Assessment opinion intelligent generation method and system based on multi-source data fusion

Abstract

The invention provides an assessment opinion intelligent generation method and system based on multi-source data fusion, and the method comprises the steps: obtaining a plurality of pieces of text evaluation information of an assessment object, each piece of text evaluation information corresponding to different evaluators, and the evaluators need to come from at least three different position types; performing feature extraction on all the text evaluation information, and performing clustering according to the extracted feature information to obtain a plurality of class clusters; and analyzing the plurality of class clusters, dividing the class clusters into a target class cluster and a boundary class cluster according to an analysis result, calculating an original comprehensive score of each assessment object according to the target class cluster and the boundary class cluster, and generating assessment suggestions of each assessment object according to the original comprehensive score. According to the method, a technical framework of multi-level semantic extraction and multi-granularity fusion analysis is constructed, intelligent assessment suggestions with quantitative scoring, grade evaluation and personalized improvement suggestions are automatically generated from multi-source text evaluation information, and the objectivity and efficiency of assessment work are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of assessment technology, and more specifically, to a method and system for intelligently generating assessment opinions based on multi-source data fusion. Background Technology

[0002] With the deepening of the refinement and digital transformation of human resource management, employee performance appraisal systems based on multi-source feedback are becoming increasingly popular. Traditional text-based evaluation methods, which rely primarily on manual reading, summarization, and scoring, have significant shortcomings: Highly subjective and inefficient: A large number of unstructured text evaluations require manual reading and understanding of each line, which is easily affected by the reviewer's personal experience, attention and emotions, resulting in subjective and unstable summarization results, and is time-consuming and labor-intensive.

[0003] Information integration is one-sided: Evaluations from different roles (superiors, peers, and subordinates) have different perspectives and expressions. Manual processing makes it difficult to systematically quantify, compare, and deeply integrate the information. It is easy to miss key information or get bogged down in details, making it difficult to form a comprehensive and balanced three-dimensional profile.

[0004] Difficulty in uncovering deep consensus and disagreement: Traditional methods struggle to automatically and accurately identify mainstream consensus viewpoints and ambiguous, contradictory, marginal opinions in massive amounts of text, and are unable to effectively distinguish and utilize these different types of information.

[0005] Insufficient decision support: It can usually only output qualitative summaries or simple quantitative scores, lacking deeper mining of evaluation data, unable to automatically locate shortcomings in capabilities, simulate the effect of improvement, and provide targeted data-driven suggestions for the personalized development of employees.

[0006] Therefore, there is an urgent need for a technical solution that can automatically and intelligently process multi-source text evaluation information to achieve deep semantic understanding, accurate opinion clustering, quantitative scoring, and diagnostic decision support. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for intelligently generating assessment opinions based on multi-source data fusion, so as to improve the above-mentioned problems.

[0008] To achieve the above objectives, the embodiments of this application provide the following technical solutions: On the one hand, embodiments of this application provide a method for intelligently generating assessment opinions based on multi-source data fusion, the method comprising: Obtain multiple text evaluation messages of the candidates being evaluated. Each text evaluation message corresponds to a different evaluator, and the evaluators must come from at least three different job types. Feature extraction is performed on all text evaluation information, and clustering is performed based on the extracted feature information to obtain multiple clusters; Multiple clusters are analyzed, and the clusters are divided into target clusters and boundary clusters based on the analysis results. The original comprehensive score of each examinee is calculated based on the target cluster and boundary cluster, and the assessment opinion of each examinee is generated based on the original comprehensive score.

[0009] Secondly, embodiments of this application provide an intelligent assessment opinion generation system based on multi-source data fusion, the system comprising: The acquisition module is used to acquire multiple text evaluation messages of the assessment object. Each text evaluation message corresponds to a different evaluator, and the evaluators must come from at least three different job types. The clustering module is used to extract features from all text evaluation information and cluster them based on the extracted features to obtain multiple clusters. The generation module is used to analyze multiple clusters, divide the clusters into target clusters and boundary clusters based on the analysis results, calculate the original comprehensive score for each examinee based on the target clusters and boundary clusters, and generate the assessment opinion for each examinee based on the original comprehensive score.

[0010] Thirdly, embodiments of this application provide an intelligent assessment opinion generation device based on multi-source data fusion, the device including a memory and a processor. The memory is used to store a computer program; the processor is used to execute the computer program to implement the steps of the above-described intelligent assessment opinion generation method based on multi-source data fusion.

[0011] Fourthly, embodiments of this application provide a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent generation method for assessment opinions based on multi-source data fusion.

[0012] The beneficial effects of this invention are as follows: 1. This invention effectively overcomes the limitations of a single model in text semantic understanding by constructing a parallel dual-branch feature extraction architecture and designing a multi-level feature interaction and fusion strategy. By combining and concatenating shallow fused features with deep local / global features, and finally pooling them into feature vectors, a more powerful and information-rich text feature representation is generated. Furthermore, surface text similarity based on edit distance and semantic similarity based on deep features are integrated as clustering metrics to improve clustering quality. Finally, through agglomerative hierarchical clustering algorithm, automated and high-precision topic sorting of multi-source, unstructured text evaluation information is achieved, laying a high-quality data foundation for subsequent analysis.

[0013] 2. This invention proposes an automatic boundary cluster identification mechanism. By quantitatively analyzing the internal consistency and external distinguishability of each cluster, it automatically filters out evaluation sets with loose internal opinions and ambiguous boundaries with mainstream opinions. The identified boundary clusters are not simply discarded, but effectively utilized. Specifically, by calculating the credibility coefficient of the boundary clusters and combining it with the correlation between the text and each feature dimension, an adjustment value for the ability evaluation of each capability dimension is calculated. This adjustment value is then used to correct the initial capability score based on mainstream consensus (target cluster). This mechanism ensures that the final comprehensive score is not only based on clear mainstream opinions, but also prudently and quantitatively incorporates the influence of marginal and ambiguous opinions, thereby significantly improving the robustness, noise resistance, and decision credibility of the scoring results, making the assessment results more reflective of complex realities.

[0014] 3. Building upon the generation of comprehensive scores and grades, this invention further provides in-depth diagnostic and analytical functions. When a capability deficiency is detected (such as a low score in any dimension or a poor score in a core dimension), an anomaly detection process is automatically triggered. This process uses pre-constructed positive text describing an ideal state to sequentially replace the actual evaluation datasets corresponding to each capability dimension, and simulates and calculates the improvement effect (improvement value) of the overall comprehensive score after the replacement. The capability dimension corresponding to the maximum improvement value is identified as the feature dimension label for key improvement. This method extends the assessment from simply "evaluating the past" to "guiding the future," clearly revealing which dimension's improvement will bring the greatest marginal benefit to overall performance. This provides clear capability development suggestions for both the assessed individuals and managers, enhancing the action guidance value of the assessment system.

[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the intelligent generation method for assessment opinions based on multi-source data fusion as described in this embodiment of the invention; Figure 2This is a schematic diagram of the intelligent assessment opinion generation system based on multi-source data fusion as described in this embodiment of the invention; Figure 3 This is a schematic diagram of the intelligent assessment opinion generation device based on multi-source data fusion as described in this embodiment of the invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] It should be noted that similar reference numerals or letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] Example 1

[0021] like Figure 1 As shown, this embodiment provides an intelligent method for generating assessment opinions based on multi-source data fusion, which includes steps S1, S2 and S3.

[0022] Step S1: Obtain multiple text evaluation messages of the assessment object. Each text evaluation message corresponds to a different evaluator, and the evaluators must come from at least three different job types. In this step, evaluators assess the candidate, generating text evaluation information. The requirement that evaluators come from at least three different job types means that all evaluators providing text evaluation information to the candidate must belong to at least three different job types after deduplication. For example, if there are five evaluators with corresponding job types of Director, Director, Deputy Director, Assistant, and Assistant, the duplicate job types "Director" and "Assistant" are each counted as one. Therefore, the total number of job types corresponding to the five evaluators is three (Director, Deputy Director, Assistant).

[0023] This step can be understood as follows: In the performance evaluation of an employee "Zhang San", anonymous text evaluation information from colleagues was obtained. For example: The first text evaluation message (from colleague 1, whose job type is department head): "He has a clear approach to the diagnosis and treatment of acute and critical illnesses, is calm and decisive during rescue, and has strong clinical decision-making ability." The second text evaluation message (from colleague 2, job type: attending physician): "Good teamwork spirit, willing to share experience during departmental consultations." The third text evaluation message (from colleague 3, job type: resident physician): "He is patient and meticulous in his teaching, often taking the initiative to guide us in standardized operations, and has a high reputation among young physicians." Step S2: Extract features from all text evaluation information and cluster them according to the extracted features to obtain multiple clusters; The specific implementation steps of this step include step S21; Step S21: Encode the text evaluation information into a word vector sequence; simultaneously and independently input the word vector sequence into the parallel first local feature extraction branch and the first global context feature extraction branch to extract the first local feature matrix and the first global context feature matrix respectively; add the first local feature matrix and the first global context feature matrix element by element to generate an intermediate fusion feature matrix; simultaneously input the intermediate fusion feature matrix into the second local feature extraction branch and the second global context feature extraction branch to extract the second deep local feature matrix and the second deep global context feature matrix respectively; calculate the feature vector corresponding to each text evaluation information based on the intermediate fusion feature matrix, the second deep local feature matrix, and the second deep global context feature matrix, and cluster all text evaluation information based on the feature vector corresponding to each text evaluation information to obtain multiple clusters.

[0024] In this step, conventional methods can be used to encode the text evaluation information into a word vector sequence, which will not be elaborated further. Simultaneously, the word vector sequence is input into the first local feature extraction branch (a one-dimensional convolutional neural network (CNN)) to obtain the first local feature matrix. The same word vector sequence is then input into the first global context feature extraction branch (a Transformer encoder) to obtain the first global context feature matrix. The intermediate fused feature matrix is ​​then simultaneously input into the second local feature extraction branch (a one-dimensional convolutional neural network (CNN)) and the second global context feature extraction branch (a Transformer encoder) to extract the second deep local feature matrix and the second deep global context feature matrix, respectively. The structures of the two one-dimensional convolutional neural networks can be the same or different. For example, the second one-dimensional convolutional neural network may have a wider kernel width; for instance, if the kernel width of the first one-dimensional convolutional neural network is 3, then the kernel width of the second one-dimensional convolutional neural network may be 5 or 7. The two Transformer encoders can also have the same or different structures; for example, the second Transformer encoder may have more attention heads. In this step, a dual-branch architecture combining CNN and Transformer was constructed to extract multi-level, multi-granular semantic features from text evaluation information. The CNN branch excels at capturing local phrase patterns and keyword collocations, while the Transformer branch excels at modeling long-range dependencies and global context. This synergistic and complementary architecture yields a richer and more discriminative feature matrix than a single model, laying a high-quality feature foundation for subsequent accurate clustering.

[0025] In addition, in this step, the feature vector corresponding to each text evaluation information is calculated based on the intermediate fusion feature matrix, the second deep local feature matrix, and the second deep global context feature matrix. The specific implementation steps for clustering all text evaluation information to obtain multiple clusters are steps S211 and S212. Step S211: Calculate the similarity between pairs of text evaluation information, wherein the pairs of text evaluation information are denoted as the first text evaluation information and the second text evaluation information, respectively. Calculate the edit distance between the first text evaluation information and the second text evaluation information, and simultaneously count the string lengths of the first text evaluation information and the second text evaluation information. Divide the edit distance by the maximum value between the two string lengths to obtain the first value. The difference between the preset second value and the first value is taken as the surface text similarity between the pairs of text evaluation information. In this step, the second value is preset to 1; the surface text similarity is obtained by subtracting the first value from 1; the edit distance between the first text evaluation information and the second text evaluation information can be calculated using the Levenstein distance algorithm known in the art.

[0026] Step S212: Concatenate the first local feature matrix, the first global context feature matrix, and the second deep local feature matrix along the feature dimension to obtain the first combined feature; concatenate the first local feature matrix, the first global context feature matrix, and the second deep global context feature matrix along the feature dimension to obtain the second combined feature; concatenate the first combined feature and the second combined feature along the feature dimension, and perform global average pooling on the concatenation result to generate the feature vector corresponding to each text evaluation information; calculate the cosine similarity between the feature vectors corresponding to the first text evaluation information and the second text evaluation information as the deep semantic similarity; weight the sum of the surface text similarity and the deep semantic similarity to obtain the similarity between pairs of text evaluation information; use agglomerative hierarchical clustering algorithm to cluster the text evaluation information, where the inter-cluster similarity is defined as the average of the similarities of all text evaluation information pairs between two clusters; set a merging threshold, and iteratively merge cluster pairs whose inter-cluster similarity is greater than or equal to the merging threshold; stop when there is no inter-cluster similarity greater than or equal to the merging threshold, and output the final clusters.

[0027] In this step, concatenation along the feature dimension means connecting multiple matrices along the feature dimension (number of columns) while keeping the sequence length (number of rows) unchanged, forming a new comprehensive feature matrix with a wider feature dimension. The weight of surface text similarity can be 0.3, and the weight of deep semantic similarity can be 0.7. An agglomerative hierarchical clustering algorithm is used, which takes each text as an initial cluster and iteratively merges two clusters with a similarity greater than or equal to a preset merging threshold. The similarity between clusters is defined as the average of the similarities of all text evaluation information pairs (composed by taking one text evaluation information from each of the two clusters) between the two clusters. When no cluster pair meets the merging condition, the iteration stops and the clusters are output.

[0028] Steps S211 and S212 involve weighted fusion of surface text similarity (edit distance) and deep semantic similarity (deep feature cosine similarity) to serve as a clustering metric. This avoids the problems of relying solely on deep learning models being overly sensitive to differences in expression, or relying solely on literal matching failing to understand semantics. Ultimately, multiple clusters are obtained through agglomerative hierarchical clustering, achieving automated processing of text evaluation information.

[0029] Step S3: Analyze multiple clusters, divide the clusters into target clusters and boundary clusters based on the analysis results, calculate the original comprehensive score for each subject based on the target clusters and boundary clusters, and generate assessment opinions for each subject based on the original comprehensive scores.

[0030] The specific implementation steps of this step include step S31 and step S32; Step S31: For each cluster, calculate the first similarity between any two text evaluation messages in that cluster, and average all first similarities to obtain the first average similarity corresponding to that cluster; calculate the second similarity between each text evaluation message in that cluster and each text evaluation message in other clusters, and average all second similarities corresponding to each text evaluation message to obtain the second average similarity corresponding to each text evaluation message; average all second average similarities corresponding to all text evaluation messages in that cluster to obtain the third average similarity corresponding to that cluster; determine whether the first average similarity is less than a preset first similarity threshold and whether the third average similarity is greater than a preset second similarity threshold. If both are true, then that cluster is considered a boundary cluster; the first similarity threshold is greater than the second similarity threshold. In this step, the first similarity between pairwise text evaluation information in the cluster can be calculated according to the methods in steps S211 and S212 above; if the first average similarity is less than the preset first similarity threshold and the third average similarity is greater than the preset second similarity threshold, then the cluster is regarded as a boundary cluster. Boundary clusters represent groups with internal disagreements and indistinguishable characteristics from external clusters, potentially containing ambiguous, neutral, or contradictory comments. Traditional clustering typically treats all clusters equally. This invention automatically filters out ambiguous evaluation sets that exhibit internal disagreements and are not easily distinguishable from mainstream opinions. Identifying boundary clusters has a dual significance: first, it distinguishes them from target clusters representing mainstream consensus, ensuring the representativeness of core scores; second, it provides input for subsequent steps to refine scores using this edge information, increasing the system's decision-making depth.

[0031] Step S32: Record the clusters that are not boundary clusters as target clusters, calculate the original comprehensive score for each examinee based on the target clusters and boundary clusters, and generate the assessment opinion for each examinee based on the original comprehensive ability score.

[0032] The specific implementation steps of this step include step S321 and step S322; Step S321: Obtain the feature dimension labels assigned to the target clusters based on the text evaluation information contained in each target cluster. Collect all the text evaluation information corresponding to each feature dimension label to form a feature dataset corresponding to each feature dimension label. At the same time, count the number of text evaluation information corresponding to each feature dimension label and take the feature dimension label with the largest number of text evaluation information as the core dimension label. Input each text evaluation information in each feature dataset into the preset ability evaluation model and output the ability score corresponding to each text evaluation information. Calculate the average of all ability scores to obtain the initial ability score corresponding to each feature dataset. Use the boundary clusters to correct the initial ability score corresponding to each feature dimension label to obtain the dimension comprehensive score corresponding to each feature dimension label. Sum the dimension comprehensive scores by weight to obtain the original comprehensive score. Determine the assessment level of the assessment object based on the original comprehensive score. In this step, feature dimension labels are assigned manually. For example, feature dimension labels can be assigned based on the high-frequency keywords contained in each target cluster. For example, if the high-frequency keyword in target cluster A is "team collaboration", then "team collaboration" can be used as the feature dimension label for this target cluster. All text evaluation information corresponding to each feature dimension label is also all text evaluation information contained in the target cluster corresponding to this feature dimension label. In this step, the method for generating the competency evaluation model can be as follows: obtain multiple historical text evaluation information from the past, label each historical text evaluation information with a competency score, and use the labeled information as a sample to train the convolutional neural network model to obtain the competency evaluation model. In this step, the weighted sum of the dimensional comprehensive scores is used to obtain the raw comprehensive score. The assessment level of the candidate is then determined based on this raw comprehensive score. The weight of each dimensional comprehensive score, i.e., the weight of each feature dimension label, can be customized as needed. Simultaneously, a mapping table between raw comprehensive scores and assessment levels can be preset in the system, where one raw comprehensive score range corresponds to one assessment level. After calculating the raw comprehensive score, the corresponding assessment level can be obtained by referring to the mapping table stored in the system. Meanwhile, in this step, the initial capability score corresponding to each feature dimension label is corrected by using boundary clusters to obtain the comprehensive dimension score corresponding to each feature dimension label. The specific implementation steps include steps S3211 and S3212. Step S3211: Obtain multiple keywords and their corresponding weight values ​​for each feature dimension label, where the weight values ​​reflect the importance of the keywords in the corresponding feature dimension label; This step can be understood as follows: For example, for the feature dimension label "team collaboration", the keyword weight dictionary can be preset as follows: {"collaboration": 0.35, "cooperation": 0.25, "communication": 0.20, "support": 0.15, "cooperation": 0.05}, where the sum of all weight values ​​is 1; Step S3212: For each text evaluation message in each boundary cluster, perform the following operations: perform word segmentation on the text evaluation message to obtain the word segmentation result; for each feature dimension label, traverse the keyword set corresponding to that feature dimension label; if the word segmentation result of the text evaluation message contains keywords, accumulate the weight values ​​corresponding to the keywords to obtain the weight contribution value of the text evaluation message to that feature dimension label; sum the weight contribution values ​​of the text evaluation message to each feature dimension label to obtain the total weight value of the text evaluation message; divide the weight contribution value corresponding to each feature dimension label by the total weight value to obtain the relevance between the text evaluation message and that feature dimension label; calculate the dimension comprehensive score corresponding to each feature dimension label based on the relevance between the text evaluation message and each feature dimension label in the boundary cluster.

[0033] In this step, for each feature dimension label, the keyword set corresponding to that feature dimension label is traversed. If the word segmentation result of the text evaluation information contains keywords, the weight values ​​corresponding to the keywords are accumulated to obtain the weight contribution value of the text evaluation information to that feature dimension label. This can be understood as follows: if the word segmentation result of a text evaluation information contains keywords 1 and 2 corresponding to a certain feature dimension label, the weight values ​​corresponding to keywords 1 and 2 are added together to obtain the weight contribution value of the text evaluation information to that feature dimension label. Meanwhile, in this step, the specific implementation steps for calculating the comprehensive score of each feature dimension label based on the correlation between the text evaluation information in the boundary cluster and each feature dimension label include steps S32121, S32122 and S32123. Step S32121: For each boundary cluster, divide the number of text evaluation information entries contained in the boundary cluster by the total number of text evaluation information entries of the assessment object to obtain the first proportion A; obtain the ideal proportion B, and calculate the quantity reduction factor P. The process involves counting the number of text evaluation messages corresponding to the same job type in the boundary cluster, obtaining the number of evaluation messages corresponding to different job types, calculating the ratio between the number of evaluation messages corresponding to different job types and the total number of text evaluation messages in the boundary cluster, and obtaining the second proportion of different job types. This second proportion is then multiplied by its logarithm to obtain the first product. All first products corresponding to the second proportion are summed, and the result is inversely represented to obtain the information entropy. The total number of job types for all evaluators in the boundary cluster is obtained, and the logarithm of the total number of job types is calculated to obtain the first logarithm. The information entropy is divided by the first logarithm to obtain the source balance factor. Finally, the quantity reduction factor is multiplied by the source balance factor to obtain the credibility coefficient of the boundary cluster. In this step, the ideal percentage can be customized, representing the reasonable proportion of boundary-based, ambiguous, or contradictory text evaluation information in all text evaluation information under ideal conditions. For example, it can be set to 10% to 20%. Counting the number of text evaluation information entries corresponding to the same job type evaluator in the boundary cluster can be understood as follows: for example, if there are three job types, namely director, deputy director, and assistant, then count the number of text evaluation information entries corresponding to evaluators with the job types of director / deputy director / assistant respectively. In this step, the second proportion is multiplied by its logarithm to obtain the first product. The first products corresponding to all second proportions are summed, and the result is inversely represented to obtain the information entropy. The total number of job types for all evaluators in the boundary cluster is obtained, and the logarithm of the total number of job types is calculated to obtain the first logarithm. The information entropy is divided by the first logarithm to obtain the source balance factor, which can be understood as: Assuming the second proportion is r, then The sum is then inversely represented; if the sum is V, then the information entropy is -V. The total number of job types for all evaluators in the boundary cluster is the number of unique job types held by all evaluators in the boundary cluster. For example, if the boundary cluster contains four evaluators with the corresponding positions of Director, Head, Deputy Head, and Assistant, then the total number of job types is 3. If the total number of job types for evaluators in the boundary cluster is n, then the first logarithm is... ; Step S32122: For each boundary cluster, calculate its capability evaluation adjustment value for each feature dimension label. Specifically, collect the text evaluation information in the boundary cluster whose relevance to the feature dimension label is greater than a preset relevance threshold to obtain a text evaluation information set. Input each text evaluation information in the text evaluation information set into the capability evaluation model. Calculate the mean of the obtained capability scores to obtain the boundary dimension score. Multiply the boundary dimension score by the credibility coefficient of the boundary cluster to obtain the capability evaluation adjustment value of the boundary cluster for the feature dimension label. Step S32123: Add up the capability evaluation adjustment value and the initial capability score for each feature dimension label to obtain the comprehensive dimension score corresponding to each feature dimension label.

[0034] In this step, after calculating the capability evaluation adjustment value for each boundary cluster for each feature dimension label, for each feature dimension label, all corresponding capability evaluation adjustment values ​​are added to the initial capability score to obtain the dimension comprehensive score. Step S322: Set Rule 1 and Rule 2. Rule 1 is that in all the comprehensive scores of all dimensions, at least one dimension's comprehensive score is less than the preset first ability score threshold. Rule 2 is that the comprehensive score of the dimension corresponding to the core dimension label is less than the preset second ability score threshold. Judge all the comprehensive scores of the dimensions. When any rule is met, perform anomaly detection to determine the feature dimension label that needs to be improved. Finally, output the original comprehensive score, assessment level, and feature dimension label that needs to be improved as assessment opinions. If no anomaly detection is performed, directly output the original comprehensive score and assessment level as assessment opinions. The first ability score threshold is less than the second ability score threshold.

[0035] In step S322, the specific implementation steps for anomaly detection to determine the feature dimension labels for key improvements include steps S3223-S3225; Step S3223: Obtain multiple positive text evaluation information corresponding to each feature dimension label, and form replacement text evaluation information corresponding to each feature dimension label; In this step, positive text evaluation information refers to standard text snippets used to describe specific behaviors, performances, or traits when a certain feature dimension label reaches an excellent or ideal state. For example, if the feature dimension label is teamwork, then the positive text evaluation information would be: Positive text evaluation information 1: "Demonstrated excellent teamwork in cross-departmental diagnosis and treatment, proactively communicated patient conditions and treatment plans, and actively cooperated with various departments to complete joint treatment." Positive text evaluation information 2: "Willing to cooperate with the team's work arrangements, actively support colleagues in daily work, and actively cooperate to complete various medical tasks." Positive text evaluation information 3: "In teamwork, she emphasizes communication and collaboration, effectively cooperates with colleagues from different professional backgrounds, provides professional support in complex case discussions, and promotes teamwork." Step S3224: For each feature dimension label: each time, select one corresponding replacement text evaluation information as its corresponding feature dataset, while the feature datasets corresponding to other feature dimension labels remain unchanged; after replacement, still input each feature dataset into the preset ability evaluation model, output the ability score corresponding to each feature dataset, and sum the ability scores corresponding to each feature dataset by weight to obtain the comprehensive score; after calculating the comprehensive score corresponding to each replacement feature, calculate the average of all comprehensive scores to obtain the average comprehensive score of each feature dimension label; subtract the initial ability score from the average comprehensive score of each feature dimension label to obtain the improvement value; In this step, the specific calculation method for "after replacement, each feature dataset is still input into the preset ability evaluation model, and the ability score corresponding to each feature dataset is output" is the same as the method in step S322, that is: input each text evaluation information in each feature dataset into the preset ability evaluation model, output the ability score corresponding to each text evaluation information, and calculate the average of all ability scores to obtain the ability score corresponding to each feature dataset. Step S3225: Select the feature dimension label corresponding to the maximum improvement value as the feature dimension label for key improvement.

[0036] Example 2 like Figure 2 As shown in the figure, this embodiment provides an intelligent assessment opinion generation system based on multi-source data fusion. The system includes an acquisition module 1, a clustering module 2, and a generation module 3.

[0037] Module 1 is used to obtain multiple text evaluation information of the assessment object. Each text evaluation information corresponds to a different evaluator, and the evaluators must come from at least three different job types. Clustering module 2 is used to extract features from all text evaluation information and cluster them according to the extracted features to obtain multiple clusters; Module 3 is used to analyze multiple clusters, divide the clusters into target clusters and boundary clusters based on the analysis results, calculate the original comprehensive score for each subject based on the target clusters and boundary clusters, and generate assessment opinions for each subject based on the original comprehensive scores.

[0038] In one specific embodiment of this disclosure, the clustering module 2 further includes an extraction unit 21.

[0039] Extraction unit 21 is used to encode text evaluation information into a word vector sequence; input the word vector sequence simultaneously and independently into the parallel first local feature extraction branch and the first global context feature extraction branch to extract the first local feature matrix and the first global context feature matrix respectively; add the first local feature matrix and the first global context feature matrix element by element to generate an intermediate fusion feature matrix; input the intermediate fusion feature matrix simultaneously into the second local feature extraction branch and the second global context feature extraction branch to extract the second deep local feature matrix and the second deep global context feature matrix respectively; calculate the feature vector corresponding to each text evaluation information based on the intermediate fusion feature matrix, the second deep local feature matrix and the second deep global context feature matrix, and cluster all text evaluation information based on the feature vector corresponding to each text evaluation information to obtain multiple clusters.

[0040] In one specific embodiment of this disclosure, the extraction unit 21 further includes a first calculation unit 211 and a clustering unit 212.

[0041] The first calculation unit 211 is used to calculate the similarity between pairs of text evaluation information, wherein the pairs of text evaluation information are respectively denoted as first text evaluation information and second text evaluation information, the edit distance between the first text evaluation information and the second text evaluation information is calculated, and the string lengths of the first text evaluation information and the second text evaluation information are respectively counted. The edit distance is divided by the maximum value between the two string lengths to obtain a first value, and the difference between the preset second value and the first value is used as the surface text similarity between the pairs of text evaluation information. Clustering unit 212 is used to concatenate the first local feature matrix, the first global context feature matrix, and the second deep local feature matrix along the feature dimension to obtain the first combined feature; concatenate the first local feature matrix, the first global context feature matrix, and the second deep global context feature matrix along the feature dimension to obtain the second combined feature; concatenate the first combined feature and the second combined feature along the feature dimension, and perform global average pooling on the concatenation result to generate the feature vector corresponding to each text evaluation information; calculate the cosine similarity between the feature vectors corresponding to the first text evaluation information and the second text evaluation information as the deep semantic similarity; weighted sum the surface text similarity and the deep semantic similarity to obtain the similarity between pairs of text evaluation information; use agglomerative hierarchical clustering algorithm to cluster the text evaluation information, where the inter-cluster similarity is defined as the average of the similarities of all text evaluation information pairs between two clusters; set a merging threshold, and iteratively merge cluster pairs with inter-cluster similarity greater than or equal to the merging threshold; stop when there is no inter-cluster similarity greater than or equal to the merging threshold, and output the final clusters.

[0042] In one specific embodiment of this disclosure, the generation module 3 further includes a second calculation unit 31 and a generation unit 32.

[0043] The second calculation unit 31 is used to calculate, for each cluster, the first similarity between any two text evaluation information in that cluster, and to calculate the average of all first similarities to obtain the first average similarity corresponding to that cluster; calculate the second similarity between each text evaluation information in that cluster and each text evaluation information in other clusters, and to calculate the average of all second similarities corresponding to each text evaluation information to obtain the second average similarity corresponding to each text evaluation information; and to calculate the average of all second average similarities corresponding to all text evaluation information in that cluster to obtain the third average similarity corresponding to that cluster; determine whether the first average similarity is less than a preset first similarity threshold and whether the third average similarity is greater than a preset second similarity threshold; if both are true, then that cluster is considered a boundary cluster; the first similarity threshold is greater than the second similarity threshold. The generation unit 32 is used to mark non-boundary clusters as target clusters, calculate the original comprehensive score of each examinee based on the target clusters and boundary clusters, and generate the assessment opinion of each examinee based on the original comprehensive ability score.

[0044] In one specific embodiment of this disclosure, the generation unit 32 further includes an acquisition unit 321 and an output unit 322.

[0045] The acquisition unit 321 is used to acquire feature dimension labels assigned to target clusters based on the text evaluation information contained in each target cluster, collect all text evaluation information corresponding to each feature dimension label to form a feature dataset corresponding to each feature dimension label; simultaneously, count the number of text evaluation information corresponding to each feature dimension label, and take the feature dimension label with the largest number of text evaluation information as the core dimension label; input each text evaluation information in each feature dataset into a preset ability evaluation model, output the ability score corresponding to each text evaluation information, calculate the average of all ability scores to obtain the initial ability score corresponding to each feature dataset; use boundary clusters to correct the initial ability score corresponding to each feature dimension label to obtain the dimension comprehensive score corresponding to each feature dimension label; perform weighted summation of the dimension comprehensive scores to obtain the original comprehensive score, and determine the assessment level of the assessment object based on the original comprehensive score. Output unit 322 is used to set rule one and rule two. Rule one is that in all the comprehensive scores of all dimensions, at least one dimension's comprehensive score is less than a preset first ability score threshold. Rule two is that the comprehensive score of the dimension corresponding to the core dimension label is less than a preset second ability score threshold. The comprehensive scores of all dimensions are judged. When any rule is met, anomaly detection is performed to determine the feature dimension label that needs to be improved. Finally, the original comprehensive score, assessment level, and feature dimension label that needs to be improved are output as assessment opinions. If no anomaly detection is performed, the original comprehensive score and assessment level are directly output as assessment opinions. The first ability score threshold is less than the second ability score threshold.

[0046] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0047] Example 3 Corresponding to the above method embodiments, this disclosure also provides an intelligent assessment opinion generation device based on multi-source data fusion. The intelligent assessment opinion generation device based on multi-source data fusion described below and the intelligent assessment opinion generation method based on multi-source data fusion described above can be referred to in correspondence with each other.

[0048] Figure 3 This is a block diagram illustrating an intelligent assessment feedback generation device 300 based on multi-source data fusion, according to an exemplary embodiment. Figure 3 As shown, the intelligent assessment feedback generation device 300 based on multi-source data fusion may include: a processor 301 and a memory 302. The intelligent assessment feedback generation device 300 may also include one or more of the following: a multimedia component 303, an I / O interface 304, and a communication component 305.

[0049] The processor 301 controls the overall operation of the intelligent assessment opinion generation device 300 based on multi-source data fusion to complete all or part of the steps in the aforementioned intelligent assessment opinion generation method based on multi-source data fusion. The memory 302 stores various types of data to support the operation of the intelligent assessment opinion generation device 300 based on multi-source data fusion. This data may include, for example, instructions for any application or method operating on the intelligent assessment opinion generation device 300 based on multi-source data fusion, as well as application-related data, such as contact data, sent and received messages, images, audio, video, etc. The memory 302 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 303 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 302 or transmitted via the communication component 305. The audio component also includes at least one speaker for outputting audio signals. I / O interface 304 provides an interface between processor 301 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 305 is used for wired or wireless communication between the multi-source data fusion-based intelligent assessment opinion generation device 300 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 305 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0050] In an exemplary embodiment, the intelligent assessment opinion generation device 300 based on multi-source data fusion may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned intelligent assessment opinion generation method based on multi-source data fusion.

[0051] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the above-described intelligent assessment opinion generation method based on multi-source data fusion. For example, the computer-readable storage medium may be the memory 302 including the program instructions, which may be executed by the processor 301 of the intelligent assessment opinion generation device 300 based on multi-source data fusion to complete the above-described intelligent assessment opinion generation method based on multi-source data fusion.

[0052] Example 4 Corresponding to the above method embodiments, this disclosure also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the above-described intelligent generation method for assessment opinions based on multi-source data fusion.

[0053] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the intelligent generation method for assessment opinions based on multi-source data fusion described in the above method embodiments.

[0054] The readable storage medium can specifically be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligently generating assessment opinions based on multi-source data fusion, characterized in that, include: Obtain multiple text evaluation messages of the candidates being evaluated. Each text evaluation message corresponds to a different evaluator, and the evaluators must come from at least three different job types. Feature extraction is performed on all text evaluation information, and clustering is performed based on the extracted feature information to obtain multiple clusters; Multiple clusters are analyzed, and the clusters are divided into target clusters and boundary clusters based on the analysis results. The original comprehensive score of each examinee is calculated based on the target cluster and boundary cluster, and the assessment opinion of each examinee is generated based on the original comprehensive score.

2. The intelligent generation method for assessment opinions based on multi-source data fusion according to claim 1, characterized in that, Feature extraction is performed on all text evaluation information, and clustering is performed based on the extracted feature information to obtain multiple clusters, including: The text evaluation information is encoded into a sequence of word vectors. The word vector sequence is simultaneously and independently input into the first local feature extraction branch and the first global context feature extraction branch in parallel to extract the first local feature matrix and the first global context feature matrix, respectively. The first local feature matrix and the first global context feature matrix are added element-wise to generate an intermediate fusion feature matrix. The intermediate fusion feature matrix is ​​simultaneously input into the second local feature extraction branch and the second global context feature extraction branch to extract the second deep local feature matrix and the second deep global context feature matrix, respectively. The feature vector corresponding to each text evaluation information is calculated based on the intermediate fusion feature matrix, the second deep local feature matrix, and the second deep global context feature matrix. All text evaluation information is clustered based on the feature vector corresponding to each text evaluation information to obtain multiple clusters.

3. The intelligent generation method for assessment opinions based on multi-source data fusion according to claim 2, characterized in that, Based on the intermediate fusion feature matrix, the second deep local feature matrix, and the second deep global context feature matrix, the feature vector corresponding to each text evaluation message is calculated. Then, based on the feature vectors corresponding to each text evaluation message, all text evaluation messages are clustered to obtain multiple clusters, including: Calculate the similarity between pairs of text evaluation information, where each pair of text evaluation information is denoted as the first text evaluation information and the second text evaluation information. Calculate the edit distance between the first text evaluation information and the second text evaluation information. At the same time, count the string lengths of the first text evaluation information and the second text evaluation information respectively. Divide the edit distance by the maximum value between the two string lengths to obtain the first value. The difference between the preset second value and the first value is used as the surface text similarity between the pairs of text evaluation information. The first local feature matrix, the first global context feature matrix, and the second deep local feature matrix are concatenated along the feature dimension to obtain the first combined feature. The first local feature matrix, the first global context feature matrix, and the second deep global context feature matrix are then concatenated along the feature dimension to obtain the second combined feature. The first combined feature and the second combined feature are concatenated along the feature dimension, and global average pooling is performed on the concatenation result to generate a feature vector corresponding to each text evaluation information. The cosine similarity between the feature vectors corresponding to the first and second text evaluation information is calculated as the deep semantic similarity. The surface text similarity and the deep semantic similarity are weighted and summed to obtain the similarity between pairs of text evaluation information. A cohesive hierarchical clustering algorithm is used to cluster the text evaluation information, where the inter-cluster similarity is defined as the average of the similarities of all text evaluation information pairs between two clusters. A merging threshold is set, and cluster pairs with inter-cluster similarity greater than or equal to the merging threshold are iteratively merged. The process stops when no inter-cluster similarity greater than or equal to the merging threshold exists, and the final clusters are output.

4. The intelligent generation method for assessment opinions based on multi-source data fusion according to claim 1, characterized in that, Multiple clusters are analyzed, and based on the analysis results, the clusters are divided into target clusters and boundary clusters. An initial comprehensive score is calculated for each examinee based on the target and boundary clusters. An assessment opinion is then generated for each examinee based on the initial comprehensive score, including: For each cluster, calculate the first similarity between any two text evaluation messages in that cluster, and average all first similarities to obtain the first average similarity for that cluster. Calculate the second similarity between each text evaluation message in that cluster and each text evaluation message in other clusters, and average all second similarities for each text evaluation message to obtain the second average similarity for each text evaluation message. Average all second average similarities for all text evaluation messages in that cluster to obtain the third average similarity for that cluster. Determine whether the first average similarity is less than a preset first similarity threshold and whether the third average similarity is greater than a preset second similarity threshold. If both are true, then that cluster is considered a boundary cluster. The first similarity threshold is greater than the second similarity threshold. Clusters that are not boundary clusters are designated as target clusters. The original comprehensive score for each examinee is calculated based on the target clusters and boundary clusters, and an assessment opinion for each examinee is generated based on the original comprehensive ability score.

5. The intelligent generation method for assessment opinions based on multi-source data fusion according to claim 4, characterized in that, Calculate the original comprehensive score for each examinee based on the target cluster and boundary cluster, and generate an assessment opinion for each examinee based on the original comprehensive ability score, including: The process involves: acquiring feature dimension labels for each target cluster based on the text evaluation information contained within it; collecting all text evaluation information corresponding to each feature dimension label to form a feature dataset for each feature dimension label; counting the number of text evaluation information entries corresponding to each feature dimension label and using the feature dimension label with the largest number of entries as the core dimension label; inputting each text evaluation information entry from each feature dataset into a pre-defined ability evaluation model to output the ability score corresponding to each text evaluation information; calculating the average of all ability scores to obtain the initial ability score for each feature dataset; refining the initial ability score for each feature dimension label using boundary clusters to obtain a comprehensive dimension score for each feature dimension label; weighting and summing the comprehensive dimension scores to obtain the original comprehensive score; and determining the assessment level of the examinee based on the original comprehensive score. Rule 1 and Rule 2 are set. Rule 1 requires that at least one dimension's overall score is lower than a preset first capability score threshold. Rule 2 requires that the overall score of the dimension corresponding to the core dimension label is lower than a preset second capability score threshold. All dimension scores are evaluated. If either rule is met, anomaly detection is performed to determine the key feature dimension labels for improvement. Finally, the original overall score, assessment level, and key feature dimension labels for improvement are output as assessment comments. If no anomaly detection is performed, the original overall score and assessment level are directly output as assessment comments. The first capability score threshold must be lower than the second capability score threshold.

6. An intelligent assessment opinion generation system based on multi-source data fusion, characterized in that, include: The acquisition module is used to acquire multiple text evaluation messages of the assessment object. Each text evaluation message corresponds to a different evaluator, and the evaluators must come from at least three different job types. The clustering module is used to extract features from all text evaluation information and cluster them based on the extracted features to obtain multiple clusters. The generation module is used to analyze multiple clusters, divide the clusters into target clusters and boundary clusters based on the analysis results, calculate the original comprehensive score for each examinee based on the target clusters and boundary clusters, and generate the assessment opinion for each examinee based on the original comprehensive score.

7. The intelligent assessment opinion generation system based on multi-source data fusion according to claim 6, characterized in that, The clustering module includes: An extraction unit is used to encode text evaluation information into a sequence of word vectors; the word vector sequence is simultaneously and independently input into a first local feature extraction branch and a first global context feature extraction branch in parallel to extract a first local feature matrix and a first global context feature matrix, respectively; the first local feature matrix and the first global context feature matrix are added element-wise to generate an intermediate fusion feature matrix; the intermediate fusion feature matrix is ​​simultaneously input into a second local feature extraction branch and a second global context feature extraction branch to extract a second deep local feature matrix and a second deep global context feature matrix, respectively; the feature vector corresponding to each text evaluation information is calculated based on the intermediate fusion feature matrix, the second deep local feature matrix, and the second deep global context feature matrix, and all text evaluation information is clustered based on the feature vector corresponding to each text evaluation information to obtain multiple clusters.

8. The intelligent assessment opinion generation system based on multi-source data fusion according to claim 7, characterized in that, Extraction unit, including: The first calculation unit is used to calculate the similarity between pairs of text evaluation information. The pairs of text evaluation information are denoted as the first text evaluation information and the second text evaluation information, respectively. The edit distance between the first text evaluation information and the second text evaluation information is calculated. At the same time, the string lengths of the first text evaluation information and the second text evaluation information are counted respectively. The edit distance is divided by the maximum value between the two string lengths to obtain the first value. The difference between the preset second value and the first value is used as the surface text similarity between the pairs of text evaluation information. The clustering unit is used to concatenate the first local feature matrix, the first global context feature matrix, and the second deep local feature matrix along the feature dimension to obtain the first combined feature; concatenate the first local feature matrix, the first global context feature matrix, and the second deep global context feature matrix along the feature dimension to obtain the second combined feature; concatenate the first combined feature and the second combined feature along the feature dimension, and perform global average pooling on the concatenation result to generate the feature vector corresponding to each text evaluation information; calculate the cosine similarity between the feature vectors corresponding to the first text evaluation information and the second text evaluation information as the deep semantic similarity; weighted sum the surface text similarity and the deep semantic similarity to obtain the similarity between pairs of text evaluation information; use agglomerative hierarchical clustering algorithm to cluster the text evaluation information, where the inter-cluster similarity is defined as the average of the similarities of all text evaluation information pairs between two clusters; set a merging threshold, and iteratively merge cluster pairs with inter-cluster similarity greater than or equal to the merging threshold; stop when there is no inter-cluster similarity greater than or equal to the merging threshold, and output the final clusters.

9. The intelligent assessment opinion generation system based on multi-source data fusion according to claim 6, characterized in that, The generation module includes: The second calculation unit is used to calculate the first similarity between any two text evaluation messages in each cluster, and to calculate the average of all first similarities to obtain the first average similarity corresponding to the cluster; calculate the second similarity between each text evaluation message in the cluster and each text evaluation message in other clusters, and to calculate the average of all second similarities corresponding to each text evaluation message to obtain the second average similarity corresponding to each text evaluation message; and to calculate the average of all second average similarities corresponding to all text evaluation messages in the cluster to obtain the third average similarity corresponding to the cluster; determine whether the first average similarity is less than a preset first similarity threshold and whether the third average similarity is greater than a preset second similarity threshold; if both are true, then this cluster is considered a boundary cluster; the first similarity threshold is greater than the second similarity threshold. The generation unit is used to designate non-boundary clusters as target clusters, calculate the original comprehensive score for each examinee based on the target clusters and boundary clusters, and generate an assessment opinion for each examinee based on the original comprehensive ability score.

10. The intelligent assessment opinion generation system based on multi-source data fusion according to claim 9, characterized in that, The generation unit includes: The acquisition unit is used to acquire feature dimension labels assigned to target clusters based on the text evaluation information contained in each target cluster. It collects all text evaluation information corresponding to each feature dimension label to form a feature dataset for each feature dimension label. Simultaneously, it counts the number of text evaluation information entries corresponding to each feature dimension label and uses the feature dimension label with the largest number of entries as the core dimension label. Each text evaluation information entry in each feature dataset is input into a preset ability evaluation model, outputting an ability score corresponding to each text evaluation information. The average of all ability scores is calculated to obtain the initial ability score for each feature dataset. The initial ability score for each feature dimension label is corrected using boundary clusters to obtain a comprehensive dimension score for each feature dimension label. The comprehensive dimension scores are then weighted and summed to obtain the original comprehensive score. The assessment level of the assessment subject is determined based on the original comprehensive score. The output unit is used to set Rule 1 and Rule 2. Rule 1 is that in all the comprehensive scores of all dimensions, at least one dimension's comprehensive score is less than a preset first ability score threshold. Rule 2 is that the comprehensive score of the dimension corresponding to the core dimension label is less than a preset second ability score threshold. All the comprehensive scores of the dimensions are judged. When any rule is met, anomaly detection is performed to determine the feature dimension labels that need to be improved. Finally, the original comprehensive score, assessment level, and feature dimension labels that need to be improved are output as assessment opinions. If no anomaly detection is performed, the original comprehensive score and assessment level are directly output as assessment opinions. The first ability score threshold is less than the second ability score threshold.

Images