a normalization factor to ensure that the sum of probabilities is 1; T represents transpose; Indicates status To status The state transition probability.

[0057] Step 3: The feature extraction layer extracts features from the training feature extraction vector based on the feature extraction weight and feature extraction bias, and uses the Sigmoid activation function to obtain a high-level feature representation; The feature vector after feature extraction is used as input to initialize the deep learning-based feature extraction layer, which is responsible for capturing complex patterns and relationships in the input data. Suppose the input feature vector is , the high-level features are expressed as , then the process of feature extraction layer processing can be expressed as: ; In the formula, represents the Sigmoid activation function, and They represent feature extraction weight and feature extraction bias respectively.

[0058] In weight The update is performed through the back propagation algorithm. For a given loss function , the weight update method can be expressed as: ; In the formula, is the weight update learning rate, is the loss function with respect to weights The gradient of the loss function is the cross entropy loss. In this embodiment, Set to 0.01.

[0059] Step 4: Construct a reward function based on the labels predicted by the conditional random field model and the real labels of the text data used for training; the real labels are entity categories; Specifically, a reinforcement learning strategy is used to optimize the parameters of the conditional random field model based on a predefined reward function. Preferably, the reward function is an accuracy calculation function for entity recognition.

[0060] This process is repeated multiple times. In each iteration, the model tries different parameter configurations and adjusts them according to the effect of entity recognition. Evaluating model output With the true label The matching degree of strategy optimization aims to maximize the expected reward , which can be expressed as: ; in, express expectations; Step 5: Update the state transition probability and feature function weight based on the reward function; the update formula is: ; ; in, and are the first parameter learning rate and the second parameter learning rate, respectively. For a given input Time state sequence The conditional probability of and For the expected reward and The gradient of and They are the updated state transition probability matrix and characteristic function weight respectively; Step 6: Repeat iterative steps 3-5 until the maximum number of iterations is reached to obtain a trained conditional random field model; input the feature extraction vector into the trained policy-optimized conditional random field model, and output the entities and their categories in the text data.

[0061] In this embodiment, the maximum number of iterations adopted above is set to 1000 times.

[0062] The policy-optimized conditional random field model further adjusts the model parameters through policy optimization, which can perform entity recognition and classification more accurately and optimize the accuracy and efficiency of the labeling process.

[0063] The trained model is used to process new samples and complete the knowledge graph entity recognition task based on multi-granularity features. In the embodiment, in the task of building a knowledge graph of market subject dishonesty information, a large amount of corporate information and judicial documents are collected as data sources. The data source contains rich entity information, such as company name, establishment date, legal representative, etc. These entities need to be accurately identified and added to the knowledge graph.

[0064] In entity recognition, the input is a new enterprise information set, which contains unprocessed text data. The text data is first preprocessed, and then the text data is input into the trained sequence concatenation layer to be converted into vectorized form. Furthermore, the quantized features are input into the trained feature extraction model to obtain the feature-extracted vector. Furthermore, the feature-extracted vector is input into the trained policy-optimized conditional random field model to identify the entities and their categories in the text.

[0065] In one of the entity recognition tasks in this embodiment, from a text about corporate information, the model identified "Beijing ×××× Technology Co., Ltd." as the company name, "April 9, 2002" as the establishment date, and "Yang Mou" as the legal representative.

[0066] The market dishonesty information knowledge graph entity recognition method provided in this embodiment has the following beneficial effects: 1. It can more accurately identify subtle entities in the text, significantly improving the accuracy of entity recognition.

[0067] 2. The data generated by the topologically evolved generative adversarial network is of high quality and can be effectively used for training, reducing overfitting and enhancing the generalization ability of the model.

[0068] 3. The adaptive fractal exploration optimization algorithm improves the global search capability, making the model more likely to escape the local optimal solution.

[0069] 4. The conditional random field model for strategy optimization improves processing speed and operating efficiency through automatic feature extraction and rapid parameter adjustment.

[0070] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A method for identifying entities in a knowledge graph of market dishonesty information, characterized in that: include: S1: Obtain text data of market dishonesty information to be identified; S2: inputting the text data into a sequence concatenation layer for multi-granularity processing to obtain vocabulary vectors of different structures, and concatenating the vocabulary vectors of different structures to obtain a text vector; S3: inputting the text vector into a feature extraction model based on an adaptive fractal exploration optimization algorithm, and outputting a feature extraction vector; S4: Input the feature extraction vector into a conditional random field model based on strategy optimization, and output entities and their categories in the text data.

2. The market dishonesty information knowledge graph entity recognition method according to claim 1 is characterized in that: The inputting of the text data into the sequence concatenation layer for multi-granularity processing to obtain vocabulary vectors of different structures includes: The sequence splicing layer includes four structures: B, M, E, and S; The B structure is used to match the text data with characters c i The first word; The M structure is used to match characters in the text data. c i The vocabulary that is in the middle; The E structure is used to match characters in the text data c i The words that are at the end; The S structure is used to represent characters c i itself; Different structures of vocabulary vector expressions include: ； ； ； ； in, Representing characters c i The vocabulary vector matched to the B structure; Representing characters c i The vocabulary vector matched to the M structure; Representing characters c i The vocabulary vector that matches the E structure; Representing characters c i The vocabulary vector matched to the S structure; c i Represents the i-th character; Represents the word whose i-th character comes before the k-th character in the text data; Indicates the word whose i-th character is between the j-th character and the k-th character in the text data; represents the vocabulary whose i-th character comes after the j-th character in the text data; L represents the dictionary; and n represents the text data sequence.

3. The market dishonesty information knowledge graph entity recognition method according to claim 2 is characterized in that: The concatenation of vocabulary vectors of different structures to obtain a text vector comprises: The vocabulary vectors of each structure are normalized and integrated. The calculation formula includes: ； ； in, represents normalization and integration; s represents any structure; w represents the vocabulary vector of any structure; S represents the vocabulary vector in the text data that conforms to the S structure; B represents the vocabulary vector in the text data that conforms to the B structure; M represents the vocabulary vector in the text data that conforms to the M structure; E represents the vocabulary vector in the text data that conforms to the E structure; Represents the word frequency of vocabulary vector w; represents the vocabulary embedding matrix of the vocabulary vector w; Z represents the sum of the word frequencies of the vocabulary vector w in the vocabulary vectors of the four structures; The vocabulary vectors of all structures after normalization and integration are concatenated to obtain the text vector. The calculation formula is: ； in, represents a text vector; Represents the vocabulary vector of the normalized and integrated B structure; Represents the vocabulary vector of the normalized and integrated M structure; Represents the vocabulary vector of the normalized and integrated E structure; Represents the vocabulary vector of the normalized and integrated S structure.

4. The market dishonesty information knowledge graph entity recognition method according to claim 3 is characterized in that: It also includes training the feature extraction model, and the training process includes: Step 1: Collect text data for training and preprocess it; Step 2: Input the preprocessed training text vector into the feature extraction model; Step 3: Select an initial parameter set of the feature extraction model and initialize the search granularity and search radius; the initial parameter set includes the weight and bias of the neural network; Step 4: Generate a set of self-similar exploration points based on the parameter point of the current iteration; all the exploration points are distributed around the parameter point of the current iteration according to the fractal rule based on the search granularity and the search radius; Step 5: Calculate the loss function value of all exploration points, calculate the information entropy of all exploration points based on the loss function value, and construct a search granularity adjustment function based on the information entropy; Step 6: dynamically adjusting the search granularity based on the search granularity adjustment function; Step 7: Select the exploration point with the smallest loss function value as the parameter point of the next iteration, and dynamically adjust the search radius according to the loss between the parameter point of the current iteration and the parameter point of the next iteration; Step 8: Repeat iterative steps 4-7 until the maximum number of iterations is reached to obtain a trained feature extraction model; input the training text vector into the trained feature extraction model based on the adaptive fractal exploration optimization algorithm, and output the training feature extraction vector.

5. The market dishonesty information knowledge graph entity recognition method according to claim 4 is characterized in that: The pre-processing comprises: The text data used for training is passed through the sequence concatenation layer to obtain the text vector for training; The generative adversarial network based on topological phase evolution performs data expansion on the training text vector to obtain the training text vector.

6. The market dishonesty information knowledge graph entity recognition method according to claim 5 is characterized in that: The training process of the generative adversarial network based on topological phase evolution includes: Step 1: Initialize the network parameters of the generator and discriminator; Step 2: Adjust the generator's generation strategy based on the real dataset and the generated dataset of the current iteration; Step 3: Use the adjusted generator to generate a batch of fake data from random noise; Step 4: Input the false data and the real data into the discriminator to obtain the discrimination result; Step 5: Calculate the discrimination loss according to the discrimination result, and update the discriminator parameters based on the discrimination loss; Step 6: Fix the updated discriminator, calculate the generation loss, and update the generator parameters based on the generation loss; Step 7: Adaptively adjust the corresponding training frequency based on the discrimination loss and the generation loss; Step 8: Repeat steps 2-7 until the maximum number of iterations is reached to obtain a trained generative adversarial network; input the training text vector into the trained generative adversarial network based on topological phase evolution to obtain the training text vector.

7. The market dishonesty information knowledge graph entity recognition method according to claim 5 is characterized in that: The training process of the policy optimization-based conditional random field model includes: Step 1: Input the training feature extraction vector into the conditional random field model, and initialize the feature extraction layer in the conditional random field model; Step 2: Initialize the parameters of the conditional random field model, including the state transition probability and the characteristic function weight; Step 3: The feature extraction layer extracts features from the training feature extraction vector based on the feature extraction weight and feature extraction bias, and uses the Sigmoid activation function to obtain a high-level feature representation; Step 4: Construct a reward function based on the labels predicted by the conditional random field model and the real labels of the text data used for training; the real labels are entity categories; Step 5: Update the state transition probability and feature function weight based on the reward function; Step 6: Repeat iterative steps 3-5 until the maximum number of iterations is reached to obtain a trained conditional random field model; input the feature extraction vector into the trained policy-optimized conditional random field model, and output the entities and their categories in the text data.

8. The market dishonesty information knowledge graph entity recognition method according to claim 7 is characterized in that: The maximum number of iterations is set to 1000.

9. The market dishonesty information knowledge graph entity recognition method according to claim 4 is characterized in that: The feature extraction model based on the adaptive fractal exploration optimization algorithm includes a three-layer fully connected neural network.

10. The market dishonesty information knowledge graph entity identification method according to claim 4 is characterized in that: The sources of text data used for training include public knowledge bases and professional websites; the text data used for training is stored in JSON format.

Images