A teaching intention recognition method based on teaching situation perception and pseudo-label optimization
By constructing teaching scenarios to enhance representation and optimizing pseudo-label screening, the problems of scarce labeled samples and semantic ambiguity in classroom teaching intention recognition are solved, and stable and reliable recognition is achieved under low labeling conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-02
AI Technical Summary
In real classroom teaching, the identification of teachers' teaching intentions is unstable due to the scarcity of labeled samples, semantic ambiguity, and context dependence, which leads to unreliable pseudo-labels and the accumulation of noise.
By integrating the semantic features of teachers' discourse, the contextual features of teachers and students, and prior teaching metadata, an enhanced representation of the teaching context is constructed. Confidence constraints and cross-iteration consistency checks are used to filter out pseudo-labels and optimize the model training process.
It improved the accuracy and stability of teaching intention recognition, enhanced the model's adaptability and generalization ability to complex classroom contexts, and suppressed the accumulation of noise and false labels.
Smart Images
Figure CN122132834A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence education application technology, and relates to a teaching intention recognition method based on teaching context perception and pseudo-label optimization. Background Technology
[0002] Instructional intent refers to the teaching objectives and strategies demonstrated by teachers through language and teaching behaviors during classroom instruction. It is a crucial component of teaching process analysis, teaching quality evaluation, and teacher professional development research. Accurately identifying teachers' instructional intent helps in deepening understanding of the classroom teaching process, supports teaching quality assessment and teaching behavior analysis, and is of great significance in smart classrooms and intelligent educational applications.
[0003] In real classroom teaching scenarios, the expression of teachers' teaching intentions often exhibits significant context dependence and semantic ambiguity. The same teacher's utterances may correspond to different teaching intentions at different teaching stages, in different subject backgrounds, or in different teacher-student interaction contexts. Furthermore, different teachers, due to differences in teaching styles and expression habits, exhibit significant variations in the form of their utterances and the way they express their intentions. This makes it easy to generate ambiguity and misjudgment when relying solely on isolated utterances or partial semantic information for teaching intention identification. In addition, teaching intention identification tasks commonly face the problem of scarce labeled samples in practical applications. Labeling real classroom teaching corpora requires the participation of personnel with educational backgrounds; the labeling process is complex, costly, and time-consuming, making it difficult to obtain large-scale, high-quality labeled data. Under conditions of insufficient labeled samples, supervised learning-based teaching intention identification models are prone to overfitting, and their generalization ability and recognition stability are difficult to guarantee.
[0004] To alleviate the scarcity of labeled samples, some existing methods attempt to introduce semi-supervised learning or pseudo-label learning strategies, using unlabeled samples to assist model training. However, in the scenario of recognizing teaching intentions, due to the strong ambiguity and context-dependent nature of teacher utterance semantics, pseudo-labels that have not been effectively filtered often contain a lot of noise. As training iterations continue to accumulate, this noise weakens model performance, leading to unstable recognition results. Furthermore, existing methods generally ignore the macro-level teaching context information contained in teaching metadata, such as subject attributes, grade level characteristics, teacher information, and class sessions. They fail to effectively integrate these prior teaching contexts with the semantics and contextual information of teacher utterances, limiting the model's ability to understand complex classroom teaching contexts.
[0005] Therefore, there is an urgent need for a technical solution that can fully integrate teaching context information under conditions of scarce labeled samples, and suppress noise accumulation through an effective pseudo-label optimization mechanism, thereby improving the accuracy and stability of recognizing real classroom teaching intentions. Summary of the Invention
[0006] In view of this, the purpose of this invention is to provide a teaching intention recognition method based on teaching context perception and pseudo-label optimization, so as to solve the problem that under the condition of scarce labeled samples in real classroom education data, the semantic ambiguity and strong context dependence of teachers' speech make pseudo-labels unreliable and noise easily accumulates, which leads to confusion of intention representation and insufficient recognition stability. Thus, it can achieve stable and reliable recognition of teachers' teaching intentions under low labeling conditions.
[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying teaching intentions based on teaching context perception and pseudo-label optimization, comprising: Acquire teacher discourse data and the corresponding teacher-student context sequence, and simultaneously acquire teaching metadata; the teaching metadata includes subject information, grade level information, teacher gender information, and classroom video identification information; Semantic encoding is performed on teacher discourse and teacher-student context sequences, and situation modeling is performed based on the context sequences to obtain semantic features of teacher discourse and context features reflecting teacher-student interaction relationships. The teaching metadata is mapped to contextual prior representations and fused with the semantic features of teacher discourse and contextual features to construct an enhanced representation of the teaching context. The teaching context augmentation representation is used as input, and the teaching intention recognition model is trained based on labeled samples. Then, the trained teaching intention recognition model is used to generate pseudo-labels for unlabeled samples. The pseudo-labels are sequentially screened by confidence constraints and cross-iteration consistency checks to obtain highly reliable pseudo-label samples and participate in model updates. The enhanced representation of the teaching context corresponding to the teacher's discourse to be identified is input into the updated teaching intention recognition model, and the teaching intention recognition result of the teacher's discourse is output.
[0008] Furthermore, acquiring teacher discourse data and the corresponding teacher-student context sequences includes, for the target teacher discourse to be predicted... Construct a context window that includes the preceding and following statements from both the teacher and the student:
[0009] In the formula, The radius of the context window; As a context sequence.
[0010] Semantic encoding of teacher discourse and teacher-student context sequences includes encoding any discourse within the context sequence. A pre-trained BERT semantic encoder is used to obtain sentence-level semantic representations, and the [CLS] vector is used as the semantic feature of the utterance:
[0011] A bidirectional long short-term memory network is used to model the semantic feature sequence to obtain the contextual features of the target teacher's discourse: .
[0012] Furthermore, the teaching metadata is mapped into trainable embedding vectors to form a prior representation of the teaching context:
[0013] In the formula, , , and These represent the embedding vectors corresponding to the grade level, subject, teacher gender, and classroom video identifier information, respectively.
[0014] By splicing and fusing the semantic features of teacher discourse, contextual features, and prior representations of the teaching context, an enhanced representation of the teaching context is constructed:
[0015] In the formula, To enhance representation in teaching contexts, Semantic features of teacher discourse as a target. It is a contextual feature.
[0016] Furthermore, the confidence-constrained screening of pseudo-labels includes introducing a category-aware adaptive threshold strategy, assigning a dynamic pseudo-label screening threshold to each category based on the difficulty of category learning. Indicates the first In the next iteration, the category The average prediction confidence is updated by smoothing the class difficulty using momentum:
[0017] In the formula, The momentum coefficient is used; the category threshold is calculated using the following formula:
[0018] In the formula, Based on the threshold, This is the adjustment coefficient; For a given unlabeled sample, the instructional intent recognition model in the _ In the next iteration, the predicted probability that the unlabeled sample belongs to a different category is: , , For the number of categories of teaching intentions, if the following conditions are met:
[0019] Then add the sample to the pseudo-label candidate set. ;in For category The corresponding category threshold.
[0020] Cross-iteration consistency verification and filtering of pseudo-labels includes filtering the candidate set of pseudo-labels. The samples in the data maintain their past The prediction history of each iteration; Indicates sample No. Class probability distribution of the next iteration Indicates the highest probability A set consisting of category indexes; the overlap ratio of the sets is used to characterize the consistency between the current prediction set and the historical prediction set:
[0021] The consistency scores were then aggregated using a time-weighted method to obtain the sample-level semantic stability score.
[0022] In the formula, Time weighting; when stability score When the threshold is exceeded, the sample will be... Included in the final pseudo-tag set .
[0023] Furthermore, during the model update process, a joint objective function of supervised and unsupervised loss is used to optimize the teaching intent recognition model. The joint objective function is expressed as:
[0024] In the formula, These are the weighting coefficients used to balance supervised loss and unsupervised loss; Monitoring losses Defined as the cross-entropy loss of the labeled sample batch:
[0025] In the formula, To label sample batches, For the number of categories of teaching intentions, For the first The labeled sample in the first The actual label indicator on the class, For the model to the first The predicted value for the nth sample is... The probability of a class; Unsupervised loss The mean squared error loss is defined as the difference between the predicted distribution of the unlabeled sample batch and the pseudo-label.
[0026] In the formula, For samples without batch number labeling For the model to the first The category score vectors output by each unlabeled sample. This refers to the label vector corresponding to the highly reliable pseudo-labels obtained through pseudo-label filtering. It is an L2 norm.
[0027] The beneficial effects of this invention are as follows: (1) By integrating the semantic features of teachers’ discourse, the contextual features of teachers and students and the prior information of teaching metadata, we can construct an enhanced representation of teaching context, which can improve the separability of teaching intention representation and adaptability to complex classroom contexts, and reduce category confusion caused by semantic ambiguity and context differences.
[0028] (2) Under the condition of scarce labeled samples, pseudo-labels are generated by unlabeled samples and used for training. Without increasing the cost of manual labeling, the effective training signal sources can be expanded, thereby improving the training effectiveness and generalization ability of the model.
[0029] (3) The pseudo-label optimization mechanism, which combines confidence constraint screening and cross-iteration consistency verification screening, effectively suppresses the accumulation of noise pseudo-labels with iteration, and improves the stability of the training process and the reliability of multi-class decision-making under low label conditions.
[0030] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of a teaching intent recognition method based on teaching context perception and pseudo-label optimization provided in an embodiment of the present invention. Detailed Implementation
[0032] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0033] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0034] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0035] like Figure 1 As shown, this is an embodiment of the teaching intent recognition method based on teaching context perception and pseudo-label optimization provided by the present invention. It addresses the problems of strong context dependence of teacher's discourse intent, semantic ambiguity, and unstable recognition caused by the accumulation of pseudo-label noise in semi-supervised training under the condition of scarce real classroom labeled samples. The method constructs a context-enhanced representation by integrating teacher discourse semantics, teacher-student context, and prior teaching metadata, and selects high-confidence pseudo-labels to participate in model updates by combining confidence constraints and cross-iteration consistency verification, thereby improving the separability of teaching intent representation and recognition stability under low label conditions.
[0036] The teacher's teaching intentions include, but are not limited to: Spark interest: Use emotional or motivating language to arouse students' curiosity and enhance their intrinsic interest in the learning content.
[0037] Emotional support and feedback: Enhance students' self-confidence and maintain a positive emotional experience through encouragement, praise, or emotional reassurance.
[0038] Knowledge delivery: Conveying factual, conceptual, or procedural knowledge through explanation, description, or demonstration.
[0039] Summary and Conclusion: Review and integrate key concepts or learned content to enhance students' understanding.
[0040] Guided thinking: Guide students to engage in analytical, reflective, or higher-order thinking activities through questions or thought-provoking prompts.
[0041] Enhancing participation: Using language to guide students into the classroom interaction process and encourage them to participate in verbal responses or learning activities.
[0042] Classroom management: Organizing, regulating, and controlling classroom order or teaching process through verbal communication.
[0043] Promote collaborative learning: Through structured guidance, encourage students to engage in peer discussions, collaborative inquiry, and joint problem-solving.
[0044] In this embodiment, the method is as follows: I. Obtain teacher discourse data and the corresponding teacher-student context sequence, and simultaneously obtain teaching metadata, which includes subject, grade level, teacher gender, and classroom video identification information.
[0045] Teacher discourse on the prediction of goals Construct a context window that includes the preceding and following statements from both the teacher and the student:
[0046] in, This is the radius of the context window.
[0047] Second, semantic encoding is performed on the teacher's discourse and the teacher-student context sequence, and situation modeling is performed based on the context sequence to obtain the semantic features of the teacher's discourse and the context features reflecting the teacher-student interaction relationship.
[0048] For any utterance within the context sequence A pre-trained BERT semantic encoder is used to obtain sentence-level semantic representations, and the [CLS] vector is used as the semantic feature of the utterance: ,
[0049] To capture the temporal dependencies and teacher-student interaction relationships within the discourse sequence within the context window, a bidirectional long short-term memory network (BiLSTM) is used to model the semantic feature sequence, obtaining the contextual features of the target teacher's discourse:
[0050] Step 3: Map the teaching metadata to contextual prior representations and integrate them with the semantic features of teacher discourse and contextual features to construct an enhanced representation of the teaching context.
[0051] Specifically, firstly, the teaching metadata is mapped into trainable embedding vectors, which then form a prior representation of the teaching context:
[0052] in, , , and These represent the embedding vectors corresponding to the grade level, subject, teacher gender, and classroom video identifier information, respectively.
[0053] By splicing and fusing the semantic features and contextual features of the target teacher's discourse with the prior representation of the teaching context, an enhanced representation of the teaching context is constructed:
[0054] in The semantic features of teacher discourse as a target.
[0055] Fourth, the teaching context enhancement representation is used as input to train the teaching intention recognition model based on labeled samples. Then, the trained teaching intention recognition model is used to generate pseudo-labels for unlabeled samples.
[0056] The sample mentioned above is text data obtained by collecting teachers' voices and converting them into text.
[0057] Fifth, the pseudo-labels are sequentially screened by confidence constraints and cross-iteration consistency checks to obtain highly reliable pseudo-label samples and participate in model updates.
[0058] Specifically, a category-aware adaptive thresholding strategy is introduced, assigning a dynamic pseudo-label filtering threshold to each category based on the difficulty of category learning. Let... Indicates the first In the next iteration, the category The average prediction confidence is updated by smoothing the class difficulty using momentum:
[0059] in The momentum coefficient, Difficulty level is categorized by category.
[0060] Define the category threshold as:
[0061] in, Based on the threshold, This is the adjustment coefficient.
[0062] For a given unlabeled sample, the model at the... In the next iteration, the predicted probability that the unlabeled sample belongs to a different category is: , , For the number of categories of teaching intentions, if the following conditions are met:
[0063] Then add the sample to the pseudo-label candidate set. .in For category The corresponding category threshold.
[0064] For the pseudo-label candidate set The samples in the data maintain their past The prediction history of each iteration. Let... Indicates sample No. Class probability distribution of the next iteration Indicates the highest probability A set consisting of category indices. The overlap ratio of the sets is used to characterize the consistency between the current prediction set and the historical prediction set:
[0065] The consistency scores were then aggregated using a time-weighted method to obtain the sample-level semantic stability score.
[0066] in, Time weighting.
[0067] Finally, when the stability score When the threshold is exceeded, the sample will be... Included in the final pseudo-tag set .
[0068] During model training, a joint objective function of supervised and unsupervised loss is used to optimize the instructional intent recognition model. The joint objective function is expressed as:
[0069] in, These are the weighting coefficients used to balance supervised loss and unsupervised loss.
[0070] Monitoring losses Defined as the cross-entropy loss of the labeled sample batch:
[0071] in, To label sample batches, For the number of categories of teaching intentions, For the first The labeled sample in the first The actual label indicator on the class, For the model to the first The predicted value for the nth sample is... The probability of a class; Unsupervised loss The mean squared error loss is defined as the difference between the predicted distribution of the unlabeled sample batch and the pseudo-label.
[0072] in, For samples without batch number labeling For the model to the first The category score vectors output by each unlabeled sample. This refers to the label vector corresponding to the highly reliable pseudo-labels obtained through pseudo-label filtering. It is an L2 norm.
[0073] VI. Input the enhanced representation of the teaching context corresponding to the teacher's discourse to be identified into the updated teaching intention recognition model, and output the teaching intention recognition results of the teacher's discourse.
[0074] In summary, the teaching intent recognition method based on teaching context perception and pseudo-label optimization provided by this invention can improve the separability of teaching intent representation and its adaptability to complex classroom contexts, and reduce category confusion caused by semantic ambiguity and contextual differences. By generating pseudo-labels from unlabeled samples and participating in training, the effective training signal sources can be expanded without increasing the cost of manual labeling, thereby improving the effectiveness and generalization ability of model training. The pseudo-label optimization mechanism, which combines confidence constraint screening and cross-iteration consistency verification screening, can effectively suppress the accumulation of noisy pseudo-labels with iterations, and improve the stability of the training process and the reliability of multi-class decision-making under low-labeling conditions.
[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for identifying teaching intentions based on teaching context perception and pseudo-label optimization, characterized in that, The method includes: Acquire teacher discourse data and the corresponding teacher-student context sequence, and simultaneously acquire teaching metadata; Semantic encoding is performed on teacher discourse and teacher-student context sequences, and situation modeling is performed based on the context sequences to obtain semantic features of teacher discourse and context features reflecting teacher-student interaction relationships. The teaching metadata is mapped to contextual prior representations and fused with the semantic features of teacher discourse and contextual features to construct an enhanced representation of the teaching context. The teaching context augmentation representation is used as input, and the teaching intention recognition model is trained based on labeled samples. Then, the trained teaching intention recognition model is used to generate pseudo-labels for unlabeled samples. The pseudo-labels are sequentially screened by confidence constraints and cross-iteration consistency checks to obtain highly reliable pseudo-label samples and participate in model updates. The enhanced representation of the teaching context corresponding to the teacher's discourse to be identified is input into the updated teaching intention recognition model, and the teaching intention recognition result of the teacher's discourse is output.
2. The teaching intention recognition method according to claim 1, characterized in that, Obtaining teacher discourse data and the corresponding teacher-student context sequences includes, for the target teacher discourse to be predicted Construct a context window that includes the preceding and following statements from both the teacher and the student: In the formula, The radius of the context window; As a context sequence.
3. The teaching intention recognition method according to claim 2, characterized in that, Semantic encoding is performed on teacher discourse and the teacher-student context sequence, and any discourse within the context sequence is encoded. A pre-trained BERT semantic encoder is used to obtain sentence-level semantic representations, and the [CLS] vector is used as the semantic feature of the utterance: A bidirectional long short-term memory network is used to model the semantic feature sequence to obtain the contextual features of the target teacher's discourse: .
4. The teaching intention recognition method according to claim 1, characterized in that, The teaching metadata is mapped to trainable embedding vectors to form a prior representation of the teaching context: In the formula, , , and These represent the embedding vectors corresponding to the grade level, subject, teacher gender, and classroom video identifier information, respectively.
5. The teaching intention recognition method according to claim 4, characterized in that, By splicing and fusing the semantic features of teacher discourse, contextual features, and prior representations of the teaching context, an enhanced representation of the teaching context is constructed: In the formula, To enhance representation in teaching contexts, Semantic features of teacher discourse as a target. It is a contextual feature.
6. The teaching intention recognition method according to claim 1, characterized in that, Confidence-constrained screening of pseudo-labels includes introducing a category-aware adaptive threshold strategy, assigning a dynamic pseudo-label screening threshold to each category based on the difficulty of category learning. Indicates the first In the next iteration, the category The average prediction confidence is updated by smoothing the class difficulty using momentum: In the formula, The momentum coefficient is used; the category threshold is calculated using the following formula: In the formula, Based on the threshold, This is the adjustment coefficient; For a given unlabeled sample, the instructional intent recognition model in the _ In the next iteration, the predicted probability that the unlabeled sample belongs to a different category is: , , For the number of categories of teaching intentions, if the following conditions are met: Then add the sample to the pseudo-label candidate set. ;in For category The corresponding category threshold.
7. The teaching intention recognition method according to claim 6, characterized in that, Cross-iteration consistency verification and filtering of pseudo-labels includes filtering the candidate set of pseudo-labels. The samples in the data maintain their past The prediction history of each iteration; Indicates sample No. The class probability distribution of the next iteration. Indicates the highest probability A set consisting of category indexes; the overlap ratio of the sets is used to characterize the consistency between the current prediction set and the historical prediction set: The consistency scores were then aggregated using a time-weighted method to obtain the sample-level semantic stability score. In the formula, Time weighting; when stability score When the threshold is exceeded, the sample will be... Included in the final pseudo-tag set .
8. The method according to claim 7, characterized in that, During the model update process, the teaching intention recognition model is optimized using a joint objective function of supervised and unsupervised losses. The joint objective function is expressed as: In the formula, These are the weighting coefficients used to balance supervised loss and unsupervised loss; Monitoring losses Defined as the cross-entropy loss of the labeled sample batch: In the formula, To label sample batches, For the number of categories of teaching intentions, For the first The labeled sample in the first The actual label indicator on the class, For the model to the first The predicted value for the nth sample is... The probability of a class; Unsupervised loss The mean squared error loss is defined as the difference between the predicted distribution of the unlabeled sample batch and the pseudo-label. In the formula, For samples without batch number labeling For the model to the first The category score vectors output by each unlabeled sample. This refers to the label vector corresponding to the highly reliable pseudo-labels obtained through pseudo-label filtering. It is an L2 norm.
9. The teaching intention recognition method according to claim 1, characterized in that, The teaching metadata includes subject information, grade level information, teacher gender information, and classroom video identification information.