Assessment assisting method based on convolutional neural network recommendation algorithm

Abstract

The invention discloses an evaluation auxiliary method based on a convolutional neural network recommendation algorithm, relates to the technical field of computer application, and comprises a text similarity matching algorithm based on a convolutional neural network, an improved RankNet sorting technology and an evaluation conclusion auxiliary generation technology based on a template rule. According to the evaluation auxiliary method based on the convolutional neural network recommendation algorithm, under the condition of giving an evaluation problem, through semantic understanding modes such as word segmentation and named entity recognition on the problem, evaluation basis content most relevant to the problem is recommended, an expert can modify a generated evaluation conclusion according to the evaluation basis content, the evaluation efficiency is improved, and the evaluation accuracy is improved. The method can be applied to various evaluation auxiliary scenes, the method can be applied to an automatic evaluation system for real soldier confrontation training, aiming at a current automatic evaluation system, automatic evaluation can be carried out, a result can be generated, evaluation reference data can be given, experts can modify a generated evaluation conclusion, and the evaluation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer application technology, and in particular to an evaluation auxiliary method based on a convolutional neural network recommendation algorithm. Background Art

[0002] As an important way and means of pre-practice, simulation is widely used in conflict analysis, crisis management and other fields. Evaluation of the simulation process is a necessary step, which requires evaluation experts to conduct a comprehensive analysis of the scientificity and rationality of the simulation process and disposal, and requires a large amount of historical cases, computational models, laws and regulations and other knowledge data support. Due to the fast pace of simulation and limited on-site data, evaluation based solely on expert experience and knowledge will lead to subjective uncertainty in the evaluation results. How to use intelligent means to provide auxiliary support for evaluation experts through knowledge recommendation can effectively improve the scientific and refined level of simulation evaluation work.

[0003] In recent years, automated assessment systems have realized the automatic generation of assessment conclusions in specific simulation scenarios by pre-setting simulation plans, preparing assessment data, and setting assessment rules. For assessment tasks in general scenarios, the simulation situation changes rapidly, and the automated assessment system cannot pre-set complete assessment rules. It is necessary to conduct real-time assessment by combining various simulation data with actual conditions. At present, the full-scenario assessment capability of the adjudication system is weak and requires human intervention.

[0004] Such as the following two existing technical solutions:

[0005] 1. "Research on Active Recommendation Methods of Military Information Based on User Behavior Analysis" 2021: This paper studies how to improve the effectiveness of active recommendation of military information. At present, the total amount of information in military information systems is growing, and the content is complex. Information service systems need to obtain effective resources from massive amounts of information. This paper analyzes military needs and proposes a variety of information recommendation methods based on user long-term and short-term preference perception, user intention perception, etc., builds an active information recommendation prototype system for simulation experiments, and improves the quality of active recommendation services for military information.

[0006] 2. Research on Automated Judgement System Based on Live-action Confrontation Training, 2023: This paper proposes an automated judgement system to address the time-consuming and labor-intensive post-evaluation analysis of current live-action confrontation training. The system includes information collection subsystem, information management subsystem, judgement scheme design subsystem, and judgement analysis subsystem. The system implements functions such as judgement resource preparation, training data acquisition and analysis, real-time judgement result feedback, and judgement report generation, reducing manual intervention in the entire training process and improving judgement efficiency.

[0007] It can be seen from the above two existing technical solutions that the evaluation process involves multiple evaluation bases, resulting in a high complexity of the evaluation process. Traditional evaluation relies on expert experience and is highly subjective. The existing automated evaluation system mainly analyzes the evaluation task and sets rules based on equipment attributes and task types to achieve rapid evaluation generation. However, the current automatic evaluation system does not provide evaluation reference materials, making it difficult to assist manual evaluation analysis. The current evaluation data recommendation focuses on interest preference recommendations based on user behavior. Through user historical interaction behavior, user features are extracted based on the historical interaction matrix, and different information streams are recommended based on user features. However, this type of method does not perform semantic analysis on the search content and cannot solve the problem of analyzing user search conditions and making personalized recommendations. Therefore, an evaluation assistance method based on a convolutional neural network recommendation algorithm is proposed. Summary of the invention

[0008] In view of the deficiencies in the prior art, the present invention provides an evaluation assistance method based on a convolutional neural network recommendation algorithm, which solves the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: an evaluation auxiliary method based on a convolutional neural network recommendation algorithm, comprising the following steps:

[0010] Step 1: Maintain the basis data by evaluating the basis data management technology, build the basis data set, and store the basis data set in the basis database;

[0011] Step 2: During the evaluation, the user inputs the evaluation question, and the evaluation basis data recommendation technology is used to obtain the evaluation basis most relevant to the question, and the database provides data support for the data recommendation technology;

[0012] Step 3: Based on the existing data, generate auxiliary evaluation conclusions through auxiliary evaluation conclusion generation technology.

[0013] Optionally, the step 1: maintaining the basis data through evaluation basis data management technology, constructing a basis data set, and storing the basis data set in the process of the basis database. The evaluation basis data management technology used includes data sorting for data from different sources according to recommendation requirements, and the data sorting includes data compilation, data management, data addition, data viewing and subject management;

[0014] The data compilation integrates, expands, and stores various independent document data related to existing evaluation data, model data, and research reports in a unified manner. Since different document formats are different, the content of the documents can be extracted by OCR scanning to obtain unified text format data. Through the text summarization generation algorithm, the text content is analyzed to extract the summary information of the data files, and relevant information such as the text name, location, and role is supplemented to form complete compiled data. Association tags between the compiled data and the document data are established for the service retrieval recommendation system;

[0015] For the data management, the completed compiled data is uniformly maintained through data management;

[0016] For the data addition, the documents and the compiled data are entered into the retrieval information library, supporting batch addition;

[0017] For the data viewing, the compiled data and the source files can be viewed. For the document data that has changed, it can be edited or deleted;

[0018] For the theme management, the data is classified and labeled. The data themes can classify and query different data by theme, as well as manage functions such as adding, editing, and deleting themes.

[0019] Optionally, in step two: during the evaluation, the user enters an evaluation question, and the evaluation basis data recommendation technology is used to obtain the evaluation basis most relevant to the question. In the process where the database provides data support for the data recommendation technology, the evaluation basis data recommendation technology includes a text similarity matching algorithm based on a convolutional neural network and an improved RankNet sorting technology.

[0020] Optionally, for the text similarity matching algorithm based on a convolutional neural network, the current evaluation question and historical evaluation questions are used as inputs. Semantic understanding is used to deeply analyze the retrieval statements, extract the user's intentions and information needs, and the BERT pre-trained model is used to obtain the vectorized feature representation of the evaluation questions, which is then input into a convolutional neural network using a collaborative attention mechanism to learn the text features of the evaluation questions. The similarity between the text features of the two evaluation questions is calculated to complete the similarity matching of the two evaluation questions.

[0021] Optionally, for the improved RankNet sorting technology, the text matching results of the evaluation questions are sorted customarily, and the sorting result algorithm is returned. The goal of the RankNet algorithm is to train a relatively accurate sorting model. A probability loss function is proposed to learn sorting and sort the document set. For any pair of documents doc (U i ,U j ), the model output is a score of (s i ,s j) According to the prediction of the model, U i U is more j The probability that is more relevant to the query query is defined as:

[0022]

[0023] The typical RankNet neural network mainly consists of three parts: the input layer, the hidden layer, and the output layer. Its main feature is that there are connections between neurons in adjacent layers, but there are no connections between neurons in the same layer. When designing the structure of the neural network, the algorithm characteristics determine that the number of neurons in the input and output layers is usually a fixed value. In the RankNet algorithm, the number of neurons in the input layer is the dimension of the feature vector;

[0024] During the forward propagation process, data continuously passes from the input layer, is superimposed and then passes through the activation function to the hidden layer, and then from the hidden layer, is superimposed and passes through the activation function to reach the output layer. This process of performing logistic regression sequentially from the input layer to the output layer can be simply understood as the forward propagation of the neural network. After one iteration ends, the degree of inconsistency between the output data value and the true value is judged and enters the backpropagation;

[0025] Backpropagation: The result of the output layer is the predicted value of the model. The inconsistency between the predicted value and the true value is the error, that is, the loss function. Taking the mean squared error as an example, denoted as C, we get:

[0026]

[0027] After calculation, we can obtain:

[0028]

[0029] Among them, a i is the known input. In order to minimize the loss function, v and w are variables that the model needs to continuously correct. Therefore, the entire network is actually calculating the mapping relationship between v, w, and C. From the calculation process of the loss function C, it can be concluded that the loss function before the improvement of RankNet calculates the error by passing the difference between two output values to the cross-entropy loss function. Then, the RankNet sorting algorithm ignores the error of misjudgment of the two documents themselves in the same pair by the model, and this part of the error is not added to the loss function C. In this case, if the relevance of a single evaluation document in a pair is prejudged incorrectly, it will affect the sorting direction of the model;

[0030] The RankNet algorithm with an improved loss function considers both the prediction error of sorting and the error of each document in the pair. During training, all losses are used for average error calculation to replace the expected error:

[0031]

[0032] Optionally, in step three: based on existing data, through the technology of generating auxiliary evaluation conclusions, in the process of generating auxiliary evaluation conclusions based on template rules, according to the existing evaluation conclusion specifications and rules and user-defined evaluation rules, the user can select a template to generate an evaluation conclusion, including semantic extraction of the evaluation request, recommendation of evaluation data based on existing evaluation data, generation of evaluation conclusions, recommendation of evaluation data, generation of evaluation conclusions, and management of evaluation-generated data; for the recommendation of evaluation data, relevant evaluation basis data for evaluation is automatically screened. After a new evaluation request is added, relevant evaluation materials for this evaluation are screened from the evaluation basis dataset through the evaluation basis data recommendation method. The user can perform operations such as adding, editing, and deleting the evaluation materials according to the evaluation requirements.

[0033] Optionally, for the generation of evaluation conclusions, analyze the evaluation basis data, and based on the evaluation template rules and combined with historical evaluation conclusion data, automatically generate evaluation conclusions. Through text summary extraction, obtain the key content of the evaluation basis as the basic text for conclusion generation. Through text similarity matching, identify the most similar historical evaluation conclusion for this evaluation as the template structure for conclusion generation. Through fusion calculation, generate the evaluation conclusion for this evaluation. The user can modify based on the generated conclusion to improve the efficiency of generating evaluation conclusions.

[0034] Optionally, for the management of evaluation-generated data, maintain the evaluation problems, evaluation conclusions, and evaluation basis-related evaluation data. The evaluation application list can be displayed, showing the application content of the evaluation application, the red and blue parties, the status (pending processing, processed), the application time, and the operations (generate, view) and other related content.

[0035] The present invention provides an evaluation assistance method based on a convolutional neural network recommendation algorithm, having the following beneficial effects:

[0036] This evaluation assistance method based on a convolutional neural network recommendation algorithm focuses on, under the condition of a given evaluation problem, through semantic understanding methods such as word segmentation and named entity recognition of the problem, recommends the evaluation basis content most relevant to the problem. Experts can modify the generated evaluation conclusions based on the evaluation basis content to improve the evaluation efficiency. This method can be applied to various evaluation assistance scenarios. This method can be applied to the automated evaluation system for actual combat confrontation training. For the current automated evaluation system, it can automatically evaluate and generate results and give evaluation reference materials. Experts can modify the generated evaluation conclusions to improve the evaluation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the workflow diagram of the present invention;

[0038] Figure 2 It is a block diagram of the text similarity matching algorithm of the present invention;

[0039] Figure 3 This is a diagram of the automated judging and evaluation system based on real-life confrontation training of the present invention;

[0040] Figure 4 This is the intelligent auxiliary generation diagram of the present invention. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0042] See also Figures 1 to 2 The present invention provides a technical solution: an evaluation auxiliary method based on a convolutional neural network recommendation algorithm, comprising the following steps:

[0043] Step 1: Maintain the basis data through evaluation basis data management technology, collect and compile evaluation data, structure the evaluation data, label it, build a basis data set, store the basis data set in the basis database, and provide data support for data recommendation technology;

[0044] The evaluation is based on data management techniques, including data collation based on recommended requirements for data from different sources. Data collation includes data compilation, data management, data addition, data review and subject management;

[0045] Data compilation is to unify, expand and store all kinds of independent document data such as existing review data, model data, research reports, etc. Due to different document formats, the document content can be extracted through OCR scanning to obtain unified text format data. Through the text summary generation algorithm, the text content is analyzed, the summary information of the data file is extracted, and the text name, location, role and other related information are supplemented to form a complete compilation data. The association tags between the compilation data and the document data are formed for the service retrieval recommendation system;

[0046] Data management: unified maintenance of compiled data through data management;

[0047] Data addition: entering documents and compiled data into the search information base, supporting batch addition;

[0048] Data viewing: you can view the compiled data and source files, and edit or delete the changed document data;

[0049] Subject management, classifying and labeling data. Data subjects can classify and query different data by subject, as well as manage functions such as adding, editing, and deleting subjects;

[0050] Step 2: When the user conducts an evaluation question retrieval, semantic analysis is performed on the input evaluation question through text processing methods such as word segmentation and named entity recognition to extract the text feature vector of the evaluation question, perform text similarity matching calculations with historical evaluation questions, then perform sorting calculations on the matching calculation results, and use the evaluation basis data recommendation technology to obtain the evaluation basis most relevant to the question. The database provides data support for the data recommendation technology;

[0051] The evaluation basis data recommendation technology includes a text similarity matching algorithm based on a convolutional neural network and an improved RankNet sorting technology;

[0052] The text similarity matching algorithm based on a convolutional neural network takes the current evaluation question and historical evaluation questions as inputs, deeply analyzes the retrieval statement using semantic understanding, extracts the user's intention and information needs, obtains the vectorized feature representation of the evaluation question using the BERT pre-trained model, and inputs it into a convolutional neural network using a collaborative attention mechanism to learn the text features of the evaluation question, calculates the similarity of the text features of the two evaluation questions, and completes the similarity matching of the two evaluation questions;

[0053] Based on the improved RankNet sorting technology, custom sorting is performed on the text matching results of the evaluation questions, and the sorting result algorithm is returned, thereby improving the work efficiency of the arbitration users. The goal of the RankNet algorithm is to train a relatively accurate sorting model. The core is to propose a probability loss function to learn sorting and sort the document set. For any pair of documents doc (U i , U j ), the model outputs scores (s i , s j ). According to the model's prediction, the probability that U i is more relevant to the query query than U j is defined as:

[0054]

[0055] The typical RankNet neural network mainly consists of three parts: the input layer, the hidden layer, and the output layer. Its main feature is that there are connections between neurons in adjacent layers, but neurons in the same layer are not connected to each other. When designing the structure of the neural network, the algorithm characteristics determine that the number of neurons in the input and output layers is usually a fixed value. In the RankNet algorithm, the number of neurons in the input layer is the dimension of the feature vector;

[0056] During the forward propagation process, data continuously passes from the input layer, is stacked and then passes through the activation function to the hidden layer, and then from the hidden layer, is stacked and passes through the activation function to reach the output layer. This process of performing logistic regression sequentially from the input layer to the output layer can be simply understood as the forward propagation of the neural network. After one iteration, the degree of inconsistency between the output data value and the true value is judged to enter the backpropagation;

[0057] Backpropagation: The result of the output layer is the predicted value of the model. The inconsistency between the predicted value and the true value is the error, that is, the loss function. Taking the mean squared error as an example, denoted as C, we get:

[0058]

[0059] It can be calculated that:

[0060]

[0061] Among them, a i is the known input. In order to minimize the loss function, v and w are the variables that the model needs to continuously correct. Therefore, the entire network is actually calculating the mapping relationship between v, w, and C. From the calculation process of the loss function C, it can be obtained that the loss function before the improvement of RankNet is to pass the difference between the two output values to the cross-entropy loss function for error calculation. Then, the RankNet sorting algorithm ignores the error of misjudgment of the two documents themselves in the same pair by the model, and this part of the error is not added to the loss function C. In this case, if the relevance of a single evaluation document in a pair is misjudged in advance, it will affect the sorting direction of the model;

[0062] The RankNet algorithm with the improved loss function considers both the prediction error of sorting and the error of each document in the pair. During training, all losses are used for average error calculation to replace the expected error:

[0063]

[0064] The improved loss function consists of an empirical risk part and a regularization term part. The loss function calculation in the average error method considers more factors, and the regularization term is used to solve problems such as overfitting;

[0065] Step 3: Through text-assisted generation technology, recommend evaluation bases and generate evaluation conclusions;

[0066] The technology for generating auxiliary evaluation conclusions based on template rules can generate evaluation conclusions by selecting a template according to existing evaluation conclusion specifications and rules as well as user-defined evaluation rules. It includes semantic extraction of evaluation requests, recommendation and generation of evaluation conclusions based on existing evaluation data, recommendation of evaluation data, generation of evaluation conclusions, and management of evaluation-generated data. For the recommendation of evaluation data, it automatically filters the evaluation basis data relevant to the evaluation. After adding a new evaluation request, it screens the evaluation materials relevant to this evaluation from the evaluation basis dataset through the evaluation basis data recommendation method. Users can perform operations such as adding, editing, and deleting evaluation materials according to evaluation requirements.

[0067] For generating evaluation conclusions, it analyzes the evaluation basis data, combines historical evaluation conclusion data on the basis of the evaluation template rules, and automatically generates evaluation conclusions. Through text summary extraction, it obtains the key content of the evaluation basis as the basic text for conclusion generation. Through text similarity matching, it identifies the most similar historical evaluation conclusion for this evaluation as the template structure for conclusion generation. Through fusion calculation, it generates the evaluation conclusion for this time. Users can modify based on the generated conclusion to improve the efficiency of generating evaluation conclusions.

[0068] For the management of evaluation-generated data, it maintains the evaluation-related data such as evaluation questions, evaluation conclusions, and evaluation basis. It can display the evaluation application list, showing the application content, red and blue parties, status (pending, processed), application time, and operations (generate, view) of the evaluation applications.

[0069] Embodiment

[0070] Aiming at the problem that the post-event evaluation and analysis of actual combat confrontation training currently takes a lot of time and effort, the automated evaluation system formulates evaluation rules and generates evaluation results, reducing manual intervention in the entire training process and improving the evaluation efficiency. Taking the automated evaluation system based on actual combat training as an example, the system needs to formulate detailed evaluation rules. Although it can automatically generate an evaluation plan, the generated results fail to automatically and effectively provide evaluation reference materials. As shown in the appendix Figure 3 It is difficult to assist manual evaluation and analysis. The evaluation conclusion generation method based on template rules proposed by the present invention focuses on automatic evaluation generation and recommends relevant evaluation bases based on existing evaluation data. Experts can modify the generated evaluation conclusions to improve the evaluation efficiency. As shown in the appendix Figure 4 .

[0071] As mentioned above, it is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. An evaluation assistance method based on a convolutional neural network recommendation algorithm, characterized in that: The following steps are involved: Step 1: Maintain the basis data by evaluating the basis data management technology, build the basis data set, and store the basis data set in the basis database; Step 2: During the evaluation, the user inputs the evaluation question, and the evaluation basis data recommendation technology is used to obtain the evaluation basis most relevant to the question, and the database provides data support for the data recommendation technology; Step 3: Based on the existing data, generate auxiliary evaluation conclusions through auxiliary evaluation conclusion generation technology.

2. The evaluation assistance method based on the convolutional neural network recommendation algorithm according to claim 1, wherein: The step 1: maintaining the basis data through evaluation basis data management technology, constructing a basis data set, and storing the basis data set in the process of the basis database. The evaluation basis data management technology used in the process includes data sorting for data from different sources according to the recommendation requirements, and the data sorting includes data compilation, data management, data addition, data viewing and subject management; The data compilation is to unify, expand and store the existing independent document data related to the review data, model data and research reports. Due to the different document formats, the document content can be extracted by OCR scanning to obtain unified text format data. The text summary generation algorithm is used to analyze the text content, extract the summary information of the data file, supplement the text name, location, role and other related information, form a complete compilation data, and form the association tags between the compilation data and the document data for the service retrieval recommendation system. The data management is to uniformly maintain the compiled data through data management; The data addition includes entering documents and compiled data into the search information base, and supports batch addition; The data review can review the compiled data and source files, and can edit or delete the changed document data; The subject management classifies and labels the data. The data subject can classify and query different data by subject, as well as add, edit, and delete subject-related management functions.

3. An evaluation assistance method based on a convolutional neural network recommendation algorithm according to claim 1, characterized in that: The step 2: during the evaluation, the user inputs an evaluation question, and uses the evaluation basis data recommendation technology to obtain the evaluation basis most relevant to the question. The evaluation basis data recommendation technology in the process of providing data support for the data recommendation technology based on the database includes a text similarity matching algorithm based on a convolutional neural network and a sorting technology based on an improved RankNet.

4. An evaluation assistance method based on a convolutional neural network recommendation algorithm according to claim 1, characterized in that: The text similarity matching algorithm based on convolutional neural network takes the current evaluation question and the historical evaluation question as input, uses semantic understanding to perform in-depth analysis on the retrieval sentence, extracts user intention and information needs, and uses the BERT pre-trained model to obtain the vectorized feature representation of the evaluation question, and inputs it into the convolutional neural network using the collaborative attention mechanism to learn the text features of the evaluation question, calculates the similarity of the text features of the two evaluation questions, and completes the similarity matching of the two evaluation questions.

5. An evaluation assistance method based on a convolutional neural network recommendation algorithm according to claim 3, characterized in that: The improved RankNet-based sorting technology customizes the sorting for the text matching results of the evaluation problem and returns the sorting result algorithm. The goal of the RankNet algorithm is to train a relatively accurate sorting model, propose a probability loss function to learn sorting, and sort the document set. For any pair of documents doc (U i , U j ), the model outputs scores (s i , s j ). According to the model's prediction, the probability that U i is more relevant to the query than U j is defined as: The typical RankNet neural network mainly consists of three parts: the input layer, the hidden layer, and the output layer. Its main feature is that there are connections between neurons in adjacent layers, but there are no connections between neurons in the same layer. When designing the structure of the neural network, the algorithm characteristics determine that the number of neurons in the input and output layers is usually a fixed value. In the RankNet algorithm, the number of neurons in the input layer is the dimension of the feature vector; During the forward propagation process, data continuously passes from the input layer, is stacked and then passes through the activation function to the hidden layer, and then from the hidden layer, is stacked and passes through the activation function to reach the output layer. This process of performing logistic regression sequentially from the input layer to the output layer can be simply understood as the forward propagation of the neural network. After one iteration ends, the degree of inconsistency between the output data value and the true value is judged to enter the backpropagation; Backpropagation: The result of the output layer is the predicted value of the model. The inconsistency between the predicted value and the true value is the error, that is, the loss function. Taking the mean squared error as an example, denoted as C, we get: It can be calculated that: Among them, a i is the known input. To minimize the loss function, v and w are variables that the model needs to continuously correct. Therefore, the entire network is actually calculating the mapping relationship among v, w, and C. From the calculation process of the loss function C, it can be obtained that the loss function before the improvement of RankNet is to transfer the difference between the two output values to the cross-loss function for error calculation. Then, the RankNet sorting algorithm ignores the error of misjudgment of the two documents in the same pair by the model itself, and this part of the error is not added to the loss function C. In this case, if the relevance of a single evaluation document in a pair is prejudged incorrectly, it will affect the sorting direction of the model; The RankNet algorithm that improves the loss function considers both the prediction error of sorting and the error of each document in the pair. During training, all losses are used to calculate the average error to replace the expected error:

6. The evaluation assistance method based on the convolutional neural network recommendation algorithm according to claim 1, characterized in that: The third step: Based on the existing data, through the auxiliary evaluation conclusion generation technology, in the process of generating the auxiliary evaluation conclusion of the technology based on the template rules, according to the existing evaluation conclusion specifications and rules and the user-defined evaluation rules, the user can select a template to generate the evaluation conclusion, including semantic extraction of the evaluation request, recommendation and generation of evaluation conclusions based on the existing evaluation data, evaluation data recommendation, evaluation conclusion generation, and evaluation-generated data management; for the evaluation data recommendation, relevant evaluation basis data for the evaluation is automatically screened. After a new evaluation request is added, relevant evaluation materials for this evaluation are screened from the evaluation basis dataset through the evaluation basis data recommendation method. The user can perform operations such as adding, editing, and deleting the evaluation materials according to the evaluation requirements.

7. An evaluation assistance method based on a convolutional neural network recommendation algorithm according to claim 1, characterized in that: For generating the evaluation conclusion, analyze the evaluation basis data, and based on the evaluation template rules and combined with the historical evaluation conclusion data, automatically generate the evaluation conclusion. Through text summary extraction, obtain the key content of the evaluation basis as the basic text for conclusion generation. Through text similarity matching, identify the most similar historical evaluation conclusion for this evaluation as the template structure for conclusion generation. Through fusion calculation, generate the evaluation conclusion for this time. The user can modify based on the generated conclusion to improve the efficiency of evaluation conclusion generation.

8. An evaluation assistance method based on a convolutional neural network recommendation algorithm according to claim 6, characterized in that: For the evaluation-generated data management, maintain the evaluation problems, evaluation conclusions, and evaluation basis-related evaluation data. The evaluation application list can be displayed, showing the application content, red and blue parties, status (pending processing, processed), application time, operations (generate, view) and other related content of the evaluation application.

Images