A data analysis method and analysis system based on deep learning

Abstract

The application relates to the technical field of deep learning, and particularly discloses a data analysis method and system based on deep learning, which identifies and marks evaluation ability words from performance comments, extracts behavior subject words, behavior verbs and behavior object words to form behavior triplets, matches the behavior triplets with a capability-related behavior mode library, and calculates behavior evidence strength values. Standard behavior triplets of the mode library are derived from structured operation logs, and the binding relationship with the capability words is established according to a position capability model or a post competency dictionary independent of the performance comments, so as to cut off the pollution of subjective favor to the behavior evidence. Through semantic alignment judgment of the evaluation ability words and target capability words, the capability labels are divided into labels supported by comments, labels without behavior evidence support or labels with behavior evidence without comments, so as to distinguish objective behavior evidence from subjective evaluation projection, and find implicit capabilities not mentioned in the comments.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and specifically to a data analysis method and system based on deep learning. Background Technology

[0002] In corporate talent management, expert selection, and employment decisions, the automatic extraction of competency tags from performance reviews using deep learning has become a common practice. These methods typically train natural language processing models to identify competency terms such as leadership and systems thinking from unstructured comments written by superiors, thereby creating talent competency heatmaps.

[0003] However, the generation mechanism of performance review texts has inherent semantic entanglement. Superiors tend to use highly praising competency terms for employees with a good overall impression, even if there is a lack of specific behavioral facts to support it; while for those with a less favorable impression, even if successful behaviors are described, devaluation terms or competency labels may be used or omitted. Existing models only extract competency terms from the surface of the text and cannot distinguish whether the competency terms originate from the actual behavior of the person being evaluated or from the subjective projection of the evaluator's favorability. As a result, the final output competency heatmap is essentially very close to the "superior favorability heatmap," which seriously interferes with the objectivity of talent identification, expert evaluation, and staffing. Summary of the Invention

[0004] The purpose of this invention is to provide a data analysis method and system based on deep learning to solve the above-mentioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions: A data analysis method based on deep learning includes the following steps: A standardized dataset is obtained by preprocessing historical performance review texts; A deep learning model was constructed and trained using a standardized dataset. The trained deep learning model was used to identify and label evaluative competency words in performance review texts. Input the performance evaluation text A to be analyzed into the trained deep learning model to obtain the performance evaluation text A labeled with evaluative ability words; Extract the subject words, verb words, and object words from the performance evaluation text A, and combine them into behavior triples. The behavior triples are input into the capability-related behavior pattern library for matching, and the behavioral evidence strength value of the behavior triples for the target capability words is calculated. The target capability words are the capability words in the capability-related behavior pattern library that are pre-bound to the standard behavior triples that have been successfully matched. Generate a capability tag for the target capability word, perform semantic alignment determination between the evaluative capability word and the target capability word, and mark the capability tag according to the determination result.

[0006] As a further aspect of the present invention, the process of obtaining the standardized dataset is as follows: Retrieve historical performance appraisal texts, which are derived from structured performance appraisal records stored in the human resources system; The structured performance appraisal records are parsed to extract the appraisal text field and the appraisee identification field. The extracted appraisal text is then merged according to the appraisee identification field. The merged appraisal text is then segmented into sentences, deduplicated, and punctuation and non-visible characters are removed to obtain a standardized dataset.

[0007] As a further aspect of the present invention: the training process of the deep learning model is as follows: Obtain the set of ability words, which stores multiple preset evaluative ability words; The text samples in the standard dataset are segmented into words to obtain a text word sequence. Words in the text word sequence that appear in the ability word set are labeled as positive examples, and words in the text word sequence that do not appear in the ability word set are labeled as negative examples, resulting in a labeled word sequence. The combination of the text word sequence and the labeled word sequence is used as a training sample. A sequence labeling deep learning model is constructed. The sequence labeling deep learning model adopts a combination architecture of bidirectional long short-term memory network and conditional random field. The text word sequence in the training sample is used as input and the labeled word sequence in the training sample is used as output target. The sequence labeling deep learning model is trained using multiple training samples to obtain the trained deep learning model.

[0008] As a further aspect of the present invention: the process of obtaining the behavioral triplet is as follows: The performance evaluation text A consists of multiple sentences. Verb recognition is performed on each sentence to locate the core verb in each sentence. Call the verb semantic framework library to obtain the semantic framework corresponding to the core verb. The semantic framework defines the required semantic role types for the core verb. Extract specific words that fill the agent role from the sentence as the subject of the action, extract specific words that fill the patient role as the object of the action, and use the core verb as the action verb; The subject word, verb word, and object word extracted from the same sentence are combined into a behavior triplet. This process is repeated for all sentences in the performance review text A to generate a sequence of behavior triplets consisting of multiple behavior triplets.

[0009] As a further aspect of the present invention: the process of calculating the strength value of behavioral evidence is as follows: Obtain a pre-built capability-related behavior pattern library. Each record in the capability-related behavior pattern library contains a standard behavior triple and at least one capability word identifier pre-associated with the standard behavior triple. Replace the subject word in the behavior triple with a preset fixed role placeholder to obtain a role-normalized behavior triple; Map the action verbs in the normalized action triplet and the standard action verbs in the standard action triplet to the distributed semantic vector space, and calculate the first cosine value of the action verb vector and the standard action verb vector. Map the behavior object words in the normalized behavior triplet and the standard behavior object words in the standard behavior triplet to the distributed semantic vector space respectively, and calculate the second cosine value of the behavior object vector and the standard behavior object vector. The first cosine value and the second cosine value are arithmetically averaged, and the result is used as the behavior matching degree value. Pairings with behavior matching degree values ​​exceeding the preset matching threshold are judged as successful matches. The ability word identifiers bound to the standard behavior triples that are successfully matched are extracted as target ability words. All behavior matching degree values ​​pointing to the same target ability word are accumulated, and the accumulated result is used as the behavior evidence strength value for the target ability word.

[0010] As a further aspect of the present invention: the process of constructing a capability-related behavior pattern library is as follows: Retrieve structured operation logs automatically recorded by the system for all personnel within the same organization when completing predetermined tasks. The structured operation logs include operator identifier field, operation action identifier field, and operation target identifier field. Replace the operator identifier field with a preset fixed role placeholder, fill the replacement result into the standard behavior subject word slot, extract the operation action name corresponding to the operation action identifier field and fill it into the standard behavior verb slot, extract the operation target name corresponding to the operation target identifier field and fill it into the standard behavior object word slot, combine them into standard behavior triples, merge standard behavior triples that have the same structure, and keep one copy of each standard behavior triple that has the same structure. Obtain competency definition standards independent of performance evaluations. The competency definition standards include an expert-developed competency behavior anchoring table. For each standard behavior triple, find the anchor behavior description in the competency definition standards that matches the standard behavior verb and standard behavior object in that standard behavior triple. Extract the competency word corresponding to the anchor behavior description as a competency word identifier and bind it to the standard behavior triple. Collect all the bound standard behavior triples and their corresponding competency word identifiers to build a competency-related behavior pattern library.

[0011] As a further aspect of the present invention, the process of marking capability tags is as follows: Obtain the target ability words and their corresponding behavioral evidence strength values, and generate an ability label for each target ability word. The ability label contains the target ability word and its corresponding behavioral evidence strength value. Obtain all the marked evaluative competency words in the performance evaluation text A. For each evaluative competency word, determine whether it is exactly the same as a certain target competency word. If they are the same, mark the competency tag corresponding to the target competency word as having comment support and associate the evaluative competency word with the competency tag. For each evaluative ability word, if it is different from all target ability words, then the evaluative ability word is mapped to the distributed semantic vector space to obtain the evaluative ability vector, and each target ability word is mapped to the distributed semantic vector space to obtain the target ability vector. The cosine values ​​of the evaluative ability vector and each target ability vector are calculated respectively. When the maximum cosine value exceeds the preset synonym judgment threshold, the ability tag corresponding to the target ability word with the maximum cosine value is marked as having commentary support, and the evaluative ability word is associated with the ability tag; when the maximum cosine value does not exceed the preset synonym judgment threshold, the evaluative ability word is marked as having no behavioral evidence support. After iterating through all evaluative competency terms and completing the assessment, all competency tags that were not marked as having behavioral evidence but no evaluative support are marked as having behavioral evidence but no evaluative support.

[0012] A deep learning-based data analysis system includes: AI module: Preprocesses historical performance review texts to obtain a standardized dataset; A deep learning model was constructed and trained using a standardized dataset. The trained deep learning model was used to identify and label evaluative competency words in performance review texts. Input the performance evaluation text A to be analyzed into the trained deep learning model to obtain the performance evaluation text A labeled with evaluative ability words; Analysis module: Extract the subject words, verb words, and object words from the performance evaluation text A, and combine the subject words, verb words, and object words into behavior triples; The behavior triples are input into the capability-related behavior pattern library for matching, and the behavioral evidence strength value of the behavior triples for the target capability words is calculated. The target capability words are the capability words in the capability-related behavior pattern library that are pre-bound to the standard behavior triples that have been successfully matched. Tagging module: Generates a capability tag for the target capability word, performs semantic alignment determination between the evaluative capability word and the target capability word, and tags the capability tag according to the determination result.

[0013] The beneficial effects of this invention compared to the prior art are as follows: This solution constructs a competency-related behavioral pattern library independent of performance reviews. It establishes the binding relationship between standard behavioral triples and competency terms on job competency models, job competency dictionaries, or expert-developed competency-behavioral anchoring tables, rather than relying on the performance reviews of the evaluated individual. This eliminates the contamination of behavioral evidence standards by subjective evaluations. Simultaneously, it independently extracts behavioral triples from performance reviews, matches them with the competency-related behavioral pattern library, calculates the strength of behavioral evidence, and then uses semantic alignment to hierarchically associate evaluative competency terms with target competency terms. The final output includes differentiated competency labels in three states: "supported by comments," "no behavioral evidence," and "behavioral evidence without comments." This effectively distinguishes between genuine competencies supported by objective behavior and projective evaluations lacking behavioral evidence. Furthermore, it can discover implicit competencies not explicitly mentioned in comments but reflected in behavior, providing more objective and comprehensive data support for industry talent assessment, expert selection, and employment decisions. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating a data analysis method based on deep learning according to the present invention. Figure 2 This is a schematic diagram of the process for calculating the strength value of behavioral evidence according to the present invention. Detailed Implementation

[0016] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-2 As shown, this invention is a data analysis method based on deep learning, comprising the following steps: A standardized dataset is obtained by preprocessing historical performance review texts; In a preferred embodiment of the present invention, the process of obtaining a standardized dataset is as follows: Extract the comment text field and the appraisee identifier field from the structured performance appraisal records in the human resources system. Each row of the structured performance appraisal record corresponds to one comment, containing three fields: employee ID, appraisal period, and comment content. The extraction operation reads the comment content field from each row as the comment text and the employee ID field as the appraisee identifier. The two are temporarily stored as key-value pairs, with the appraisee identifier as the key and the comment text as the value.

[0018] After extracting all rows, a merge operation is performed based on the evaluated object's identifier. All key-value pairs are scanned, and multiple comments corresponding to the same evaluated object identifier are concatenated into a single text paragraph, separated by periods. For example, if employee E001 has three comments during the evaluation period: "High communication efficiency this quarter," "Coordinated with the marketing department to complete product launch," and "Timely response to customer needs," the merged text will be: "High communication efficiency this quarter. Coordinated with the marketing department to complete product launch. Timely response to customer needs."

[0019] The merged comment text is processed sequentially through sentence segmentation, deduplication, and removal of punctuation and invisible characters. Sentence segmentation uses periods, question marks, and exclamation marks as delimiters to split the text into independent sentences, with each sentence stored as an entry in a sentence list. Deduplication retains only one copy of identical sentences from the sentence list, deleting the rest. Punctuation removal iterates through each character in each sentence, replacing commas, semicolons, colons, and quotation marks with empty characters. Periods are already used as delimiters during sentence segmentation and require no further processing. Removal of invisible characters replaces newline characters, carriage returns, and tabs with empty characters. The resulting sentence list constitutes the normalized dataset.

[0020] A deep learning model was constructed and trained using a standardized dataset. The trained deep learning model was used to identify and label evaluative competency words in performance review texts. Input the performance evaluation text A to be analyzed into the trained deep learning model to obtain the performance evaluation text A labeled with evaluative ability words; In a preferred embodiment of the present invention, the training process of the deep learning model is as follows: Obtain a set of competency terms, which is stored in a text file. Each line records one evaluative competency term, such as "communication skills", "project management skills", "execution skills", "innovative thinking", "stress resistance", etc. The competency term set is constructed based on the job competency model and job competency dictionary, and is reviewed and approved by human resource management experts before the implementation of the plan.

[0021] Each text sample in the standardized dataset is segmented into a word sequence. The text sample is a processed sentence from the standardized dataset, such as "Coordinate the marketing department to complete the product launch". The segmentation operation uses a general Chinese word segmentation tool to divide the text sample into a sequence of words. The example sentence above, after segmentation, yields the word sequence "coordinate", "marketing department", "complete", "product", and "launch". The order of each word in this word sequence remains unchanged from its position in the original sentence.

[0022] Each word in the text word sequence is matched against the set of competency words one by one. The matching process iterates through each word in the text word sequence, checking whether the word appears completely in any line of the competency word set. Taking the text word sequence "coordinate," "marketing department," "complete," "product," and "release" as an example, the matching results are: "coordinate" is not in the competency word set, "marketing department" is not in the competency word set, "complete" is not in the competency word set, "product" is not in the competency word set, and "release" is not in the competency word set. Taking another text word sequence "demonstrated," "strong," "communication ability," "coordinate," "multiple departments," "complete," and "customer launch" as an example, the matching results are: "demonstrated" is not in the competency word set, "strong" is not in the competency word set, "communication ability" exists in the competency word set, "coordinate" is not in the competency word set, "multiple departments" is not in the competency word set, "complete" is not in the competency word set, and "customer launch" is not in the competency word set.

[0023] For words in the text word sequence that appear in the ability word set, these words are labeled as positive examples, using the marker B-EVAL to indicate evaluative ability words. For words in the text word sequence that do not appear in the ability word set, these words are labeled as negative examples, using the marker 0 to indicate non-evaluative words. The labeled word sequence generated using the second example above is 0, 0, B-EVAL, 0, 0, 0, 0. The position of each label in the labeled word sequence corresponds one-to-one with the position of the corresponding word in the text word sequence, and they have the same length.

[0024] A combination of a text word sequence and a labeled word sequence is used as a training sample. Each training sample contains two parallel sequences: the first is the text word sequence, and the second is the labeled word sequence. The above word segmentation, matching, and labeling operations are performed on all text samples in the standardized dataset one by one to generate multiple training samples. All training samples constitute the training sample set. The first 80% of the training sample set, arranged chronologically, is used as the actual training data, and the last 20% is used as the validation data.

[0025] A sequence labeling deep learning model is constructed, employing a combined architecture of a bidirectional long short-term memory network (LSTM) and a conditional random field (CRF), comprising a three-layer structure. The first layer is the embedding layer, which maps each word in the text word sequence to a 128-dimensional floating-point vector. The embedding layer parameters are initialized using pre-trained Chinese word vectors before training. The second layer is the bidirectional LSTM layer, which contains two sub-layers: a forward LSTM network and a backward LSTM network. Each sub-layer has 256 hidden units. The forward sub-layer processes the word vector sequence from left to right, and the backward sub-layer processes it from right to left. At each word position, the hidden outputs of the forward and backward sub-layers are concatenated to form a 512-dimensional context representation vector. The third layer is the CRF layer, which receives the context representation vector sequence output from the bidirectional LSTM layer and outputs a label sequence. When predicting the label of the current word, the CRF layer considers the transition probabilities between the labels of adjacent words. The output label types include B-EVAL and O.

[0026] The training process uses real training data to iterate the model multiple times. In each iteration, the real training data is divided into batches of 32 training samples each. After each batch is input into the model, the negative log-likelihood loss of the conditional random field layer is calculated. The Adam optimizer is used to update all model parameters, with the learning rate set to 0.001. After each iteration, the sequence labeling accuracy of the model on the validation data is calculated. Training stops when the model's labeling accuracy on the validation data no longer improves after 5 consecutive iterations. The model parameters at the time of stopping training are saved, resulting in the trained deep learning model.

[0027] The input to the trained deep learning model is a new performance review text. Before input, the same word segmentation operation as during training is performed on the performance review text to obtain a sequence of words to be predicted. The model outputs a predicted label for each word in the sequence of words to be predicted, which is either B-EVAL or O. Words with the predicted label B-EVAL are labeled as evaluative competency words in the performance review text. For example, if the input performance review text is "possesses strong project management skills", and the model outputs a label sequence of O, O, O, B-EVAL after word segmentation, then "project management skills" in the text is labeled as an evaluative competency word.

[0028] It should be noted that this sequence labeling deep learning model formalizes the task of identifying evaluative competency words into a word-by-word classification problem. The essence of this is that the position of competency words in performance evaluations is not fixed and may be embedded in any part of the sentence as independent terms. Therefore, using a sequence labeling architecture to independently predict each word is a direct means of achieving word-level labeling.

[0029] The embedding layer maps words to 128-dimensional floating-point vectors and loads pre-trained Chinese word vectors. The pre-trained word vectors contain general semantic priors, enabling the model to capture semantic relationships between words even with a limited training sample size. For example, it can distinguish the distribution differences in semantic space between action verbs like "coordinate" and ability descriptors like "communication ability".

[0030] The bidirectional long short-term memory network layer serves as the feature encoder. The forward sub-layer processes the word vector sequence from left to right and the backward sub-layer processes it from right to left. Each word position is fused with the bidirectional context information to form a 512-dimensional representation. This design allows the judgment of whether the current word is an evaluative ability word to refer to the semantic environment of the entire sentence, avoiding misjudgments based solely on the isolated form of the word.

[0031] The Conditional Random Field (CRF) layer receives a sequence of context representations and outputs a sequence of labels. When predicting the current word label, it explicitly models the transition dependencies between adjacent labels, thereby suppressing the occurrence of irrational label sequences, such as avoiding disjointed and isolated competency word annotation fragments in adjacent positions. During the training phase, positive and negative example labels are automatically generated using an expert-approved competency word set. This competency word set itself originates from the job competency model and the job competency dictionary, and the annotation is based on performance reviews independent of any single individual, ensuring the consistency of the training objectives and avoiding mislabeling competency words as subjective language preferences of the review writers.

[0032] Training employs the Adam optimizer with a learning rate of 0.001 for parameter updates, terminating training when the validation accuracy fails to improve for several consecutive rounds. This prevents the model from overfitting to specific expression patterns in the training samples, thereby enhancing its generalization ability to recognize competency words in unseen comments. During actual inference, the trained model performs word-by-word prediction on the input comments, and the position of the output B-EVAL label represents the location of the evaluative competency word. This achieves automated and standardized extraction of evaluative competency words, providing clean evaluative word input for subsequent behavior triple extraction and behavioral evidence strength calculation. This eliminates the need for manual review in competency label generation, and the recognition process is mechanistically isolated from the subjective rhetoric in the comments.

[0033] Extract the subject words, verb words, and object words from the performance evaluation text A, and combine them into behavior triples. In a preferred embodiment of the present invention, the process of obtaining behavioral triples is as follows: Verb identification is performed on each sentence in the performance evaluation text A, which consists of multiple sentences, each of which has been segmented into independent entries during the preprocessing stage. Verb identification uses a Chinese part-of-speech tagging tool to tag each word in the sentence. Words tagged as verbs are marked as candidate verbs. The rule for locating the core verb among multiple candidate verbs is to select the verb that serves as the root node or subject-predicate verb in the syntactic dependency analysis results. If there is only one candidate verb in the sentence, it is directly identified as the core verb. For example, in the sentence "Zhang San coordinated the R&D department to complete the customer's online launch," after part-of-speech tagging, "coordinate" is tagged as a verb, and "complete" is tagged as a verb. Syntactic dependency analysis shows that "coordinate" is the root node verb of the sentence; therefore, "coordinate" is located as the core verb.

[0034] The semantic framework for the core verb is retrieved by calling the verb semantic framework library. This library is a pre-built verb framework mapping table that uses the morphology of the core verb as the lookup key and stores the semantic framework definitions activated by that verb. Each semantic framework defines the required semantic role types for the core verb, with each semantic role type corresponding to an abstract identity of the action participant. Taking the core verb "coordination" as an example, the semantic framework corresponding to coordination is retrieved from the library. This framework defines the agent and patient roles as required semantic roles, where the agent role represents the subject performing the coordination action, and the patient role represents the object or matter being coordinated. If no corresponding record for the core verb is found in the verb semantic framework library, the sentence is skipped and no action triple is generated.

[0035] The process extracts specific words that fill the agent role from the sentence as the subject words and specific words that fill the patient role as the object words. The extraction process uses a semantic role labeling tool, centered on the core verb, to classify the semantic roles of each phrase component in the sentence. The core nouns of phrases that assume the agent role are extracted as the subject words, and phrases that assume the patient role are extracted as the object words. Taking the sentence "Zhang San coordinated the R&D department to complete the customer launch" as an example, the semantic role labeling results show that Zhang San assumes the agent role, and the customer launch assumes the patient role. Therefore, Zhang San is extracted as the subject word, the customer launch as the object word, and the core verb "coordinate" as the action verb.

[0036] The subject word, verb word, and object word extracted from the same sentence are combined into an action triple. The combination method involves filling the three words into the corresponding slots of a pre-defined triple structure: the subject word in the first slot, the verb word in the second slot, and the object word in the third slot, forming a triple representation of "subject word, verb word, object word". The example sentence above, after combination, yields the action triple "Zhang San, coordinate, customer online". All sentences in the performance review text A are iterated through, and the above operation is performed on each sentence from which complete agent and patient roles can be extracted. The action triples generated from each sentence are stored sequentially in a list, forming an action triple sequence. If the semantic role annotation result of the core verb in a sentence lacks an agent or patient role, the sentence is skipped, no action triple is generated, and it is not included in the action triple sequence.

[0037] The behavior triples are input into the capability-related behavior pattern library for matching, and the behavioral evidence strength value of the behavior triples for the target capability words is calculated. The target capability words are the capability words in the capability-related behavior pattern library that are pre-bound to the standard behavior triples that have been successfully matched. In a preferred embodiment of the present invention, the process of calculating the strength value of behavioral evidence is as follows: Retrieve a pre-built capability-related behavior pattern library, which is stored in a data file in a key-value pair structure. Each record contains a standard behavior triplet and at least one capability word identifier associated with that standard behavior triplet. The standard behavior triplet consists of three fields: standard behavior subject word, standard behavior verb, and standard behavior object word. The standard behavior subject word is uniformly a preset fixed role placeholder, such as "employee". The standard behavior verb and standard behavior object word are derived from the actual operation action name and operation target name in the structured operation log. The capability word identifier field stores one or more capability words. For example, if the standard behavior triplet of a record is "employee, host, cross-departmental needs alignment meeting", its associated capability word identifier field includes "cross-departmental collaboration capability" and "meeting organization capability".

[0038] Each behavior triple in the sequence is extracted sequentially as the current behavior triple to be processed. Role normalization is then performed on the current behavior triple: the subject word in the behavior triple is replaced with the preset fixed role placeholder "employee," while the action verb and action object remain unchanged, resulting in a role-normalized behavior triple. For example, the behavior triple "Zhang San, Coordination, Customer Online" becomes "Employee, Coordination, Customer Online" after role normalization. The purpose of role normalization is to ensure that the subject words of the behavior triples extracted from the comments are consistent in form with the subject words of the standard behavior triples in the ability-related behavior pattern library, avoiding the impact of differences in names or personal pronouns on subsequent matching calculations.

[0039] The action verbs in the role-normalized action triples are mapped to a distributed semantic vector space. This mapping operation uses a pre-trained Chinese word vector model (familiar to those skilled in the art, and will not be elaborated upon here). This model contains 300-dimensional floating-point vector representations of approximately 300,000 Chinese words. The action verb is input into the word vector model as a query term. If the word exists in the vocabulary of the word vector model, its corresponding 300-dimensional vector is extracted as the action verb vector. If the word is not in the vocabulary, it is split into individual characters, and the average of the vectors of each character is taken as the vector representation of the word. Simultaneously, the standard action verbs in the currently compared standard action triples are mapped to 300-dimensional standard action verb vectors in the same way. The cosine value of the action verb vector and the standard action verb vector is calculated. The cosine value is calculated by dividing the inner product of the two vectors by the product of their magnitudes; the result is a floating-point number from -1 to +1, denoted as the first cosine value. For example, after mapping the action verb "coordinate" and the standard action verb "preside over" to vectors, the first cosine value is calculated to be 0.68, and the first cosine value is calculated to be 1.00 for the action verb "coordinate" and the standard action verb "coordinate".

[0040] The same mapping and calculation operations are applied to behavior object words. Behavior object words in the role-normalized behavior triples are mapped to 300-dimensional behavior object vectors, and standard behavior object words in the currently compared standard behavior triples are mapped to 300-dimensional standard behavior object vectors. The cosine value between the two is calculated and denoted as the second cosine value. For example, the second cosine value of the behavior object word "customer online" and the standard behavior object word "cross-departmental requirement alignment meeting" is 0.42, and the second cosine value of the behavior object word "customer online" and the standard behavior object word "customer deployment" is 0.85.

[0041] The arithmetic mean of the first and second cosine values ​​is calculated by adding the two values ​​and dividing by 2. The result is used as the behavior matching score between the current character's normalized behavior triplet and the current standard behavior triplet. For example, if the first cosine value is 0.68 and the second cosine value is 0.42, the arithmetic mean yields a behavior matching score of 0.55. This score is then compared to a preset matching threshold of 0.70. If the score exceeds 0.70, the current character's normalized behavior triplet is considered a successful match with the current standard behavior triplet. All ability word identifiers bound to this standard behavior triplet in the ability-related behavior pattern library are extracted and used as target ability words. If the score does not exceed 0.70, it is considered a mismatch, and the comparison continues with the next record's standard behavior triplet in the ability-related behavior pattern library.

[0042] For example, after role normalization, the behavior triple "Zhang San, coordination, customer onboarding" is compared with the standard behavior triple "employee, coordination, customer deployment" in terms of behavior matching degree. The first cosine value of the behavior verb "coordination" is 1.00, and the second cosine value of the behavior object words "customer onboarding" and "customer deployment" is 0.85. The arithmetic mean yields a behavior matching degree of 0.925, which exceeds the preset matching threshold of 0.70. The match is successful, and the capability word identifier "project management capability" bound to this standard behavior triple is extracted as the target capability word.

[0043] After comparing the current behavior triple with all standard behavior triples in the capability-related behavior pattern library, multiple successfully matched target capability words may be obtained. The same behavior triple may also match multiple different standard behavior triples. All successfully matched target capability words are aggregated and grouped by target capability word. The behavior matching scores of all behaviors pointing to the same target capability word are accumulated, and the accumulated result is used as the behavior evidence strength value for that target capability word. For example, if three different behavior triples match the target capability word "cross-departmental collaboration ability" with matching scores of 0.78, 0.82, and 0.71 respectively, the accumulated behavior evidence strength value for "cross-departmental collaboration ability" is 2.31. After performing all the above matching and accumulation operations on all behavior triples in the behavior triple sequence, each target capability word and its corresponding behavior evidence strength value are output.

[0044] In a preferred embodiment, the process of constructing the capability-related behavior pattern library is as follows: This system retrieves data automatically recorded by structured operation logs from all personnel within the same organization as they complete assigned tasks. These logs are automatically generated by the organization's internal business systems when employees perform work tasks. Each log entry corresponds to an operation performed by a single employee and includes three fields: Operator Identifier, Operation Action Identifier, and Operation Target Identifier. The Operator Identifier stores the unique ID of the person performing the operation; the Operation Action Identifier stores the standard action name; and the Operation Target Identifier stores the name of the specific object targeted by the operation. For example, one log entry might have an Operator Identifier of E001, an Operation Action Identifier of "Create," and an Operation Target Identifier of "Project Progress Tracking Table." Another log entry might have an Operator Identifier of E002, an Operation Action Identifier of "Moderate," and an Operation Target Identifier of "Cross-Departmental Requirements Alignment Meeting." The log data collection cycle is consistent with the performance evaluation cycle, and data is exported in batches on a quarterly or annual basis.

[0045] Replace the operator identifier field with a preset fixed role placeholder. This placeholder is a uniform identifier string, such as "employee." It does not contain any specific personnel identification information; it only indicates that the operation was performed by a member of the organization. During the replacement, directly replace the operator number corresponding to the operator identifier field in each log entry with the placeholder string "employee," and fill the standard action subject slot with the replacement result. For example, in the log, operator identifier field E001 is replaced with "employee," and E002 is also replaced with "employee."

[0046] Extract the operation action name corresponding to the operation action identifier field and fill it into the standard behavior verb slot. The operation action identifier field stores the name text of the operation action, so this text is directly extracted and filled into the standard behavior verb slot. Extract the operation target name corresponding to the operation target identifier field and fill it into the standard behavior object word slot, again by directly extracting the name text of the operation target identifier field and filling it in. Taking the two logs above as examples, after processing, the standard behavior subject word slot of the first log is "employee", the standard behavior verb slot is "create", and the standard behavior object word slot is "project progress tracking table", which is combined into the standard behavior triple "employee, create, project progress tracking table"; the second log is combined into the standard behavior triple "employee, host, cross-departmental requirement alignment meeting".

[0047] Merging identical standard behavior triples involves iterating through all generated standard behavior triples. Triples with identical standard behavior verbs and standard behavior object words are considered to have the same structure. Only one copy of each identical standard behavior triple is retained, and all duplicate records are deleted. For example, in the log data, employee E005 also performed the operation of hosting a cross-departmental requirements alignment meeting. This log also generated the standard behavior triple "employee, host, cross-departmental requirements alignment meeting". After merging, only one record of this structure is retained, discarding duplicates.

[0048] To obtain competency definition standards independent of performance reviews, these standards are behavioral competency mapping documents developed by human resource management experts before program implementation. Each row of the table contains an anchor behavior description and the corresponding competency term. The anchor behavior description is a textual summary of a typical work behavior, including the behavioral verb and the typical object of the behavior; the competency term is the organization-approved competency name. For example, the competency behavior anchoring table might contain a row with the anchor behavior description "chairing or organizing cross-departmental meetings to coordinate resources from multiple parties to achieve common goals," corresponding to the competency term "cross-departmental collaboration ability"; and another row with the anchor behavior description "developing or creating project-level planning documents, schedules, or task schedules," corresponding to the competency term "project management ability."

[0049] For each retained standard behavior triple after merging, a matching anchor behavior description is found in the capability behavior anchoring table. The matching operation concatenates the standard behavior verb and standard behavior object in the standard behavior triple into a query text. All anchor behavior descriptions in the capability behavior anchoring table are traversed, and text similarity calculation is used to determine if they match. A match is considered successful if the similarity exceeds a preset threshold (the text similarity calculation method is also common knowledge in the field and will not be elaborated here). For example, the standard behavior triple "employee, chair, cross-departmental needs alignment meeting" concatenates to obtain "chair cross-departmental needs alignment meeting," which matches the anchor behavior description "chair or organize cross-departmental level meetings to coordinate resources from multiple parties to achieve common goals." The standard behavior triple "employee, create, project progress tracking table" concatenates to obtain "create project progress tracking table," which matches the anchor behavior description "develop or create project-level planning documents, schedules, or task schedules."

[0050] Extract the competency words corresponding to the successfully matched anchored behavior descriptions as competency word identifiers and bind them to the current standard behavior triples. For example, "employee, host, cross-departmental needs alignment meeting" is bound to the competency word identifier "cross-departmental collaboration ability," and "employee, create, project progress tracking table" is bound to the competency word identifier "project management ability." A standard behavior triple may match multiple anchored behavior descriptions; therefore, all successfully matched competency words are bound to that standard behavior triple as competency word identifiers. Write all bound standard behavior triples and their corresponding competency word identifiers to a data file one by one. Each binding record becomes a record in the competency-related behavior pattern library. After all binding records are written, a complete competency-related behavior pattern library is formed. The entire construction process of this library does not involve any performance evaluation data of the evaluated individuals.

[0051] Generate a capability tag for the target capability word, perform semantic alignment determination between the evaluative capability word and the target capability word, and mark the capability tag according to the determination result.

[0052] In a preferred embodiment of the present invention, the process of marking capability tags is as follows: The process retrieves target competency terms and their corresponding behavioral evidence strength values. These terms, output from the previous step, are passed in as a list of key-value pairs. A competency tag is generated for each target competency term. The competency tag is a two-field data structure: the first field stores the text string of the target competency term, and the second field stores the corresponding behavioral evidence strength value, which is a floating-point number. For example, if the behavioral evidence strength value for the target competency term "cross-departmental collaboration ability" is 2.31, the generated competency tag would be "cross-departmental collaboration ability, 2.31"; if the behavioral evidence strength value for the target competency term "project management ability" is 0.85, the generated competency tag would be "project management ability, 0.85". All target competency terms are generated with independent competency tags in this manner.

[0053] Retrieve all the evaluative competency words that have been tagged in the performance evaluation text A. These evaluative competency words are the words corresponding to the B-EVAL labels output by the deep learning model that has been trained during the model recognition stage. They are stored in list form and retrieved one by one for processing.

[0054] For each retrieved evaluative competency term, it is determined whether it is identical to the text string of a target competency term. The string of the current evaluative competency term is compared one by one with the strings of all target competency terms. If every character of a target competency term's string is identical to that of the current evaluative competency term's string, it is considered identical. If identical target competency terms exist, the supporting status field of the competency tag corresponding to that target competency term is marked as having comment support, and an association is established between the evaluative competency term and that competency tag. This association is recorded as a pointer or index, indicating that the evaluative competency term has behavioral evidence support and has appeared explicitly in the comment. For example, if the evaluative competency term is "cross-departmental collaboration ability," and "cross-departmental collaboration ability" happens to exist in the target competency term list, then the competency tag "cross-departmental collaboration ability, 2.31" is marked as having comment support, and the evaluative competency term "cross-departmental collaboration ability" is associated with that tag.

[0055] If the current evaluative ability word is different from all target ability words, then the evaluative ability word is mapped to a distributed semantic vector space to obtain an evaluative ability vector. The mapping operation uses a pre-trained Chinese word vector model, the same model used in the ability-related behavior pattern library construction phase. The evaluative ability word is input into the word vector model as a query word. If the word exists in the vocabulary of the word vector model, its 300-dimensional vector is extracted as the evaluative ability vector; if the word is not in the vocabulary, the word is split into individual characters, and the average of the vectors of each character is taken as the evaluative ability vector. Simultaneously, each target ability word is mapped to a 300-dimensional target ability vector in the same way.

[0056] Calculate the cosine value of the evaluation capability vector and each target capability vector separately. The calculation method is to take the inner product of the two vectors and divide it by the product of their magnitudes; the result is a floating-point number ranging from negative to positive 1. For example, if the evaluation capability term is "communication and coordination ability," after mapping it to the evaluation capability vector, the cosine value calculated with the target capability vector of the target capability term "cross-departmental collaboration ability" is 0.78, with the target capability vector of the target capability term "project management ability" it is 0.41, and with the target capability vector of the target capability term "customer orientation" it is 0.35. Take the maximum value from these cosine values; in this example, the maximum value is 0.78. Compare the maximum value with a preset near-similar judgment threshold, which is set to 0.70.

[0057] If the maximum cosine value exceeds the preset synonym determination threshold of 0.70, the supporting status field of the ability tag corresponding to the target ability word with the largest cosine value is marked as having comment support, and the current evaluative ability word is associated with that ability tag. In the example above, the maximum value of 0.78 exceeds 0.70, so the ability tag corresponding to "cross-departmental collaboration ability" is marked as having comment support, and the evaluative ability word "communication and coordination ability" is associated with that tag, indicating that the "communication and coordination ability" explicitly used in the comment is semantically synonymous with "cross-departmental collaboration ability," and is regarded as the same ability and associated with the corresponding behavioral evidence.

[0058] If the maximum cosine value does not exceed the preset synonym judgment threshold of 0.70, the evaluative ability term is marked as having no behavioral evidence to support it. To save this judgment result, a record without a behavioral evidence strength value is generated for the evaluative ability term. The record contains the text of the evaluative ability term and the status label "no behavioral evidence to support it". For example, if the evaluative ability term is "strategic thinking", and the maximum cosine value calculated with all target ability terms is 0.35, which does not exceed 0.70, then "strategic thinking" is marked as having no behavioral evidence to support it, indicating that the ability term appears in the evaluation but no corresponding objective behavioral evidence is found.

[0059] Following the above method, iterate through all the marked evaluative competency words in the performance evaluation text A, and complete the judgment, association, or marking operations one by one.

[0060] After iterating through all evaluative competency terms and completing the assessment, the supporting status fields of all generated competency tags are checked. Competency tags whose supporting status fields are not marked as having comment support are filtered out, and their supporting status fields are marked as having behavioral evidence but no comment support. For example, the target competency term "customer-oriented" generates the competency tag "customer-oriented, 0.91". During the iteration of evaluative competency terms, no evaluative competency terms are associated with this tag, and its supporting status is not marked as having comment support. Therefore, it is marked as having behavioral evidence but no comment support, indicating that the employee's objective operational log behavior reflects this competency, but the competency term is not explicitly mentioned in the performance evaluation. At this point, the marking operation for all competency tags is complete, and all competency tags and their corresponding supporting status tags are output.

[0061] A data analysis system based on deep learning, characterized in that it includes: AI module: Preprocesses historical performance review texts to obtain a standardized dataset; A deep learning model was constructed and trained using a standardized dataset. The trained deep learning model was used to identify and label evaluative competency words in performance review texts. Input the performance evaluation text A to be analyzed into the trained deep learning model to obtain the performance evaluation text A labeled with evaluative ability words; Analysis module: Extract the subject words, verb words, and object words from the performance evaluation text A, and combine the subject words, verb words, and object words into behavior triples; The behavior triples are input into the capability-related behavior pattern library for matching, and the behavioral evidence strength value of the behavior triples for the target capability words is calculated. The target capability words are the capability words in the capability-related behavior pattern library that are pre-bound to the standard behavior triples that have been successfully matched. Tagging module: Generates a capability tag for the target capability word, performs semantic alignment determination between the evaluative capability word and the target capability word, and tags the capability tag according to the determination result.

[0062] This invention identifies and tags evaluative competency terms from performance reviews. Simultaneously, it independently extracts behavioral subject words, behavioral verbs, and behavioral object words from the review text to form behavioral triples. These behavioral triples are input into a competency-related behavioral pattern library for matching and calculating the strength of behavioral evidence. The standard behavioral triples in this library are derived from structured operation logs. The binding relationship between the standard behavioral triples and competency terms is established based on a job competency model or job competency dictionary independent of performance reviews, and does not rely on the performance reviewer's own comments, thus eliminating the contamination of behavioral evidence standards by superiors' subjective favoritism. Subsequently, competency tags are generated for each target competency term obtained through behavioral matching. Evaluative competency terms are semantically aligned with target competency terms. Evaluative competency terms that are completely matched or semantically similar are associated with their corresponding competency tags and marked as having comment support. Evaluative competency terms that do not semantically match are marked as having no behavioral evidence support. Competency tags not associated with any evaluative competency term are marked as having behavioral evidence but no comment support. As a result, it is clear whether each ability word in the evaluation is supported by objective behavioral facts. Implicit abilities that are not mentioned in the evaluation but are reflected in the behavior can also be discovered in reverse. The final output ability label effectively distinguishes between abilities based on real behavioral evidence and evaluative statements from superiors' subjective projections.

[0063] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A data analysis method based on deep learning, characterized in that, Includes the following steps: A standardized dataset is obtained by preprocessing historical performance review texts; A deep learning model was constructed and trained using a standardized dataset. The trained deep learning model was used to identify and label evaluative competency words in performance review texts. Input the performance evaluation text A to be analyzed into the trained deep learning model to obtain the performance evaluation text A labeled with evaluative ability words; Extract the subject words, verb words, and object words from the performance evaluation text A, and combine them into behavior triples. The behavior triples are input into the capability-related behavior pattern library for matching, and the behavioral evidence strength value of the behavior triples for the target capability words is calculated. The target capability words are the capability words in the capability-related behavior pattern library that are pre-bound to the standard behavior triples that have been successfully matched. Generate a capability tag for the target capability word, perform semantic alignment determination between the evaluative capability word and the target capability word, and mark the capability tag according to the determination result.

2. A data analysis method based on deep learning according to claim 1, characterized in that, The process of obtaining a standardized dataset is as follows: Retrieve historical performance appraisal texts, which are derived from structured performance appraisal records stored in the human resources system; The structured performance appraisal records are parsed to extract the appraisal text field and the appraisee identification field. The extracted appraisal text is then merged according to the appraisee identification field. The merged appraisal text is then segmented into sentences, deduplicated, and punctuation and non-visible characters are removed to obtain a standardized dataset.

3. A data analysis method based on deep learning according to claim 1, characterized in that, The training process of a deep learning model is as follows: Obtain the set of ability words, which stores multiple preset evaluative ability words; The text samples in the standard dataset are segmented into words to obtain a text word sequence. Words in the text word sequence that appear in the ability word set are labeled as positive examples, and words in the text word sequence that do not appear in the ability word set are labeled as negative examples, resulting in a labeled word sequence. The combination of the text word sequence and the labeled word sequence is used as a training sample. A sequence labeling deep learning model is constructed. The sequence labeling deep learning model adopts a combination architecture of bidirectional long short-term memory network and conditional random field. The text word sequence in the training sample is used as input and the labeled word sequence in the training sample is used as output target. The sequence labeling deep learning model is trained using multiple training samples to obtain the trained deep learning model.

4. A data analysis method based on deep learning according to claim 1, characterized in that, The process of obtaining the behavior triples is as follows: The performance evaluation text A consists of multiple sentences. Verb recognition is performed on each sentence to locate the core verb in each sentence. Call the verb semantic framework library to obtain the semantic framework corresponding to the core verb. The semantic framework defines the required semantic role types for the core verb. Extract specific words that fill the agent role from the sentence as the subject of the action, extract specific words that fill the patient role as the object of the action, and use the core verb as the action verb; The subject word, verb word, and object word extracted from the same sentence are combined into a behavior triplet. This process is repeated for all sentences in the performance review text A to generate a sequence of behavior triplets consisting of multiple behavior triplets.

5. A data analysis method based on deep learning according to claim 1, characterized in that, The process of calculating the strength value of behavioral evidence is as follows: Obtain a pre-built capability-related behavior pattern library. Each record in the capability-related behavior pattern library contains a standard behavior triple and at least one capability word identifier pre-associated with the standard behavior triple. Replace the subject word in the behavior triple with a preset fixed role placeholder to obtain a role-normalized behavior triple; Map the action verbs in the normalized action triplet and the standard action verbs in the standard action triplet to the distributed semantic vector space, and calculate the first cosine value of the action verb vector and the standard action verb vector. Map the behavior object words in the normalized behavior triplet and the standard behavior object words in the standard behavior triplet to the distributed semantic vector space respectively, and calculate the second cosine value of the behavior object vector and the standard behavior object vector. The first cosine value and the second cosine value are arithmetically averaged, and the result is used as the behavior matching degree value. Pairings with behavior matching degree values ​​exceeding the preset matching threshold are judged as successful matches. The ability word identifiers bound to the standard behavior triples that are successfully matched are extracted as target ability words. All behavior matching degree values ​​pointing to the same target ability word are accumulated, and the accumulated result is used as the behavior evidence strength value for the target ability word.

6. A data analysis method based on deep learning according to claim 1, characterized in that, The process of building a capability-related behavior pattern library is as follows: Retrieve structured operation logs automatically recorded by the system for all personnel within the same organization when completing predetermined tasks. The structured operation logs include operator identifier field, operation action identifier field, and operation target identifier field. Replace the operator identifier field with a preset fixed role placeholder, fill the replacement result into the standard behavior subject word slot, extract the operation action name corresponding to the operation action identifier field and fill it into the standard behavior verb slot, extract the operation target name corresponding to the operation target identifier field and fill it into the standard behavior object word slot, combine them into standard behavior triples, merge standard behavior triples that have the same structure, and keep one copy of each standard behavior triple that has the same structure. Obtain competency definition standards independent of performance evaluations. The competency definition standards include an expert-developed competency behavior anchoring table. For each standard behavior triple, find the anchor behavior description in the competency definition standards that matches the standard behavior verb and standard behavior object in that standard behavior triple. Extract the competency word corresponding to the anchor behavior description as a competency word identifier and bind it to the standard behavior triple. Collect all the bound standard behavior triples and their corresponding competency word identifiers to build a competency-related behavior pattern library.

7. A data analysis method based on deep learning according to claim 1, characterized in that, The process of tagging ability labels is as follows: Obtain the target ability words and their corresponding behavioral evidence strength values, and generate an ability label for each target ability word. The ability label contains the target ability word and its corresponding behavioral evidence strength value. Obtain all the marked evaluative competency words in the performance evaluation text A. For each evaluative competency word, determine whether it is exactly the same as a certain target competency word. If they are the same, mark the competency tag corresponding to the target competency word as having comment support and associate the evaluative competency word with the competency tag. For each evaluative ability word, if it is different from all target ability words, then the evaluative ability word is mapped to the distributed semantic vector space to obtain the evaluative ability vector, and each target ability word is mapped to the distributed semantic vector space to obtain the target ability vector. The cosine values ​​of the evaluative ability vector and each target ability vector are calculated respectively. When the maximum value of the cosine value exceeds the preset synonym judgment threshold, the ability tag corresponding to the target ability word with the largest cosine value is marked as having commentary support, and the evaluative ability word is associated with the ability tag. When the maximum value of the cosine value does not exceed the preset synonym judgment threshold, the evaluative ability word is marked as having no behavioral evidence to support it. After iterating through all evaluative competency terms and completing the assessment, all competency tags that were not marked as having behavioral evidence but no evaluative support are marked as having behavioral evidence but no evaluative support.

8. A data analysis system based on deep learning, characterized in that, include: AI module: Preprocesses historical performance review texts to obtain a standardized dataset; A deep learning model was constructed and trained using a standardized dataset. The trained deep learning model was used to identify and label evaluative competency words in performance review texts. Input the performance evaluation text A to be analyzed into the trained deep learning model to obtain the performance evaluation text A labeled with evaluative ability words; Analysis module: Extract the subject words, verb words, and object words from the performance evaluation text A, and combine the subject words, verb words, and object words into behavior triples; The behavior triples are input into the capability-related behavior pattern library for matching, and the behavioral evidence strength value of the behavior triples for the target capability words is calculated. The target capability words are the capability words in the capability-related behavior pattern library that are pre-bound to the standard behavior triples that have been successfully matched. Tagging module: Generates a capability tag for the target capability word, performs semantic alignment determination between the evaluative capability word and the target capability word, and tags the capability tag according to the determination result.

Images