A privacy protection method and system for unstructured teacher comment text

Abstract

The present application relates to the technical field of data privacy protection, in particular to a privacy protection method and system for unstructured teacher comment text. The present application comprises text preprocessing and pre-training model labeling to generate candidate labels; vectorization and semantic encoding to extract context dependence; generating comment level representation, enhancing and constraining output feature sequence through role perception; bidirectional LSTM modeling time sequence, fusing pre and post context information; conditional random field decoding, introducing context constraint to modulate label transition, outputting entity labels; judging and evaluating behavior based on enhanced features, reconstructing state and checking, mapping to desensitized text, and outputting after consistency check. The present application can identify sensitive evaluation content of teacher comments in a complex semantic environment, and can balance the integrity of educational evaluation during privacy protection to adapt to the actual needs of safe processing of teacher comments in the educational informationization scene.

Description

Technical Field

[0001] This invention relates to the field of data privacy protection technology, specifically to a privacy protection method and system for unstructured teacher comment texts. Background Technology

[0002] With the continuous improvement of educational informatization, the development of educational informatization has promoted the large-scale collection and analysis of educational data, making data privacy protection an increasingly prominent issue. Existing research and technological applications mainly focus on the privacy protection of structured data such as grades and attendance, paying insufficient attention to the privacy risks of unstructured text data such as teacher comments. Teacher comments are characterized by strong semantic implicitness and high context dependence, often implying sensitive information such as students' personality and behavioral habits through overall semantics, making their privacy boundaries difficult to define clearly. Named entity recognition, as a fundamental technology for text privacy protection, currently relies heavily on the fusion of pre-trained language models and sequence labeling models, and has shown effectiveness in general domains. Meanwhile, text anonymization techniques mainly originate from fields such as finance and healthcare, often employing methods such as sensitive field identification, keyword replacement, content masking, or deletion to process structured or semi-structured data.

[0003] However, the aforementioned technologies still have some technical shortcomings: On the one hand, in the context of teacher comments, regarding named entity recognition, teacher comments often contain colloquial expressions, subjective modifications, and polysemous terms, making it difficult for general semantic models to stably distinguish semantic functions in different evaluation contexts; the entity categories in comments are actually semantic entities centered on the teaching evaluation dimension, and general models cannot effectively model this type of structural information; in addition, negative evaluations are often implicit in the overall semantics rather than specific terms, and relying solely on lexical or syntactic features can easily lead to unstable recognition results and semantic deviations. On the other hand, in text anonymization, teacher comments, as highly unstructured text, have dispersed and implicit privacy risks, making it difficult for anonymization methods based on fixed rules to fully cover them; simple word replacement or sentence rewriting can easily destroy the original evaluation logic and educational orientation of the comments, leading to semantic confusion and reduced usability; while directly using deletion or masking methods can reduce privacy risks, they often simultaneously lose key evaluation information, failing to achieve a balance between privacy protection and the preservation of educational value, and thus failing to meet the actual needs of educational scenarios for comments to be "both safe and usable".

[0004] Therefore, a privacy protection method and system for unstructured teacher comment texts is needed. Summary of the Invention

[0005] This invention provides a privacy protection method and system for unstructured teacher evaluation texts, aiming to solve the technical problems of existing natural language processing technology in the privacy protection scenario of teacher evaluation texts, which is difficult to accurately identify implicit evaluation risks and whose desensitization processing easily destroys the semantic structure of educational evaluation.

[0006] The technical solution of this invention is as follows: A privacy protection method for unstructured teacher comments includes the following steps: Step S1. Collect the original teacher comment texts in different teaching scenarios and perform preprocessing and standardization. After preliminary entity annotation, perform candidate trigger marking on the evaluation-related segments in the teacher comment texts to generate evaluation entity candidate marking information and standardized teacher comment text sequences. Step S2. Convert the standardized teacher comment text sequence into a vectorized input format for semantic encoding. Extract the contextual dependencies between terms through multi-layer attention processing to obtain a contextual semantic vector sequence. Step S3. Based on the context semantic vector sequence, perform pooling and nonlinear mapping to generate a comment-level semantic representation, identify the evaluation behavior and generate an evaluation dimension vector and an evaluation target pointing vector; predefine a set of semantic role types, and construct a role embedding matrix for each role type; calculate the matching score between the term and each semantic role and normalize it to obtain the role distribution, generate a role perception enhancement representation, and introduce the implicit evaluation object embedding vector; apply role constraints to subjective modification terms and ambiguous terms, and output an enhanced semantic feature sequence; Step S4. Based on the output enhanced semantic feature sequence, a bidirectional long short-term memory network is used to perform temporal dependency modeling from both the forward and backward directions, concatenate the bidirectional hidden states, and output a context feature sequence that integrates the preceding and following context information. Step S5. Based on the context feature sequence, construct the state feature function and the transition feature function. In the label transition matrix of the conditional random field decoding process, introduce the evaluation context constraint parameter to perform weighted modulation on the label transition matrix. After decoding, output the optimal entity label sequence as the teaching evaluation entity label result corresponding to each term. Step S6. Receive the obtained context feature sequence and entity label sequence, construct the evaluation behavior state vector, determine the evaluation behavior type based on the evaluation behavior state vector, reconstruct the evaluation behavior state, introduce institutionalized teaching evaluation expression state constraints to verify the legality of the reconstructed evaluation behavior state, and then map the reconstructed evaluation behavior state to desensitized evaluation paragraph text. After passing the teaching evaluation consistency test, output the desensitization result.

[0007] Furthermore, in step S3, the implicit evaluation object embedding vector is introduced for cases where the evaluation object does not appear explicitly in the teacher's comments. When no explicit evaluation object role is detected in the role distribution, the system injects the implicit evaluation object embedding vector into the evaluation structure representation space as a semantic anchor point between the evaluation behavior and the potential evaluated object.

[0008] Furthermore, in step S3, the subjective modifier terms are determined based on the pre-constructed evaluation modifier term list or the output of the role classifier. The role distribution is constrained to be normalized only within the evaluation adjustment role subspace. For ambiguous terms with multiple potential semantic functions, their role distribution calculation is always completed under the joint constraint of the evaluation dimension vector and the evaluation target pointing vector.

[0009] Furthermore, in step S6, the evaluation behavior type is determined for the located evaluation segment based on the evaluation behavior state vector, according to the joint distribution of the following three types of state features, as detailed below: a. Evaluation-oriented state characteristics: used to determine whether the evaluation is directly targeted at the individual student; b. Evaluation of stability characteristics: used to determine whether the evaluation has a tendency towards fixed abilities or qualitative personality traits; c. Evaluation functional state characteristics: used to distinguish between descriptive, judgmental, or developmental guidance evaluation behaviors; Then, after fusing the above three types of state features, the evaluation behavior type label is output.

[0010] Furthermore, the specific process of reconstructing the evaluation behavior state in step S6 is as follows: Based on the semantic role binding matrix, the set of role relationships in the evaluation behavior is analyzed to determine whether the evaluation object is a student individual entity and whether the entity carries the function of negative evaluation. When the evaluation behavior type is determined to have privacy risks, the evaluation pointing to state mapping operation is performed. At the same time, the semantic role binding relationship is rebounded.

[0011] Furthermore, the institutionalized teaching evaluation expression constraints described in step S6 are specifically implemented as follows: The expression structure of evaluation behavior is limited by state constraints, including: a. Evaluation starting state constraints, which check whether the starting state of the evaluation behavior points to the learning situation or development stage; b. Evaluation core state constraints, which check whether the core state of the evaluation behavior is consistent with the original evaluation dimension; c. Evaluation closing state constraints, which check whether the closing state of the evaluation behavior is improvement or development-oriented. The expression constraints of institutionalized teaching evaluation act as a state filtering function on the reconstructed evaluation behavior state.

[0012] Furthermore, in step S6, the evaluation behavior state vector is constructed by combining the contextual semantic representation of the evaluation segment, the evaluation dimension vector, the evaluation target pointing vector, and the semantic role binding matrix.

[0013] Furthermore, in step S3, the system calculates the matching score of the term with each semantic role under the constraints of the evaluation context, obtains the distribution of the term under each role after normalization, and performs weighted fusion of the role embedding based on the role distribution to generate the role-aware representation of the term. At the same time, it is fused with the original semantic representation of the term to obtain the enhanced semantic representation.

[0014] Furthermore, in step S2, the standardized teacher comment text sequence is converted into a vectorized input format adapted to the pre-trained language model and semantically encoded. The vectorized input includes term embedding, position embedding, and segment embedding, which are used to represent the basic semantic information of terms in the teacher comment text and their relative position information in the text.

[0015] A privacy-preserving system for unstructured teacher comments, comprising: The input module collects raw teacher comment texts from different teaching scenarios and performs preprocessing and standardization. After initial entity annotation, it performs candidate trigger marking on evaluation-related segments in the teacher comment texts to generate evaluation entity candidate marking information and standardized teacher comment text sequences. The embedding module converts the standardized teacher comment text sequence into a vectorized input format for semantic encoding. It extracts the contextual dependencies between terms through multi-layer attention processing to obtain a contextual semantic vector sequence. The semantic enhancement module generates a comment-level semantic representation by pooling and non-linear mapping based on the context semantic vector sequence, identifies the evaluation behavior, and generates an evaluation dimension vector and an evaluation target pointing vector; it predefines a set of semantic role types and constructs a role embedding matrix for each role type; it calculates the matching score between terms and each semantic role and normalizes it to obtain the role distribution, generates a role-aware enhanced representation, and introduces the implicit evaluation object embedding vector; it applies role constraints to subjective modifier terms and ambiguous terms, and outputs an enhanced semantic feature sequence. The encoding module, based on the output enhanced semantic feature sequence, uses a bidirectional long short-term memory network to model temporal dependencies from both the forward and backward directions, concatenates the bidirectional hidden states, and outputs a context feature sequence that integrates the preceding and following contextual information. The decoding module constructs state feature functions and transition feature functions based on context feature sequences. In the label transition matrix of the conditional random field decoding process, evaluation context constraint parameters are introduced to weight and modulate the label transition matrix. After decoding, the optimal entity label sequence is output as the teaching evaluation entity label result corresponding to each term. The desensitization module receives the received context feature sequence and entity label sequence, constructs an evaluation behavior state vector, determines the evaluation behavior type based on the evaluation behavior state vector, reconstructs the evaluation behavior state, introduces institutionalized teaching evaluation expression state constraints to verify the legality of the reconstructed evaluation behavior state, and then maps the reconstructed evaluation behavior state to desensitized evaluation text segments. After passing the teaching evaluation consistency check, the desensitization result is output. The beneficial effects of this invention are as follows: 1. This invention improves the stability and accuracy of teaching evaluation entity recognition in teacher comment texts by introducing an evaluation structure-driven semantic enhancement module, which explicitly models evaluation dimensions, evaluation objectives and semantic role information at the intermediate representation level, and effectively reduces the impact of subjective modification and semantic ambiguity on the recognition results.

[0016] 2. This invention achieves technical dissolution of the individual student orientation in negative evaluations by performing state mapping desensitization processing in the evaluation structure representation space, avoiding the destruction of the evaluation semantic structure by desensitization methods based on word replacement or text deletion.

[0017] 3. This invention constructs a complete technical process from evaluation entity recognition to evaluation structure reconstruction-based desensitization, achieving a balance between privacy protection and the preservation of teaching evaluation functions. It is applicable to the secure processing and application of teacher comment texts in educational informatization scenarios. Detailed Implementation

[0018] To better understand the above technical solutions, specific implementation methods will be described in detail below. It should also be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This embodiment provides a privacy protection method for unstructured teacher comment text, including the following steps: S1. Collect original teacher comment texts from different teaching scenarios and perform preprocessing and standardization. After preliminary entity annotation, perform candidate trigger marking on the evaluation-related segments in the teacher comment text to generate evaluation entity candidate marking information and standardized teacher comment text sequences.

[0020] In this embodiment of the invention, original teacher comments texts from various teaching scenarios are collected, the collected texts are cleaned to remove noise content and retain the original format symbols, and the sentences are made fluent through text normalization. At the same time, missing values, duplicate values ​​and outliers in the text are processed.

[0021] Preliminary entity annotations were performed on the cleaned text data, and standardized annotation datasets were formed by combining manual verification and correction.

[0022] This step performs evaluation entity candidate triggering labeling on evaluation-related segments in the teacher's comments text. This processing is based on the text's lexical statistical features, contextual semantic similarity features, and syntactic structure features to generate candidate labeling information for positions that may involve teaching evaluation entities, which is used to guide the subsequent model's attention weight distribution on evaluation-related segments.

[0023] S2. Convert the standardized teacher comment text sequence into a vectorized input format for semantic encoding. Extract the contextual dependencies between terms through multi-layer attention processing to obtain a contextual semantic vector sequence.

[0024] In this embodiment of the invention, this step performs contextual semantic encoding on the standardized teacher comment text sequence to generate an initial vector representation that reflects the general semantic features of the teacher comment text.

[0025] Specifically, this step converts the standardized teacher comment text sequence into a vectorized input format adapted to the pre-trained language model and performs semantic encoding processing. The vectorized input includes at least term embeddings, position embeddings, and segment embeddings, used to represent the basic semantic information of terms in the teacher comment text and their relative position information in the text.

[0026] During the encoding process of the pre-trained language model, the contextual dependencies between words in the teacher's comments text are modeled through multi-layer attention computing units, thereby achieving weighted aggregation of contextual semantic information.

[0027] After the encoding process, the output is a sequence of term-level vectors containing global context information, namely, a sequence of contextual semantic vectors, which is used to represent the initial semantic representation of the teacher's comments in the general semantic space.

[0028] S3. Based on the context semantic vector sequence, perform pooling and nonlinear mapping to generate a comment-level semantic representation, identify the evaluation behavior and generate an evaluation dimension vector and an evaluation target pointing vector; predefine a set of semantic role types and construct a role embedding matrix for each role type; calculate the matching score between the term and each semantic role and normalize it to obtain the role distribution, generate a role perception enhancement representation, and introduce the implicit evaluation object embedding vector; apply role constraints to subjective modifier terms and ambiguous terms, and output an enhanced semantic feature sequence.

[0029] In this embodiment of the invention, the semantic enhancement module is positioned between the embedding module and the encoding module, serving as an intermediate semantic representation reconstruction module. Its input is the sequence of context semantic vectors output by the embedding module. : Among them, among them, Let represent the semantic vector of the input context corresponding to the t-th term in the teacher's comment text, where T represents the term length of the teacher's comment text, and d represents the dimension of the semantic vector. The domain is represented by real numbers. The technical goal of the semantic enhancement module is to explicitly introduce the implicit institutionalized evaluation behavior structure in teacher comments into the intermediate representation space, and to enable the subsequent encoding module to perform sequence modeling under the pragmatic constraints of evaluation by performing representation reconstruction operations on the semantic representation of terms.

[0030] The semantic enhancement layer first bases itself on the context semantic vector sequence. Comment-level semantic representations are generated through pooling and nonlinear mapping. : Based on the global representation, the system uses an evaluation behavior discrimination network to determine whether the current text constitutes a teaching evaluation behavior, and generates a corresponding evaluation context condition vector when the evaluation behavior is valid.

[0031] After confirming that the evaluation behavior is valid, the system generates an evaluation dimension vector and an evaluation target orientation vector, respectively. Evaluation Dimension Vector : Evaluation target pointing vector : in, and This is a parameterized mapping function used to map the overall semantics of a comment to the evaluation dimension space and the evaluation orientation space. Represents the dimension of a spatial vector. Describe a vector space consisting of n real numbers. The dimension of the evaluation context representation vector is indicated. and Together, they constitute a global conditional representation of the evaluation behavior context structure, serving as constraint inputs for subsequent semantic role reconstruction. This evaluation behavior context structure is not obtained based on manually preset semantic rules or keyword triggering methods, but rather generated through parameterized mapping of the context semantic vector. The evaluation dimension vector and the evaluation target pointing vector are automatically learned by the model during training based on the statistical semantic distribution of the teacher's comment text; their values ​​exist in continuous vector form and do not correspond to fixed text rules or manual semantic judgments. The system predefines a set of semantic role types for teacher evaluation pragmatics. : in, Represents the first in the set of semantic role types A semantic role type, This represents the total number of predefined semantic role types. The set of semantic role types is used to limit the functional space dimension of the model's intermediate representation. Its purpose is to provide a finite and controllable representation subspace for modeling evaluation behavior, thus avoiding semantic drift in the model under unconstrained conditions. The specific size and dimension of the role set can be automatically adjusted based on the training data, and it does not inherently limit the meaning of specific language rules. A role embedding matrix is ​​constructed for each role type. ; For any term vector, the system calculates its matching score with each semantic role under the constraints of the evaluation context: in, Indicates the first The term and the first Matching scores between semantic role types; Indicates the position index of the word in the input sequence; Indicates the semantic role type index; Represents the role embedding matrix The corresponding number in the middle Role embedding vectors of semantic role types; , , The above formula represents the learnable parameter matrix, mapping term semantics, evaluation dimensions, and evaluation targets to the same computational space. Normalization yields the distribution of terms under each role. : =softmax ; in, Indicates the first The term in the first The system then performs a weighted fusion of role embeddings based on the role distribution across semantic role types to generate a role-aware representation vector for each term. : ; in, This represents the semantic role type index. It is then fused with the original term semantic representation to obtain an enhanced semantic representation. : ; in, This represents a vector concatenation operation. To address situations where the evaluated object is not explicitly present in teacher comments, an implicit evaluation object embedding vector is introduced. When no explicit evaluation object role is detected in the role distribution, the system injects the implicit evaluation object embedding vector into the evaluation structure representation space, serving as a semantic anchor between the evaluation behavior and the potential evaluated object. To address situations where the evaluated object is not explicitly mentioned in teacher comments, an implicit evaluation object embedding vector is introduced. : When no explicit evaluation object role is detected in the role distribution, the system will embed the implicit evaluation object into the vector. The injected evaluation structure representation space serves as a semantic anchor between the evaluation behavior and the potential evaluated object, maintaining the integrity of the evaluation structure. The implicit evaluation object embedding only serves as a placeholder vector in the semantic representation space for subsequent calculations, does not correspond to specific natural language terms, and does not introduce additional inference rules. Its role is to maintain the integrity of the evaluation behavior structure in the intermediate representation space. Subjective modifier terms are determined based on a pre-constructed evaluation modifier lexicon or the output of a role classifier. Their role distribution is constrained to be normalized only within the evaluation moderation role subspace, thus limiting their participation to evaluation intensity modeling and excluding entity boundary and entity category determination. For ambiguous terms with multiple potential semantic functions, their role distribution calculation is always performed within the evaluation moderation role subspace. and This is accomplished under joint constraints, avoiding semantic shifts caused by relying solely on local contextual similarity; The final output is an enhanced semantic feature sequence. : in, Indicates the first The enhanced semantic representation vectors corresponding to each term; Representation of enhanced semantic representation vector Dimensions.

[0032] Through the aforementioned role constraint calculation and representation fusion process, the semantic enhancement module performs an intermediate representation space reconstruction operation on the contextual semantic representation output by the embedding module, resulting in different enhanced representation results for the same term under different evaluation behavior contexts. This reconstruction operation changes the input feature distribution of subsequent encoding modules, enabling the sequence modeling process to be carried out under the constraints of evaluation behavior conditions, rather than in the unconstrained semantic space of a general language model.

[0033] S4. Based on the output enhanced semantic feature sequence, a bidirectional long short-term memory network is used to perform temporal dependency modeling from both the forward and backward directions, concatenate the bidirectional hidden states, and output a context feature sequence that integrates the preceding and following contextual information.

[0034] In this embodiment of the invention, this step does not directly model the original word vectors, but performs temporal dependency modeling constrained by the evaluation context for the enhanced semantic feature sequence that has explicitly introduced evaluation dimensions, evaluation target orientation and semantic role constraints.

[0035] The enhanced semantic feature sequence output in step S3 is represented as follows: : in, Indicates the first The enhanced semantic representation vectors corresponding to each term; Representation of enhanced semantic representation vector The dimension; every moment It is composed of a general contextual semantic representation, semantic role encoding under the current evaluation behavior, and global contextual constraint information fused with evaluation dimensions and evaluation goals. The enhanced feature sequence serves as the input to BiLSTM, enabling it to perceive the evaluation behavior structure during the temporal modeling stage, rather than relying solely on word order relationships. A bidirectional long short-term memory (BiLSTM) network is used to perform temporal modeling of the enhanced feature sequence from both the forward and backward directions. Taking the forward LSTM as an example, its hidden state vector at sequence position t is... : in, This represents the hidden state vector of the forward LSTM at sequence position t-1.

[0036] The backward LSTM models the sequence in reverse order in the same way, obtaining the hidden state vector at position t in the sequence. : in, This represents the hidden state vector at position t+1 of the backward LSTM sequence.

[0037] The two-way hidden states are concatenated at the same time, fusing the contextual representation with information from the preceding and following text. : in, This indicates a vector concatenation operation.

[0038] Since the input features explicitly encode evaluation roles and evaluation orientation information, BiLSTM can, during the modeling process,: strengthen the temporal dependencies between terms related to the core evaluation, suppress the interference of terms that only perform tone modification or emotion regulation functions, and maintain the structural independence of different evaluation segments in scenarios with multiple evaluation dimensions. The contextual feature sequence output by this step, which integrates the preceding and following contextual information, is represented as follows: : in, This represents the context feature vector output by BiLSTM at sequence position t; Represents the context feature vector The sequence serves as the input feature for subsequent steps in the structured decoding of teaching evaluation entities.

[0039] S5. Based on the context feature sequence, construct the state feature function and the transition feature function. In the label transition matrix of the conditional random field decoding process, introduce the evaluation context constraint parameter to perform weighted modulation on the label transition matrix. After decoding, output the optimal entity label sequence as the teaching evaluation entity label result corresponding to each term.

[0040] In this embodiment of the invention, this step introduces a label transfer constraint mechanism for teacher evaluation texts on the basis of the traditional conditional random field decoding mechanism, so that the entity decoding process is constrained by the evaluation context and evaluation structure, rather than relying solely on statistical transfer probabilities.

[0041] During the CRF (Conditional Random Field) decoding process, for each position t in the sequence, a state feature function is constructed to characterize the matching relationship between the current position's context features and the candidate entity label; a transition feature function is constructed to characterize the neighboring labels. Towards The rationale for the transfer. The overall conditional probability is expressed as... : in, is the normalization factor; A represents the label transition matrix; Indicates label Transfer to label The corresponding transfer score; When the sequence position t is labeled The corresponding state score at that time.

[0042] To address the entity structure characteristics of teacher evaluation texts, this invention introduces evaluation context constraint parameters into the CRF transition matrix to modulate label transition paths that do not conform to the pragmatic structure of teaching evaluation. Specifically: for label transition paths that do not conform to the norms of teacher evaluation entity structure, their transition weights are reduced; for label sequences that conform to the evaluation context structure, their transition probabilities are increased. These constraints are implemented through weighted adjustments to the transition matrix parameters, rather than through hard rule masking, thereby maintaining the model's trainability and generalization ability. Under the combined influence of the aforementioned state features, transition features, and evaluation context constraints, the Viterbi algorithm is used to search for the entity label sequence with the highest conditional probability. : The output entity label sequence is the teaching evaluation entity label result corresponding to each word in the teacher's comment text.

[0043] S6. Receive the obtained context feature sequence and entity label sequence, construct the evaluation behavior state vector, determine the evaluation behavior type based on the evaluation behavior state vector, reconstruct the evaluation behavior state, introduce institutionalized teaching evaluation expression state constraints to verify the legality of the reconstructed evaluation behavior state, and then map the reconstructed evaluation behavior state to desensitized evaluation paragraph text. After passing the teaching evaluation consistency test, output the desensitization result.

[0044] In this embodiment of the invention, the context feature sequence output in S4 and the entity label sequence output in S5 are received as inputs for this step. The entity label sequence is used to identify the entity category and boundary information of each term in the teacher's comments, and participates in the location of the evaluation object and the construction of semantic role binding relationships in the evaluation segment, thereby providing a structured semantic foundation for subsequent evaluation behavior state modeling and sensitive segment identification.

[0045] This step does not directly perform word replacement, sentence rewriting, or deletion operations on the original text. Instead, it uses the reconstruction of the intermediate representation of the evaluation behavior state as the core technical means. By parametrically reconstructing the "evaluation dimension—evaluation orientation—evaluation role binding relationship" in the evaluation behavior, it achieves technical desensitization of privacy risk evaluation. Unlike existing desensitization methods based on keyword filtering, replacement, or general text generation, this invention models the desensitization process as a mapping problem in the evaluation behavior state space. That is, while keeping the evaluation dimension and teaching objectives unchanged, it performs controlled transfer of the evaluation orientation state, thereby achieving the technical resolution of privacy risks.

[0046] Specifically, the input is the evaluation segment state representation constructed from the context feature sequence output in S4, where the context feature sequence has been integrated with the introduced evaluation dimensions, evaluation orientation, and semantic role constraint information. The state representation includes the identified teaching evaluation entities and their corresponding entity type labels. The entity type labels are derived from the entity label sequence output in S5 and are used to identify potential evaluation objects and their semantic category structures in the evaluation segment. Negative evaluation candidate entity types are used to identify evaluation segment locations that may involve privacy risks, but they are not directly used as triggering conditions for desensitization mapping. Instead, they constitute the evaluation behavior state representation composed of the following intermediate state vectors: a. Contextual semantic representation of the evaluation segment; b. Evaluation dimension vector output by the semantic enhancement module; c. Evaluation target orientation vector; d. Semantic role binding matrix. The above information collectively constitutes the evaluation behavior state vector. : in, Indicates the status of the evaluation dimension. Indicates the evaluation direction state. This indicates the evaluation of the role binding relationship. It represents the semantic representation of the context at the segment level.

[0047] The system is based on the evaluation behavior state vector The system determines the type of evaluation behavior for each located evaluation segment. This type of evaluation behavior refers to the structured classification of the pragmatic attributes of evaluation behaviors in teacher comments, used to characterize the comprehensive semantic features of evaluation behaviors in terms of their target audience, stability tendency, and functional attributes. This is reflected in the evaluation behavior state vector. During the construction process, the entity label sequence output by the decoding layer participates in the localization of evaluation objects and the generation of semantic role binding relationships, thereby affecting the formation of evaluation behavior state features and providing structured semantic support for subsequent evaluation behavior type determination. The system first bases its evaluation behavior state vector on... Extract evaluation-oriented state features Evaluation of stability state characteristics and evaluation of functional status characteristics ,in, a. Evaluation of state characteristics Used to determine whether the evaluation directly targets the individual student; b. Evaluate stability state characteristics Used to determine whether an evaluation has a tendency towards fixed abilities or qualitative personality traits; c. Evaluate functional state characteristics Used to distinguish between descriptive, judgmental, or developmental guidance-based evaluation behaviors.

[0048] Subsequently, the aforementioned state features are fused to form an evaluation behavior judgment feature vector. : in, This indicates a vector concatenation operation. The system will evaluate the behavior-determining feature vectors. Input evaluation behavior classification function Obtain evaluation behavior type label : in, The evaluation behavior classification model can be implemented using a multilayer perceptron classifier, a fully connected neural network classifier, or other supervised learning classification models. The evaluation behavior type label includes at least: (1) Individual attribute negation type; (2) Behavioral outcome-based blaming; (3) Capability-finalizing suggestion type; (4) Development-oriented; (5) Neutral descriptive.

[0049] The evaluation behavior type label is used for subsequent privacy risk assessment and evaluation-directed state mapping processes. When the evaluation behavior type belongs to a preset privacy risk type, the system triggers an evaluation-directed state reconstruction mechanism.

[0050] Based on the semantic role binding matrix, the set of role relationships in the evaluation behavior is parsed to determine whether the evaluation object is a student individual entity and whether the entity carries the function of negative evaluation; when the evaluation behavior type is determined to have privacy risks, the evaluation-to-state mapping operation is performed, and at the same time, the semantic role binding relationship is rebounded. The system is based on a semantic role binding matrix. Analyze the set of role relationships in evaluation behavior : in, Represents the semantic role type of the k-th class. This represents the entity node corresponding to sequence position t; This represents the binding pairs between semantic roles and evaluation entities. Through this set of relationships, it is determined whether the evaluation object is a student individual entity and whether that entity carries a negative evaluation function.

[0051] When evaluating behavior type When a system determines that a privacy risk exists, it performs an evaluation-to-state mapping operation. This operation does not modify the evaluation dimension state. Only for the evaluation of the state Perform controlled mapping: Among them, the parameterized vector mapping function Used to direct individual students to the state vector Mapping to a state as a behavioral process or a state as a developmental stage Mapping can be achieved through linear transformations, nonlinear transformations, or attention mechanisms to maintain the evaluation dimension unchanged and eliminate direct individual orientations. Simultaneously, the system adjusts the evaluation orientation state based on the mapped state. semantic role binding matrix The binding relationships related to the evaluation object are reconstructed. Specifically, while keeping the evaluation role type unchanged, the correspondence between the evaluation role and the evaluation carrier is redefined, so that the evaluation carrier relationship originally bound to the student individual entity is migrated to the behavioral process entity or developmental stage entity, generating a reconstructed semantic role binding matrix. : in, The role binding reconstruction function adjusts the binding relationship between the evaluation role and the evaluation object based on the mapped evaluation orientation state, without changing the evaluation role type or evaluation dimension state. This process only changes the binding relationship between the role and the evaluation object, without deleting or adding evaluation role types. Through the above state mapping and role rebinding, the anonymization layer completes the reconstruction of the evaluation behavior structure in the intermediate representation space, migrating negative evaluations originally carried by individual students to the behavioral process or development space, thus eliminating direct individual privacy targeting from a technical perspective.

[0052] After reconstructing the evaluation behavior state, the system introduces institutionalized teaching evaluation expression constraints to the reconstructed state. Legality is defined. Institutionalized teaching evaluation does not store specific vocabulary or sentence templates; instead, it limits the expressive structure of evaluation behavior through state-based constraints. Specific constraints include: a. Evaluation Starting State Constraints: The system checks whether the starting state of the evaluation behavior points to a learning context or development stage. If not, the system remaps or adjusts the starting state vector to point to a valid learning context or development stage representation.

[0053] b. Evaluation of core state constraints: The system checks whether the core state of the evaluated behavior is consistent with the original evaluation dimension. If the core state deviates from the original dimension, it is adjusted to a vector representation consistent with the original evaluation dimension through a state correction function.

[0054] c. Evaluate convergence state constraints: The system checks whether the convergence state of the evaluated behavior is improvement- or development-oriented. If not, it adjusts the state to be improvement- or development-oriented through the constraint mapping function.

[0055] The constraint acts as a state filtering function on the reconstructed evaluation behavior state. Output status that conforms to the standardized expression norms of institutionalized teaching evaluation. By establishing institutionalized constraints on teaching evaluation expression, the system ensures that the expression of evaluation behavior, after anonymization, maintains its teaching evaluation function while conforming to institutionalized evaluation norms.

[0056] After state reconstruction and constraint verification are completed, the system maps the evaluation behavior state to a natural language expression. This process is accomplished by a controlled natural language generation module, whose input is the reconstructed evaluation behavior state. The output is an anonymized evaluation text. The following technical constraints are applied during the generation process: a. Introducing new evaluation dimensions is prohibited; b. Adding new negative judgment states is prohibited; c. Strengthening individual student label states is prohibited. The generation module only performs the mapping operation from state to language and does not participate in the evaluation logic decision-making, thus ensuring the controllability and repeatability of the anonymization process.

[0057] The generated results must pass a consistency check module. The check includes: a. consistency of evaluation dimensions before and after anonymization; b. whether the evaluation focus has shifted from individual-oriented to development-oriented; and c. whether the pragmatic constraints of teaching evaluation are met. The consistency check function is expressed as: Results that fail the test will be fed back to the evaluation-directed state mapping process for reprocessing.

[0058] By modeling the desensitization process as a controlled mapping problem in the state space of evaluation behavior, this invention avoids the destruction of the educational semantics of teacher comments by general text desensitization methods. It achieves structured, interpretable, and verifiable privacy protection processing of sensitive evaluation content, thus ensuring the safety of students' privacy while maintaining the educational evaluation function and application value of teacher comments.

[0059] This invention provides a privacy protection system for unstructured teacher comments, including the following: The input module collects raw teacher comment texts from different teaching scenarios and performs preprocessing and standardization. After initial entity annotation, it performs candidate trigger marking on evaluation-related segments in the teacher comment texts to generate evaluation entity candidate marking information and standardized teacher comment text sequences. The embedding module converts the standardized teacher comment text sequence into a vectorized input format for semantic encoding. It extracts the contextual dependencies between terms through multi-layer attention processing to obtain a contextual semantic vector sequence. The semantic enhancement module generates a comment-level semantic representation by pooling and non-linear mapping based on the context semantic vector sequence, identifies the evaluation behavior, and generates an evaluation dimension vector and an evaluation target pointing vector; it predefines a set of semantic role types and constructs a role embedding matrix for each role type; it calculates the matching score between terms and each semantic role and normalizes it to obtain the role distribution, generates a role-aware enhanced representation, and introduces the implicit evaluation object embedding vector; it applies role constraints to subjective modifier terms and ambiguous terms, and outputs an enhanced semantic feature sequence. The encoding module, based on the output enhanced semantic feature sequence, uses a bidirectional long short-term memory network to model temporal dependencies from both the forward and backward directions, concatenates the bidirectional hidden states, and outputs a context feature sequence that integrates the preceding and following contextual information. The decoding module constructs state feature functions and transition feature functions based on context feature sequences. In the label transition matrix of the conditional random field decoding process, evaluation context constraint parameters are introduced to weight and modulate the label transition matrix. After decoding, the optimal entity label sequence is output as the teaching evaluation entity label result corresponding to each term. The desensitization module receives the obtained context feature sequence and entity label sequence, constructs an evaluation behavior state vector, determines the evaluation behavior type based on the evaluation behavior state vector, reconstructs the evaluation behavior state, introduces institutionalized teaching evaluation expression state constraints to verify the legality of the reconstructed evaluation behavior state, and then maps the reconstructed evaluation behavior state to desensitized evaluation paragraph text. After passing the teaching evaluation consistency test, the desensitization result is output.

Claims

1. A privacy protection method for unstructured teacher comments, characterized in that, Includes the following steps: Step S1. Collect the original teacher comment texts in different teaching scenarios and perform preprocessing and standardization. After preliminary entity annotation, perform candidate trigger marking on the evaluation-related segments in the teacher comment texts to generate evaluation entity candidate marking information and standardized teacher comment text sequences. Step S2. Convert the standardized teacher comment text sequence into a vectorized input format for semantic encoding. Extract the contextual dependencies between terms through multi-layer attention processing to obtain a contextual semantic vector sequence. Step S3. Based on the context semantic vector sequence, perform pooling and non-linear mapping to generate a comment-level semantic representation, identify the evaluation behavior, and generate an evaluation dimension vector and an evaluation target pointing vector; A predefined set of semantic role types is defined, and a role embedding matrix is ​​constructed for each role type. The matching scores between terms and each semantic role are calculated and normalized to obtain the role distribution, generating a role perception enhancement representation and introducing an implicit evaluation object embedding vector. Role constraints are applied to subjective modification terms and ambiguous terms, and an enhanced semantic feature sequence is output. Step S4. Based on the output enhanced semantic feature sequence, a bidirectional long short-term memory network is used to perform temporal dependency modeling from both the forward and backward directions, concatenate the bidirectional hidden states, and output a context feature sequence that integrates the preceding and following context information. Step S5. Based on the context feature sequence, construct the state feature function and the transition feature function. In the label transition matrix of the conditional random field decoding process, introduce the evaluation context constraint parameter to perform weighted modulation on the label transition matrix. After decoding, output the optimal entity label sequence as the teaching evaluation entity label result corresponding to each term. Step S6. Receive the obtained context feature sequence and entity label sequence, construct the evaluation behavior state vector, determine the evaluation behavior type based on the evaluation behavior state vector, reconstruct the evaluation behavior state, introduce institutionalized teaching evaluation expression state constraints to verify the legality of the reconstructed evaluation behavior state, and then map the reconstructed evaluation behavior state to desensitized evaluation paragraph text. After passing the teaching evaluation consistency test, output the desensitization result.

2. The privacy protection method for unstructured teacher comments according to claim 1, characterized in that, In step S3, the implicit evaluation object embedding vector is introduced for cases where the evaluation object does not appear explicitly in the teacher's comments. When no explicit evaluation object role is detected in the role distribution, the system injects the implicit evaluation object embedding vector into the evaluation structure representation space as a semantic anchor point between the evaluation behavior and the potential evaluated object.

3. The privacy protection method for unstructured teacher comments according to claim 1, characterized in that, In step S3, the subjective modifier terms are determined based on the pre-constructed evaluation modifier term list or the output of the role classifier. The role distribution is constrained to be normalized only within the evaluation adjustment role subspace. For ambiguous terms with multiple potential semantic functions, the role distribution calculation is always completed under the joint constraint of the evaluation dimension vector and the evaluation target pointing vector.

4. A privacy protection method for unstructured teacher comments according to claim 1, characterized in that, In step S6, the evaluation behavior type is determined for the located evaluation segment based on the evaluation behavior state vector, according to the joint distribution of the following three types of state features, as shown below: a. Evaluate the characteristics of the target state: Used to determine whether the evaluation directly targets the individual student; b. Evaluation of stability characteristics: used to determine whether the evaluation has a tendency towards fixed abilities or qualitative personality traits; c. Evaluation functional state characteristics: used to distinguish between descriptive, judgmental, or developmental guidance evaluation behaviors; Then, after fusing the above three types of state features, the evaluation behavior type label is output.

5. A privacy protection method for unstructured teacher comments according to claim 1, characterized in that, The specific process of reconstructing the evaluation behavior state in step S6 is as follows: Based on the semantic role binding matrix, the set of role relationships in the evaluation behavior is analyzed to determine whether the evaluation object is a student individual entity and whether the entity carries the function of negative evaluation. When the evaluation behavior type is determined to have privacy risks, the evaluation pointing to state mapping operation is performed. At the same time, the semantic role binding relationship is rebounded.

6. A privacy protection method for unstructured teacher comments according to claim 1, characterized in that, The institutionalized teaching evaluation expression constraints mentioned in step S6 are specifically described below: The expression structure of evaluation behavior is limited by state constraints, including: a. Evaluation starting state constraints, which check whether the starting state of the evaluation behavior points to the learning situation or development stage; b. Evaluation core state constraints, which check whether the core state of the evaluation behavior is consistent with the original evaluation dimension; c. Evaluation closing state constraints, which check whether the closing state of the evaluation behavior is improvement or development-oriented. The expression constraints of institutionalized teaching evaluation act as a state filtering function on the reconstructed evaluation behavior state.

7. A privacy protection method for unstructured teacher comments according to claim 1, characterized in that, In step S6, the evaluation behavior state vector is constructed by combining the contextual semantic representation of the evaluation segment, the evaluation dimension vector, the evaluation target pointing vector, and the semantic role binding matrix.

8. A privacy protection method for unstructured teacher comments according to claim 1, characterized in that, In step S3, the system calculates the matching score of the term with each semantic role under the constraint of the evaluation context, obtains the distribution of the term under each role after normalization, and performs weighted fusion of the role embedding based on the role distribution to generate the role-aware representation of the term. At the same time, it is fused with the original semantic representation of the term to obtain the enhanced semantic representation.

9. A privacy protection method for unstructured teacher comments according to claim 1, characterized in that, In step S2, the standardized teacher comment text sequence is converted into a vectorized input format adapted to the pre-trained language model and semantically encoded. The vectorized input includes term embedding, position embedding, and segment embedding, which are used to represent the basic semantic information of terms in the teacher comment text and their relative position information in the text.

10. A privacy protection system for unstructured teacher comments, applied to the privacy protection method for unstructured teacher comments as described in claim 1, characterized in that, include: The input module collects raw teacher comment texts from different teaching scenarios and performs preprocessing and standardization. After initial entity annotation, it performs candidate trigger marking on evaluation-related segments in the teacher comment texts to generate evaluation entity candidate marking information and standardized teacher comment text sequences. The embedding module converts the standardized teacher comment text sequence into a vectorized input format for semantic encoding. It extracts the contextual dependencies between terms through multi-layer attention processing to obtain a contextual semantic vector sequence. The semantic enhancement module generates a comment-level semantic representation by pooling and non-linear mapping based on the context semantic vector sequence, identifies the evaluation behavior, and generates an evaluation dimension vector and an evaluation target pointing vector. A predefined set of semantic role types is defined, and a role embedding matrix is ​​constructed for each role type. The matching scores between terms and each semantic role are calculated and normalized to obtain the role distribution, generating a role perception enhancement representation and introducing an implicit evaluation object embedding vector. Role constraints are applied to subjective modification terms and ambiguous terms, and an enhanced semantic feature sequence is output. The encoding module, based on the output enhanced semantic feature sequence, uses a bidirectional long short-term memory network to model temporal dependencies from both the forward and backward directions, concatenates the bidirectional hidden states, and outputs a context feature sequence that integrates the preceding and following contextual information. The decoding module constructs state feature functions and transition feature functions based on context feature sequences. In the label transition matrix of the conditional random field decoding process, evaluation context constraint parameters are introduced to weight and modulate the label transition matrix. After decoding, the optimal entity label sequence is output as the teaching evaluation entity label result corresponding to each term. The desensitization module receives the obtained context feature sequence and entity label sequence, constructs an evaluation behavior state vector, determines the evaluation behavior type based on the evaluation behavior state vector, reconstructs the evaluation behavior state, introduces institutionalized teaching evaluation expression state constraints to verify the legality of the reconstructed evaluation behavior state, and then maps the reconstructed evaluation behavior state to desensitized evaluation paragraph text. After passing the teaching evaluation consistency test, the desensitization result is output.

Images