A federated learning model aggregation method, system and terminal for entity extraction

Abstract

The application belongs to the technical field of natural language processing, and discloses a federated learning model aggregation method, system and terminal for entity extraction. k ; the server calculates a weight vector alpha of model aggregation based on entity information E of all clients, wherein each component alpha k represents the weight when the model is aggregated; and the server aggregates each client model based on alpha to obtain a global model. Experiments show that the federated learning model aggregation method can effectively overcome the problem of model accuracy decline under non-independent and identically distributed data, and can reduce the communication cost of model training; under the premise of reducing the risk of data privacy leakage, the model F1 value is only 1.37% lower than that of the centralized training model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of natural language processing, and particularly relates to a federated learning model aggregation method and system for entity extraction and a terminal. BACKGROUND

[0002] Currently, the entity extraction task mainly identifies words with special meanings from text data, which includes two steps: firstly, identifying the start and end positions of entities in a text sequence, and secondly, classifying the extracted entities. Entity extraction has great significance in various applications, such as medical entity extraction, which identifies disease type entities and drug type entities in electronic medical records, and plays a significant role in building drug warning systems, clinical decision support systems, and scientific research and teaching. In addition, most natural language processing technologies require entity extraction tasks as a foundation. For example, in a question and answer system, the accuracy of the system's semantic understanding and analysis of user questions depends on the entity information in the question, and the answer to the reasoning is usually composed of entities in the knowledge base.

[0003] In recent years, deep learning-based entity extraction methods have good results. Deep learning can extract more granular semantic features from text. At the same time, based on its powerful ability to automatically extract features, researchers do not need to perform a large amount of feature engineering on the text, reducing the requirement for domain knowledge of the text, such as CNN-CRF, BiLSTM-CRF, CNN-BiLSTM-CRF, BERT-CRF, and ALBERT-BiLSTM-CRF. However, these methods usually require a large amount of labeled data, and the labeled data that meets the training requirements of the entity extraction task on various platforms is limited, and manually labeling data for the entity extraction task is very expensive and time-consuming, requiring a large amount of domain-specific knowledge. In addition, text data in some fields has high privacy, such as medical fields involving patient conditions, genetic sequences, and pathology reports. And relevant laws and regulations have explicitly enacted bills that require enterprise units to protect user data privacy and security and strictly limit the scope of data transactions. This has caused various platforms to be unable to share entity labeling data, resulting in the "data island" problem, which is a great challenge to traditional entity extraction methods.

[0004] The federal learning can effectively solve the above problems, which enables each platform to conduct machine learning under the premise of protecting data privacy and meeting laws and regulations, and solves the problems of data lack and data island. The federal learning aims to coordinate multiple clients to jointly build a machine learning model in a distributed environment. And the training data set of each client does not need to be exposed to the rest of the clients, only the relevant information of model training needs to be exchanged. The performance of the final federal learning model approaches the performance of the centralized model (the machine learning model obtained by training all the training data of the clients together). In recent years, federal learning has been widely used in medical image processing, natural language processing, recommendation system and other fields, helping to solve the problem of lack of data due to data privacy protection in model training, promoting the development of artificial intelligence in the enterprise field, and bringing great commercial application value.

[0005] At present, only Ge S et al. apply the federal learning framework to the entity extraction task. This research proposes a medical entity extraction model for English corpus medical text, and builds a personalized federal learning framework based on FedAVG algorithm. The F1 values of this scheme on the data sets CADEC, ADE Corpus and SMM4H data set are 65.16, 82.57 and 32.69 respectively. The federal aggregation is to add the model parameters of each client by weighting, and obtain a global model. However, this research does not consider the change of the importance of the clients due to the difference of the training data, but simply considers that the importance of each client is equal, resulting in that the convergence speed of the global model is not fast enough and the accuracy is not high enough.

[0006] In summary, entity extraction is a key technology of natural language processing. The existing researches usually use deep learning models to process entity extraction tasks, which need sufficient entity labeled data for training. However, the entity labeled data of a single client is usually limited, and the data between clients often needs privacy protection and cannot be directly shared. The existing researches propose to use federal learning to solve the privacy security problem of sharing entity labeled data. The federal aggregation is to add the model parameters of each client by weighting, and obtain a global model. The federal learning framework of the traditional entity extraction simply considers that the importance of each client is equal, and does not consider the change of the importance of the clients due to the difference of the training data.

[0007] Through the above analysis, the problems and defects of the prior art are:

[0008] (1) The existing entity extraction method based on deep learning usually needs a large amount of labeled data, and the labeled data for training the entity extraction task on each platform is limited, and it is very expensive and time-consuming to manually label data for the entity extraction task.

[0009] (2) In deep learning-based entity extraction methods, text data in some fields are highly private, and relevant laws and regulations strictly limit the scope of data transactions. This makes it impossible for various platforms to share entity annotation data, resulting in the problem of data silos, which poses a huge challenge to traditional entity extraction methods.

[0010] (3) Existing federated learning-based entity extraction methods do not consider the changes in client importance due to different training data. They simply assume that each client is equally important, resulting in slow global model convergence and low accuracy. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention provides a federated learning model aggregation method, system, and terminal for entity extraction.

[0012] This invention is implemented as follows: a federated learning model aggregation method oriented towards entity extraction, the method comprising:

[0013] For each client Calculate entity information e k The server calculates the weight vector α of the aggregated model based on the entity information E of all clients, where each component α k When representing model aggregation The weights; the server assigns weights to each client model based on α. Aggregate to obtain the global model

[0014] Furthermore, the federated learning model aggregation method for entity extraction includes the following steps:

[0015] Step 1: Calculate the information entropy, number of entities, and entity dispersion of the entity annotation dataset. Information entropy, number of entities, and entity dispersion can all measure the importance of the dataset to the entity extraction task and are used to calculate the aggregation weight for the client.

[0016] Step two involves calculating the entropy of the concatenated text, the number of entities, and the entity dispersion to obtain the entity information vector; this is achieved by concatenating these three pieces of information.

[0017] Step 3: Normalize the three components of the entity information vector for each client. Normalization ensures that the aggregation model will not experience parameter explosion and that the sum of the proportions of each client is 1.

[0018] Step four: Perform a weighted average on the vectors obtained in step three to obtain the aggregated weights of the client model.

[0019] Further, the step one includes calculating the information entropy of the entity-labeled dataset, comprising:

[0020] (1) Traverse the dataset D k , count the occurrence of each word, and obtain the word dictionary Q;

[0021] (2) Calculate the probability of each character q in Q appearing in D, and obtain the vector R = [r1, r2, …, r k ], where r k is the probability of character q |Q| appearing;

[0022] (3) Calculate the text information entropy h of the dataset D i .

[0023] The calculation of the number of entities and the entity dispersion degree includes:

[0024] Count the occurrence of each entity in the dataset D i , and obtain the entity category vector V;

[0025] According to the following formula, calculate the number of entities γ and the entity dispersion degree σ respectively:

[0026]

[0027]

[0028] Further, in the step two, the entity information vector [h, γ, σ] is obtained by splicing the text information entropy h, the number of entities γ and the entity dispersion degree σ of the dataset D k .

[0029] Further, in the step three, according to the following formula, normalize the three components of the entity information vector [h k , γ k , σ t ] of each client in C k , and obtain the vector

[0030]

[0031] Further, in the step four, according to the following formula, weighted average the vector , and obtain the model aggregation weight α k of the client :

[0032]

[0033] Another object of the present application is to provide an entity extraction-oriented federated learning model aggregation system applying the entity extraction-oriented federated learning model aggregation method.

[0034] An entity information calculation module is configured to calculate the information entropy, the number of entities, and the entity dispersion of the entity annotation dataset.

[0035] An entity information vector calculation module is configured to calculate the information entropy, the number of entities, and the entity dispersion of the spliced text to obtain an entity information vector.

[0036] An entity information component normalization module is configured to normalize the three components of the entity information vector of each client.

[0037] A model aggregation weight calculation module is configured to perform weighted averaging on the vector obtained by the component normalization module to obtain the model aggregation weight of the client.

[0038] Another object of the present application is to provide a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to enable the processor to perform the steps of the entity extraction-oriented federated learning model aggregation method.

[0039] Another object of the present application is to provide a computer-readable storage medium storing a computer program, and the computer program is executed by a processor to enable the processor to perform the steps of the entity extraction-oriented federated learning model aggregation method.

[0040] Another object of the present application is to provide an information data processing terminal for implementing the entity extraction-oriented federated learning model aggregation system.

[0041] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present application are analyzed from the following aspects:

[0042] First, in view of the technical problems existing in the prior art and the difficulty in solving the problems, the technical solution to be protected by the present application is closely combined with the results and data in the research and development process, and the technical problems solved by the technical solution are analyzed in detail and deeply, and some creative technical effects brought by the solution of the problems are described as follows:

[0043] The application provides a federated learning model aggregation method FedEI for entity extraction.

[0044] The experimental results show that the federated learning model aggregation method can effectively overcome the problem of model accuracy decline under non-independent and identically distributed data and can reduce the communication cost of model training. Under the premise of reducing the risk of data privacy leakage, the model accuracy provided by the application is close to the centralized training model, and the model F1 value is only 1.37% lower than the centralized training model. Compared with the existing entity extraction method based on the federated learning framework FedAVG, the model F1 value of the method provided by the application on the Boson data set and the micro-blog data set is significantly improved, and the total data transmission amount of a single client is significantly reduced.

[0045] Secondly, from the perspective of the product or as a whole, the technical effect and advantages of the technical solution to be protected by the application are described as follows:

[0046] The application provides a federated learning model aggregation method FedEI for entity extraction.

[0047] The application provides a federated learning model aggregation method FedEI for entity extraction.

[0048] Thirdly, the creativity of the claims of the application is also embodied in the following important aspects:

[0049] The expected income and commercial value of the technical scheme of the present application after transformation are: (1) the existing model aggregation method simply regards each participating model aggregation client as equally important without considering the importance changes caused by different client data. The present application side evaluates the importance of each client data set without revealing the real data set, obtains the aggregation weight of each model, and improves the model convergence speed and performance of the entity extraction-based federated learning. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical scheme of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0051] Figure 1 is a federated learning model aggregation method flowchart for entity extraction provided by the embodiments of the present application;

[0052] Figure 2 is a schematic diagram of an entity information and weight calculation process provided by the embodiments of the present application;

[0053] Figure 3 is a model training F1 value curve diagram of two algorithms provided by the embodiments of the present application;

[0054] Figure 4 is a model training communication frequency diagram of two algorithms provided by the embodiments of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present application more clear and understandable, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0056] In view of the problems existing in the prior art, the present application provides a federated learning model aggregation method, system and terminal for entity extraction, which will be described in detail below in combination with the drawings.

[0057] I. Explanation of Embodiments. In order to enable those skilled in the art to fully understand how the present application is specifically implemented, this part is an explanation of the embodiments of the technical scheme of the claims.

[0058] As shown in Figure 1 The federated learning model aggregation method for entity extraction provided by the embodiments of the present application includes the following steps:

[0059] S101, calculate the information entropy, number of entities, and entity dispersion of the entity annotation dataset;

[0060] S102, calculate the entropy of the concatenated text, the number of entities, and the entity dispersion to obtain the entity information vector;

[0061] S103, normalize each of the three components of the entity information vector for each client;

[0062] S104: Perform a weighted average on the vectors obtained in S103 to obtain the client's model aggregation weights.

[0063] As a preferred embodiment, this invention proposes a model aggregation algorithm based on entity information, FedEI. First, the concept of entity information is introduced, consisting of three parts: text information entropy of the entity-annotated dataset, the number of entities, and entity dispersion. Based on this, appropriate model aggregation weights are designed. Entity information does not involve the original entity-annotated dataset; it is only an evaluation value of data quality. Therefore, there is no risk of data privacy leakage when entity information is transmitted between the server and client. The overall framework of the model aggregation algorithm based on entity information is described below. The process of model aggregation in the t-th round is shown in Algorithm 1. First, for each client... Calculate entity information e k Secondly, the server calculates the weight vector α of the aggregated model based on the entity information E of all clients, where each component α k When representing model aggregation The weights are then assigned. Finally, the server assigns weights to each client model based on α. Aggregate to obtain the global model

[0064] Table 1. The process of model aggregation in round t.

[0065]

[0066] The entity information calculation method and weight calculation method provided in the embodiments of the present invention are as follows: Figure 2 As shown. The first step is to calculate the information entropy of the entity annotation dataset. First, traverse the dataset D. k First, count the occurrences of each word to obtain the dictionary Q. Then, calculate the probability of each character q in Q appearing in D, resulting in the vector R = [r1, r2, ..., r...]. |Q| ], where r i For the character q i The probability of occurrence. Calculate the dataset D according to formula (2). k The text information entropy h. The second step is to calculate the number of entities and the entity dispersion. Statistical data set D. kThe number of occurrences of each type of entity is obtained to obtain an entity category vector V. The entity number γ and the entity dispersion σ are calculated according to formulas (3) and (1) respectively. In the third step, h, γ and σ are spliced to obtain an entity information vector [h, γ, σ]. In the fourth step, the C t In the entity information vector [h ,γ k ,σ k ,σ k ] of each client The three components of the entity information vector [h ,γ k ,σ k ,σ k ] of each client The model aggregation weight α k of the client is obtained by weighted average according to formula (5).

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] The federated learning model aggregation system for entity extraction provided by the embodiment of the application comprises:

[0073] An entity information calculation module is configured to calculate an entity annotation dataset information entropy, an entity number and an entity dispersion;

[0074] An entity information vector calculation module is configured to calculate an entity information vector by splicing a text information entropy, an entity number and an entity dispersion;

[0075] An entity information component normalization module is configured to normalize the three components of the entity information vector of each client respectively;

[0076] A model aggregation weight calculation module is configured to perform weighted average on the vector obtained by the component normalization module to obtain the model aggregation weight of the client.

[0077] II. Application Embodiment. In order to prove the creativity and technical value of the technical scheme of the application, this part is an application embodiment of the technical scheme of the claim on a specific product or related technology.

[0078] As Figure 1As shown, the importance of each client data set is evaluated from the public data set of each client, the aggregation weight of each model is obtained, and the model convergence speed and performance of federated learning based on entity extraction are improved. After the application is adopted, the various indicators of the algorithm FedEI proposed in the application are about 1% higher than FedAVG on the three data of 1998 People's Daily Chinese entity annotation data set, Boson entity annotation data set and microblog entity annotation data set. Therefore, for various federated learning tasks such as entity extraction or entity and relationship joint extraction, the importance of each client data set can be evaluated from the public data set of each client, the aggregation weight of the model is obtained, and the model convergence speed and performance of federated learning based on entity extraction are improved.

[0079] III. Evidence of the effects of the embodiments. The embodiments of the application have achieved some positive effects during research and development or use, and indeed have great advantages compared with the prior art. The following content is described in combination with experimental data, graphs and the like.

[0080] 1. Experimental data and evaluation indicators

[0081] The application uses 1998 People's Daily Chinese entity annotation data set, Boson entity annotation data set and microblog entity annotation data set.

[0082] First, the data set preprocessing method is introduced. In order to prevent the Bi-LSTM model from appearing gradient disappearance or decline due to too long input sequence, the application limits the length of all samples in the data set to 256 characters. The samples exceeding 256 characters are truncated and split into multiple samples. The samples less than 256 characters are padded with the label "O". The data set is divided into training set and test set according to the ratio of 4:1, wherein the training set contains 20864 samples and the test set contains 4636 samples.

[0083] Secondly, the simulation method of federated learning environment is introduced. First, the training set is classified. The specific method is as follows, the number of each entity category of each sentence sample is counted, the sentence is classified into the category with the most number of entities, and if the sentence does not appear entity, it is recorded as category "O". Secondly, the training set is sorted according to the classification result, and the training set except the "O" category is divided into n sub data sets as n different clients in equal order. Then, in order to simulate the inconsistent data size and low quality data set of the client, the "O" category training set is divided into n sub data sets, and each client selects a sub data set from the "O" category training set with a probability of 0.3 until the "O" category training set is empty. Finally, each client uses the local sub data set for model training and participates in federated aggregation.

[0084] The present application experiment is carried out under Ubuntu 16.04 system, using python 3.6 programming language for development. The hardware resources use 4 2.2GHz 10-core CPU, 12GB memory and GEFORCE RTX 2080 graphics card.

[0085] The present application uses precision, recall and F1 score as the evaluation indexes of entity extraction model. Define TP as the number of entities identified correctly by the model, FP as the number of entities with incorrect entity boundary or entity category identified by the model, and FN as the number of real entities not identified by the model. The formulas of the three evaluation indexes are as follows:

[0086]

[0087]

[0088]

[0089] 2. Experimental setting

[0090] This section explains the superiority of FedEI algorithm, and the selected baseline algorithm is FedAVG.

[0091] The federated learning parameter settings are as follows: the number of clients is set to 10; the proportion of selected clients per round is 0.2; the number of federated learning training rounds is 20. The model training parameter settings are as follows: the ALBERT model with base scale is adopted, Epoch is 2, BatchSize is 32, CRF layer learning rate is 10 -3 , and the neural network module learning rate is 10 -5 .

[0092] 3. Ablation experiment

[0093] The model training F1 value curves of the two algorithms are shown in Figure 3 . The model convergence speed of FedEI algorithm is faster than that of FedAVG algorithm, and the model F1 value reaches the inflection point at the 2nd round of federated communication. From the 11th round of federated communication, the model F1 value is close to convergence. The model convergence speed of FedAVG algorithm is slower, and it is close to convergence at the 14th round of federated communication. The final experimental results are shown in Table 2. Under 20 rounds of federated communication, the model F1 value of FedEI algorithm reaches 91.05%, which is about 1.35% higher than that of FedAVG algorithm. The reasons for the above experimental results are analyzed as follows.

[0094] In the model aggregation stage, the FedAVG algorithm only aggregates the models by weighting the data volume to update the global model, ignoring the influence of entity-labeled data on the performance of the model, resulting in that the weight of the model with excellent performance participating in the model aggregation is too low, further leading to the decrease of the accuracy of the model. The FedEI algorithm predicts the performance of the client model by evaluating the information entropy, entity quantity and entity dispersion of the entity-labeled data of the client, so as to allocate appropriate aggregation weight, so that the global model converges faster, and the accuracy and convergence speed of the model are improved.

[0095] Table 2 model training results of two algorithms

[0096]

[0097] Secondly, the communication cost of model training of the two algorithms is analyzed. The communication times required for the model F1 value to reach 88% are set, and the experimental results are shown in Table 2. Figure 4 The experimental results show that the model F1 value of the FedEI algorithm reaches 88% earliest, and only 11 rounds of federal communication are required, while the FedAVG needs 19 times of federal communication. It is proved that in order to reach the same model F1 value, the FedEI algorithm needs less federal communication than the baseline algorithm, the model training efficiency is high, and the training communication cost is reduced.

[0098] Further analysis of the total data transmission of a single client. In this experiment, the client is randomly selected to participate in federal learning training with a probability of 0.2 in each round of federal communication, so the total data transmission of different clients may not be consistent, therefore the average value of the total data transmission of all clients is obtained as the total data transmission of a single client, and the experimental results are shown in Table 3. The total data transmission of the FedAVG algorithm is 147.7MB, which is higher than that of the FedEI algorithm of 91.3MB. This is because the model training efficiency of the FedAVG algorithm is low, and more federal communication is required, resulting in more total data transmission uploaded by the client.

[0099] Table 3 data transmission of two algorithms

[0100]

[0101] In summary, in all evaluation module combinations, the performance of the FedEI algorithm is better than that of the baseline algorithm FedAVG. It is proved that the FedEI algorithm can improve the convergence speed of the model, reduce the federal communication rounds required for model training, and reduce the communication cost of training.

[0102] It should be noted that embodiments of the present application can be realized by hardware, software, or a combination of software and hardware. The hardware portion can be realized by a special logic; the software portion can be stored in a memory and executed by a proper instruction execution system, such as a microprocessor or a specially designed hardware. A person of ordinary skill in the art can understand that the above-mentioned apparatus and method can be realized by computer executable instructions and / or included in processor control codes, such as a carrier medium, such as a magnetic disk, CD or DVD-ROM, a programmable memory, such as a read-only memory (firmware), or a data carrier, such as an optical or electronic signal carrier. The apparatus of the present application and its modules can be realized by a hardware circuit, such as a very large scale integrated circuit or a gate array, a semiconductor, such as a logic chip, a transistor, or a programmable hardware device, such as a field programmable gate array, a programmable logic device, or the like, by software executed by various types of processors, or by a combination of the above-mentioned hardware circuit and software, such as firmware.

[0103] The above description is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any modification, equivalent replacement, and improvement within the technical range disclosed by the present application, and within the spirit and principle of the present application, should be included in the protection scope of the present application.

Claims

1. An entity extraction-oriented federated learning model aggregation method, characterized in that, The entity extraction-oriented federated learning model aggregation method comprises the following steps: For each client (0≤k≤|C t |) Calculate entity information e k The server calculates the weight vector α of the aggregated model based on the entity information E of all clients, where each component α k When representing model aggregation The weights; the server assigns weights to each client model based on α. Aggregate to obtain the global model The entity extraction-oriented federated learning model aggregation method comprises the following steps: Step one, calculate the information entropy, entity quantity and entity dispersion of the entity annotation dataset; Step two, calculate the information entropy, entity quantity and entity dispersion of the spliced text to obtain an entity information vector; Step three, normalize each of the three components of the entity information vector of each client; Step four, weighted average the vector obtained in step three to obtain the model aggregation weight of the client.

2. The federated learning model aggregation method for entity-oriented extraction according to claim 1, wherein, The calculation of the information entropy of the entity annotation dataset in step one comprises: (1) Traverse the dataset D k , count the occurrence of each word, and get the dictionary Q; (2) Calculate the probability of each character q in Q appearing in D to get vector R = [r1, r2,..., r |Q| ], where r i is the probability of character q i appearing; (3) calculating the text information entropy h of the data set D k of the data set D The calculation of the entity quantity and the entity dispersion comprises: Statistical data set D k The number of occurrences of each type of entity is obtained to obtain an entity category vector V. The entity quantity γ and the entity dispersion σ are calculated according to the following formula respectively:

3. The federated learning model aggregation method for entity-oriented extraction of claim 1, wherein, In the second step, the entity information vector [h, y, s] is obtained by splicing the text information entropy h, the entity number y and the entity dispersion degree s of the data set D k .

4. The federated learning model aggregation method for entity-oriented extraction of claim 1, wherein, In step three, C is calculated according to the following formula. t Each client entity information vector [h k ,γ k ,σ k The three components of ] are each normalized to obtain a vector.

5. The federated learning model aggregation method for entity-oriented extraction of claim 1, wherein, In step four, the vector is weighted averaged according to the following formula to obtain the model aggregation weight of the client k : 6.A federated learning model aggregation system for entity-oriented extraction applying the federated learning model aggregation method for entity-oriented extraction according to any one of claims 1 to 5. The entity extraction-oriented federated learning model aggregation system comprises: An entity information calculation module for calculating the information entropy, entity quantity and entity dispersion of the entity annotation dataset; An entity information vector calculation module for calculating the information entropy, entity quantity and entity dispersion of the spliced text to obtain an entity information vector; An entity information component normalization module for normalizing each of the three components of the entity information vector of each client; A model aggregation weight calculation module for weighted averaging the vector obtained by the component normalization module to obtain the model aggregation weight of the client.

7. A computer device, comprising: The computer device comprises a memory and a processor, and the memory stores a computer program, which, when executed by the processor, causes the processor to execute the steps of the entity extraction-oriented federated learning model aggregation method according to any one of claims 1-5.

8. A computer readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the entity extraction-oriented federated learning model aggregation method according to any one of claims 1-5.

9. An information data processing terminal, characterized by The information data processing terminal is used to implement the entity extraction-oriented federated learning model aggregation system according to claim 6.

Images