Adolescent learning emotion recognition and adaptive language content generation method

Abstract

The present application relates to the field of artificial intelligence and education technology, and discloses a learning emotion recognition and adaptive language content generation method for teenagers, which comprises: analyzing a motivation state vector corresponding to an emotion category label and an emotion category intensity, wherein the motivation state vector is a vector of motivation state, and the motivation state comprises self-efficacy, achievement motivation, goal orientation, expected value, self-determination, external reward sensitivity, reason attribution tendency, learned helplessness, growth mindset and motivation intensity moderate value; based on the motivation state vector, a language generation chain of the teenager is constructed, wherein the language generation chain comprises each intervention node and node sequence; and adaptive language content corresponding to the motivation state vector is generated at each intervention node of the language generation chain. The present application can improve the recognition ability of the motivation behind the emotion, and solve the defects that the pacification response is superficial and it is difficult to accurately alleviate the deep learning resistance psychology.

Description

Technical Field

[0001] This invention relates to a method for recognizing emotions and generating adaptive language content in adolescent learning, belonging to the fields of artificial intelligence and education technology. Background Technology

[0002] In the field of language learning for teenagers, with the development of educational technology, intelligent systems have begun to attempt to provide adaptive feedback by recognizing learners' emotional states. These systems are usually based on multimodal data collected from learners, such as analyzing facial expressions through cameras, analyzing speech through microphones, and analyzing text through interaction logs, thereby identifying emotions such as anxiety, boredom, or resistance. Building on this, existing technologies further explore linking emotional states with learning motivation. Motivation is a complex psychological construct that encompasses multiple dimensions, including an individual's confidence in their own abilities, the intrinsic drive to pursue success, and the assessment of the value of a task and the probability of success. The goal of introducing these psychological concepts is to enable systems to go beyond simple emotional responses and achieve deeper, personalized interventions aimed at maintaining and enhancing learning motivation.

[0003] However, in the process of transforming emotion recognition into effective motivational intervention, existing technologies suffer from a mismatch between intervention strategies and learners' complex and dynamic intrinsic motivational states. This results in superficial effects of reassurance and encouragement, failing to fundamentally alleviate deep-seated learning resistance. Specifically, most existing systems rely on pre-set, static rule bases for mapping emotions and responses. For example, when a system identifies high-intensity anxiety, it might mechanically trigger a pre-set, generic encouraging phrase. The drawback of this approach is that it fails to understand and respond to the different motivational roots behind the same emotion. For instance, anxiety may stem from fear of task difficulty or worry about negative evaluation. Due to a lack of sophisticated diagnostic capabilities for the multidimensional motivational states behind emotions, existing technologies cannot distinguish these intrinsic differences and can only provide a one-size-fits-all response. This general response fails to address the true root cause of learners' resistance, not only failing to effectively soothe emotions but also potentially being perceived as a formulaic perfunctory response by learners due to its lack of specificity. Consequently, it fails to achieve the ultimate goal of accurately responding to emotional needs and enhancing learning initiative.

[0004] Therefore, existing technologies suffer from insufficient ability to identify the motivations behind emotions, resulting in superficial reassurance responses that fail to accurately alleviate deep-seated learning resistance and effectively enhance learning initiative. Summary of the Invention

[0005] This invention provides a method for recognizing emotions and generating adaptive language content in adolescent learning. Its main purpose is to improve the ability to identify the motivations behind emotions and to solve the problem that soothing responses are superficial and fail to accurately alleviate deep-seated learning resistance.

[0006] To achieve the above objectives, this invention provides a method for adolescent learning emotion recognition and adaptive language content generation, comprising: When adolescents are learning a language, based on their facial expressions, speech, and text information, the system identifies the adolescents' emotion category labels and emotion category intensity. Analyze the motivational state vectors corresponding to the emotion category labels and the intensity of the emotion categories. The motivational state vectors are vectors of motivational states, which include self-efficacy, achievement motivation, goal orientation, expected value, self-determination, sensitivity to external rewards, attribution tendency, learned helplessness, growth mindset, and appropriate value of motivational intensity. Based on the motivational state vector, the language generation chain of the adolescent is constructed, wherein the language generation chain includes various intervention nodes and node order, and each intervention node includes a primary intervention node, a secondary intervention node, and a closing intervention node. Adaptive language content corresponding to the motivational state vector is generated at each intervention node of the language generation chain.

[0007] Optionally, the analysis of the motivational state vector corresponding to the emotion category label and the emotion category intensity includes: The emotion category label and the emotion category intensity are combined to form an emotion input vector, wherein the number of dimensions of the emotion input vector is equal to the number of emotion categories in the emotion category label, and each dimension value in the emotion input vector corresponds to the emotion category intensity. The emotion input vector is input into a motivational diagnostic neural network, wherein the number of neurons in the input layer of the motivational diagnostic neural network is equal to the number of dimensions of the emotion input vector, and the number of neurons in the output layer of the motivational diagnostic neural network is equal to the number of motivational states. The motivational state vector corresponding to the emotion input vector is obtained through the motivational diagnostic neural network, wherein each dimension value of the motivational state vector represents the activation intensity of each motivational state.

[0008] Optionally, obtaining the motivational state vector corresponding to the emotion input vector through the motivational diagnostic neural network includes: In the hidden layer of the motivation diagnosis neural network, the emotion input vector is nonlinearly transformed by a leaky rectified linear unit to obtain a hidden representation vector. The hidden representation vector is mapped to the output layer of the motivational diagnostic neural network; The hidden representation vector is converted into an inactive vector by using the linear weighting unit and the bias addition unit of the output layer. A range pruning operation is performed on each dimension value of the inactive vector to obtain the motivation state vector.

[0009] Optionally, the step of performing a range pruning operation on each dimension value of the inactive vector to obtain the motivational state vector includes: Identify the current dimension name for each dimension value of the inactive vector; Query the upper and lower threshold values ​​corresponding to the current dimension name in the dimension-threshold lookup table, wherein the dimension-threshold lookup table is determined by the acceptable fluctuation range of each motivation state; If the dimension value of the inactive vector is greater than the upper limit threshold, the dimension value of the inactive vector is set to the upper limit threshold to determine the first value range clipping dimension value; If the dimension value of the inactive vector is less than the lower threshold, the dimension value of the inactive vector is set to the lower threshold to determine the second value range clipping dimension value. The first domain clipping dimension value and the second domain clipping dimension value are recombined to obtain the motivation state vector.

[0010] Optionally, constructing the adolescent's language generation chain based on the motivational state vector includes: The motivational state vector is input into the strategy planner; In the strategy planner, the three dimensions with the highest activation intensity in the motivational state vector are queried; The three intervention nodes corresponding to the three dimension values ​​are matched sequentially. The three intervention nodes are arranged in descending order of activation intensity to determine the node order among the intervention nodes; The language generation chain is determined by the various intervention nodes and their order; wherein, the intervention nodes include strategy keywords and target intents.

[0011] Optionally, before constructing the adolescent's language generation chain based on the motivational state vector, the method further includes: Obtain the strategy description text; extract keywords from the strategy description text to obtain the core strategy corpus; The core corpus of the strategy is vectorized to obtain vectorized corpus; The vectorized corpus is stored in the policy embedding pool to generate policy-vector key-value pairs; The policy-vector key-value pairs are classified and mapped using the 1-nearest neighbor assignment algorithm to obtain the policy-motivation mapping table.

[0012] Optionally, generating adaptive language content corresponding to the motivational state vector at each intervention node of the language generation chain includes: The language generation chain is combined with the motivational state vector to form system prompt words; The system prompts are input into the large language model; In the large language model, adaptive language content covering each intervention node is generated based on the system prompts.

[0013] Optionally, in the large language model, generating adaptive language content covering each intervention node based on the system prompt words includes: In the large language model, initial language content covering each intervention node is generated based on the system prompt words; Identify the coverage of strategy keywords in the initial language content; When the coverage rate of the strategy keywords is lower than the preset coverage rate, the missing keywords are supplemented in the initial language content through the large language model to obtain adaptive language content.

[0014] Optionally, the step of identifying the adolescent's emotion category label and emotion category intensity based on the adolescent's facial expression information, voice information, and text information includes: The facial expression information, the voice information, and the text information are respectively input into the corresponding emotion recognition sub-model to output the preliminary emotion category and preliminary emotion intensity; The three sets of preliminary emotion categories and preliminary emotion intensities corresponding to the facial expression information, the voice information, and the text information are weighted and fused to determine the emotion category label and the emotion category intensity.

[0015] To address the aforementioned problems, this invention also provides a system for adolescent learning emotion recognition and adaptive language content generation, the system comprising: An emotion recognition module is used to identify the emotion category label and emotion category intensity of teenagers when they are learning a language, based on their facial expression information, voice information and text information. The motivation analysis module is used to analyze the motivational state vectors corresponding to the emotion category labels and the intensity of the emotion categories. The motivational state vectors are vectors of motivational states, which include self-efficacy, achievement motivation, goal orientation, expected value, self-determination, sensitivity to external rewards, attribution tendency, learned helplessness, growth mindset, and appropriate value of motivation intensity. The language chain construction module is used to construct the language generation chain of the adolescent based on the motivation state vector. The language generation chain includes various intervention nodes and node order. Each intervention node includes a primary intervention node, a secondary intervention node, and a closing intervention node. The content generation module is used to generate adaptive language content corresponding to the motivational state vector at each intervention node of the language generation chain.

[0016] Compared to the problems described in the background art, the embodiments of the present invention, by integrating three complementary information sources—facial expressions, voice, and text—overcome the limitations of single-modal perception, thereby more accurately capturing the learner's true and complex emotional state. Furthermore, by mapping surface emotions to the core, interveneable psychological dimension of learning motivation, the embodiments of the present invention essentially provide a psychological diagnosis, moving beyond simply responding to emotions to understanding the psychological causes behind them. Further, by planning multiple diagnosed motivational issues into an intervention path map with a clear sequence, the embodiments of the present invention ensure that the system output is not a collection of scattered, potentially conflicting comforting statements, but a strategic communication sequence with a logical progression. Furthermore, by transforming abstract intervention strategies and specific motivational states into natural, fluent, and warm dialogues in the learner's ears, the embodiments of the present invention ensure that the final output strictly adheres to the preset psychological intervention logic while possessing high linguistic flexibility and approachability, avoiding a mechanical and template-like feel. Therefore, the present invention can improve the ability to identify motivations behind emotions and solve the shortcomings of superficial comforting responses that fail to accurately alleviate deep-seated learning resistance. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for adolescent learning emotion recognition and adaptive language content generation according to an embodiment of the present invention. Figure 2 This is a system flowchart of a method for adolescent learning emotion recognition and adaptive language content generation according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a module for implementing the adolescent learning emotion recognition and adaptive language content generation system according to an embodiment of the present invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] This application provides a method for recognizing and adaptively generating language content for adolescent learning. The subject of this method includes, but is not limited to, at least one computer device that can be configured to perform the method provided in this application, such as a server or a terminal. In other words, the method can be implemented by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0021] Reference Figure 1 The diagram shown is a flowchart illustrating a method for recognizing and adaptively generating language content for adolescent learning, provided in an embodiment of the present invention. In this embodiment, the method includes: S1. When adolescents are learning a language, based on the adolescents' facial expressions, speech, and text information, identify the adolescents' emotion category labels and emotion category intensity.

[0022] By integrating three complementary information sources—facial expressions, speech, and text—the embodiments of this invention can overcome the limitations of single-modal perception, thereby more accurately capturing the learner's real and complex emotional state.

[0023] The facial expression information refers to the sequence of facial images of teenagers captured by a camera; the voice information refers to the voice signals of teenagers speaking captured by a microphone; the text information refers to the text content entered by teenagers during the learning interaction; the emotion category label refers to the discrete category name assigned to the learner's current emotional state after the fusion analysis of the three types of information: facial expression, voice, and text, including anxiety, boredom, resistance, etc.; and the emotion category strength refers to the normalized floating-point value assigned to each emotion category label, used to quantify the activation degree of the emotion at the current moment, with a value range of 0 to 1, and the larger the value, the stronger the emotion.

[0024] In one embodiment of the present invention, the step of identifying the emotion category label and emotion category intensity of the adolescent based on the adolescent's facial expression information, voice information, and text information includes: inputting the facial expression information, the voice information, and the text information into the corresponding emotion recognition sub-model to output preliminary emotion category and preliminary emotion intensity; and performing weighted fusion on the three sets of preliminary emotion categories and preliminary emotion intensities corresponding to the facial expression information, the voice information, and the text information to determine the emotion category label and emotion category intensity.

[0025] The emotion recognition sub-model refers to a dedicated machine learning model for extracting and classifying emotion features for a single modality. For example, a ResNet-50 convolutional network for facial expression images outputs the probabilities of seven emotion categories, such as anxiety and boredom, and their 0-1 intensities after inputting an RGB facial image; a Bi-LSTM-Attention model for speech signals outputs the emotion category and its 0-1 intensity after inputting an MFCC plus fundamental frequency and energy frame sequence; and a RoBERTa-base model with a fully connected classification head for text content outputs emotion logits after inputting word segmentation and obtains the 0-1 intensity through SoftMax.

[0026] Optionally, the process of weighted fusion of the three preliminary emotion categories and preliminary emotion intensities corresponding to the facial expression information, the voice information, and the text information is as follows: each modality has a preset reliability weight coefficient. For the same emotion category, the three preliminary emotion intensities are multiplied by their corresponding weights and then summed to obtain the fusion intensity of the emotion category. The emotion category with the highest fusion intensity is taken as the final emotion category label, and its fusion intensity is normalized to the [0, 1] interval to obtain the final emotion category intensity.

[0027] S2. Analyze the motivational state vector corresponding to the emotion category label and the emotion category intensity, wherein the motivational state vector is a vector of motivational states, and the motivational states include self-efficacy, achievement motivation, goal orientation, expected value, self-determination, sensitivity to external rewards, attribution tendency, learned helplessness, growth mindset and appropriate value of motivation intensity.

[0028] This invention maps surface emotions to the core, interventionist psychological dimension of learning motivation, which is equivalent to providing a psychological diagnosis, enabling us to move beyond simply responding to emotions and instead understand the psychological causes behind them.

[0029] Herein, self-efficacy refers to an individual's subjective confidence in their ability to successfully complete learning tasks; achievement motivation refers to an individual's intrinsic drive to pursue excellent performance and surpass others; goal orientation refers to an individual's value orientation in learning, focusing on mastering knowledge itself or on demonstrating abilities to gain praise; expected value refers to an individual's comprehensive assessment of the probability of success and importance of a task; self-determination refers to an individual's sense of autonomy, competence, and connection in learning activities; external reward sensitivity refers to an individual's sensitivity to external incentives such as grades, prizes, or praise; attribution tendency refers to the way an individual typically attributes academic success or failure to internal or external factors such as ability, effort, difficulty, or luck; learned helplessness refers to an individual's psychological state of losing the belief in controlling outcomes and giving up effort due to repeated failures; growth mindset refers to an individual's belief that abilities can be continuously improved through effort and a willingness to embrace challenges; and the appropriate level of motivation intensity refers to the ideal intensity at which an individual's current learning motivation is neither too high nor too low, ensuring continued engagement.

[0030] In one embodiment of the present invention, analyzing the motivational state vector corresponding to the emotion category label and the emotion category intensity includes: combining the emotion category label and the emotion category intensity into an emotion input vector, wherein the number of dimensions of the emotion input vector is equal to the number of emotion categories in the emotion category label, and each dimension value in the emotion input vector corresponds to the emotion category intensity; inputting the emotion input vector into a motivational diagnostic neural network, wherein the number of neurons in the input layer of the motivational diagnostic neural network is equal to the number of dimensions of the emotion input vector, and the number of neurons in the output layer of the motivational diagnostic neural network is equal to the number of motivational states; obtaining the motivational state vector corresponding to the emotion input vector through the motivational diagnostic neural network, wherein each dimension value of the motivational state vector represents the activation intensity of each motivational state.

[0031] The motivation diagnosis neural network refers to a feedforward neural network that connects the input layer, hidden layer and output layer with a feedforward structure. Its input dimension is equal to the number of emotion categories, and its output dimension is fixed as ten-dimensional motivational states. Through the collaborative calculation of leakage rectified linear units, linear weighting units and bias summing units, the emotion input vector is converted into a normalized motivational state vector.

[0032] In another embodiment of the present invention, obtaining the motivational state vector corresponding to the emotion input vector through the motivational diagnostic neural network includes: performing a nonlinear transformation on the emotion input vector in the hidden layer of the motivational diagnostic neural network using a leaky rectified linear unit to obtain a hidden representation vector; mapping the hidden representation vector to the output layer of the motivational diagnostic neural network; converting the hidden representation vector into an inactive vector through a linear weighting unit and a bias addition unit in the output layer; and performing a range pruning operation on each dimension value of the inactive vector to obtain the motivational state vector.

[0033] The leaky rectified linear unit refers to the activation function in the hidden layer used to perform nonlinear transformation on the emotion input vector. It retains a small slope in the negative interval to allow a small number of negative values ​​to pass through, thereby enhancing the network's ability to express subtle emotions and improving training stability. The linear weighting unit and bias addition unit refers to the computational component in the output layer that maps the hidden representation vector to a ten-dimensional motivational state vector. The unactivated linear result is obtained by adding the weight matrix multiplication to the learnable bias vector. The hidden layer refers to the intermediate network layer located between the input layer and the output layer, used to extract high-order abstract representations of emotion features and reduce the risk of overfitting.

[0034] In another embodiment of the present invention, the step of performing a value range pruning operation on each dimension value of the inactive vector to obtain a motivational state vector includes: identifying the current dimension name of each dimension value of the inactive vector; querying an upper limit threshold and a lower limit threshold corresponding to the current dimension name in a dimension-threshold lookup table, wherein the dimension-threshold lookup table is determined by the acceptable fluctuation range of each motivational state; if the dimension value of the inactive vector is greater than the upper limit threshold, setting the dimension value of the inactive vector to the upper limit threshold to determine a first value range pruning dimension value; if the dimension value of the inactive vector is less than the lower limit threshold, setting the dimension value of the inactive vector to the lower limit threshold to determine a second value range pruning dimension value; and recombining the first value range pruning dimension value and the second value range pruning dimension value to obtain a motivational state vector.

[0035] The current dimension name refers to the Chinese name of the motivational state corresponding to each dimension in the inactive vector, used for key-value lookup in the dimension-threshold lookup table. The dimension-threshold lookup table is a key-value mapping table pre-set by educational psychology professionals according to the acceptable fluctuation range of each motivational state. Its primary key is the current dimension name, and the value range is the lower threshold and the upper threshold. It is used to prevent abnormal activation values ​​caused by extreme emotions from entering subsequent strategy planning. The first value range pruning dimension value refers to the dimension value of the inactive vector set to the upper threshold, and the second value range pruning dimension value refers to the dimension value of the inactive vector set to the lower threshold.

[0036] Optionally, the process of recombining the first domain clipped dimension value and the second domain clipped dimension value to obtain the motivation state vector is as follows: sequentially traverse all dimensions of the inactive vector, perform upper or lower threshold clipping on each dimension, and then write the clipping results into a new ten-dimensional array in the original dimension order to complete the recombination. The final ten-dimensional array is the normalized and boundary-safe motivation state vector.

[0037] S3. Based on the motivational state vector, construct the language generation chain of the adolescent, wherein the language generation chain includes various intervention nodes and node order, and each intervention node includes a primary intervention node, a secondary intervention node, and a closing intervention node.

[0038] This invention, through its embodiments, plans multiple diagnosed motivational issues into an intervention path diagram with a clear sequence, ensuring that the system outputs not a bunch of scattered or even potentially conflicting comforting statements, but a strategic communication sequence with a logical progression.

[0039] In one embodiment of the present invention, constructing the language generation chain of the adolescent based on the motivational state vector includes: inputting the motivational state vector into a strategy planner; querying the strategy planner for the three dimensions with the highest activation intensity in the motivational state vector; sequentially matching the three intervention nodes corresponding to the three dimension values; arranging the three intervention nodes in descending order of activation intensity to determine the node order among the intervention nodes; and determining the language generation chain through the intervention nodes and the node order; wherein the intervention nodes include strategy keywords and target intentions.

[0040] The strategy planner refers to a decision-making module that receives a ten-dimensional motivational state vector and automatically generates an ordered intervention sequence based on activation intensity. It embeds a Top 3 intensity query function and a strategy-motivation mapping table to output a linked list of intervention nodes with psychological priorities. The strategy-motivation mapping table is a data dictionary with dimension names as keys and corresponding intervention strategy objects as values. It is generated by vectorizing and classifying the core strategy corpus and is used to quickly map any dimension name in the motivational state vector to a unique intervention node, thereby determining the strategy keywords and target intentions of each node in the language generation chain. The strategy keywords are those pre-annotated by educational psychology experts in the intervention nodes and must appear in the generated text to achieve the corresponding intervention. The core vocabulary set for predicting effects, such as the language generation chain [Primary Strategy: Value Reshaping], [Secondary Strategy: Attribution Adjustment], and [Closing Strategy: Goal Minimization], refers to the primary, secondary, and closing intervention nodes, respectively. Value reshaping, attribution adjustment, and goal minimization are the key strategies. The goal of value reshaping is to change students' perception of the value of the task. The goal of attribution adjustment is to guide students to attribute failure to controllable factors. The goal of goal minimization is to help students set easily achievable short-term goals to build confidence. The goal intention refers to the specific psychological adjustment purpose that the intervention node expects to achieve.

[0041] Optionally, the process of sequentially matching the three intervention nodes corresponding to the three dimension values ​​is as follows: the strategy planner sequentially traverses the Top3 list, searches for the corresponding intervention node object in the strategy-motivation mapping table using the dimension name as the key, and each object contains a strategy keyword and a target intent field to complete a one-to-one binding. Further, the process of arranging the three intervention nodes in descending order of activation intensity to determine the node order among the intervention nodes is as follows: the strategy planner inserts the three bound intervention nodes into the chain container in descending order of the original intensity of the Top3 list, first the strongest, then the second strongest, and finally the third strongest, thereby forming a fixed-order node sequence in memory. Further, the process of determining the language generation chain through the intervention nodes and the node order is as follows: the strategy planner encapsulates the chain container into a language generation chain instance, exposing an ordered list of nodes and the strategy keywords and target intents inside the nodes.

[0042] In another embodiment of the present invention, before constructing the language generation chain of the adolescent based on the motivation state vector, the method further includes: obtaining strategy description text; extracting keywords from the strategy description text to obtain strategy core corpus; vectorizing the strategy core corpus to obtain vectorized corpus; storing the vectorized corpus in a strategy embedding pool to generate strategy-vector key-value pairs; and classifying and mapping the strategy-vector key-value pairs using a 1-nearest neighbor allocation algorithm to obtain a strategy-motivation mapping table.

[0043] The strategy description text refers to raw natural language material written by educational psychology experts to describe the key points and expected effects of a particular intervention strategy. Its content typically includes a strategy definition, applicable scenarios, operational suggestions, and intervention goals. This material serves as the input source for extracting the core strategy corpus. The core strategy corpus refers to a set of high-weighted words or phrases representing the core semantics of the strategy, obtained from the strategy description text using a keyword extraction algorithm. This set is used for vectorization and classification in subsequent steps to establish a mapping relationship between the strategy and motivational dimensions. The keyword extraction algorithm, for example, is TF-IDF. TextRank, BERT-keyword extraction, etc., the policy embedding pool refers to a persistent data structure used to store policy-vector key-value pairs. It uses the policy name as the key and the vectorized representation of the policy core corpus as the value. It supports fast reading and similarity calculation, and provides a data foundation for the generation of 1-nearest neighbor assignment classification and policy-motivation mapping table. The 1-nearest neighbor assignment algorithm refers to a parameterless mapping method that finds the nearest reference vector in the set of labeled reference vectors for the vector to be classified, and directly assigns the label of the reference vector to the vector to be classified, thereby completing a one-to-one classification.

[0044] For example, the process of obtaining strategy description text; extracting keywords from the strategy description text to obtain core strategy corpus; vectorizing the core strategy corpus to obtain vectorized corpus; storing the vectorized corpus in a strategy embedding pool to generate strategy-vector key-value pairs; and classifying and mapping the strategy-vector key-value pairs using a 1-nearest neighbor assignment algorithm to obtain a strategy-motivation mapping table is exemplified as follows: First, the strategy description text pre-written by educational psychology experts is read, and after performing sentence segmentation, stop word removal, and part-of-speech tagging, a keyword extraction algorithm is used to extract high-weight words to form the core strategy corpus; subsequently, The core policy corpus is encoded into fixed-dimensional floating-point vectors using a pre-trained language model, such as BERT or Sentence-BERT, to obtain vectorized corpus. This vectorized corpus, with policy names as keys and vector values ​​as values, is batch-written into the policy embedding pool to form policy-vector key-value pairs. Finally, using the expert anchor vectors corresponding to the names of each dimension in the ten-dimensional motivation state vector as label centers, 1-nearest neighbor assignment is performed on all vectors in the policy embedding pool, assigning the nearest policy-vector key-value pairs to the corresponding dimensions, thereby generating a policy-motivation mapping table for the policy planner to quickly call during runtime.

[0045] S4. Generate adaptive language content corresponding to the motivational state vector at each intervention node of the language generation chain.

[0046] This invention transforms abstract intervention strategies and specific motivational states into natural, fluent, and warm dialogues that resonate with learners. This ensures that the final output strictly adheres to the pre-set psychological intervention logic while possessing high linguistic flexibility and approachability, avoiding a mechanical and template-like feel.

[0047] In one embodiment of the present invention, generating adaptive language content corresponding to the motivational state vector at each intervention node of the language generation chain includes: combining the language generation chain and the motivational state vector into a system prompt word; inputting the system prompt word into a large language model; and generating adaptive language content covering each intervention node in the large language model based on the system prompt word.

[0048] The system prompts refer to structured text instructions assembled from the order of intervention nodes in the language generation chain, the policy keywords of each node, and the key-value pairs of the motivational state vector according to a preset template. These instructions are used to clarify the roles, task order, and necessary semantic identifiers for the large language model, guiding the model to output text that conforms to the intervention strategy in a strongly constrained manner. Examples of system prompts include: Task order: Node 1 Value reshaping, Node 2 Attribution adjustment, Node 3 Goal minimization; Node 1 keywords: Meaning, future benefits; Node 2 keywords: Effort, controllable; Node 3 keywords: Small steps, feasible; Motivational state vector: {Self-efficacy: 0.82, Achievement motivation: 0.82, ... Machine: 0.65,…}; Requirements: Generate a coherent dialogue in sequence, with each node containing corresponding keywords. The large language model refers to a pre-trained generative Transformer network deployed in the system. After receiving system prompts, it outputs a coherent dialogue text at once and can generate compensation sentences based on supplementary prompts when keywords are missing. This is used to achieve adaptive language content generation and correction. The adaptive language content refers to the final dialogue text generated by the large language model under strong constraints of system prompts and after keyword coverage verification and missing word compensation. It covers all intervention nodes, contains all policy keywords, and can be directly presented to learners.

[0049] In another embodiment of the present invention, generating adaptive language content covering each intervention node in the large language model based on the system prompt words includes: generating initial language content covering each intervention node in the large language model based on the system prompt words; identifying the coverage rate of strategy keywords in the initial language content; and when the coverage rate of strategy keywords is lower than a preset coverage rate, supplementing missing keywords in the initial language content through the large language model to obtain adaptive language content.

[0050] The initial language content refers to the original dialogue text generated after the large language model first receives the system prompt words. It is used for subsequent policy keyword coverage verification. If the coverage meets the standard, it is directly output as adaptive language content. If it is insufficient, it enters the compensation stage. The policy keyword coverage refers to the ratio of the number of policy keywords that actually appear in the initial language content to the total number of policy keywords specified by the system prompt words. It is used to quantify the semantic fulfillment of the preset intervention strategy by the generated text. The preset coverage refers to the minimum acceptable keyword hit ratio set by the system for the intervention node. The default value in the current embodiment is 100%. When the policy keyword coverage is lower than this threshold, supplementary generation is triggered to ensure that the intervention intention is fully conveyed.

[0051] See Figure 2 The diagram shown is a system flowchart of a method for adolescent learning emotion recognition and adaptive language content generation according to an embodiment of the present invention. Figure 2First, facial expressions, speech, and text are input into the emotion sub-model. The preliminary emotion category and intensity output by the emotion sub-model are weighted and fused to obtain the final emotion label and intensity. This process belongs to S1 above. Second, the motivational state vector corresponding to the emotion label and intensity is output through the motivational diagnosis neural network. This belongs to S2 above. Third, based on the motivational state vector, the language generation chain of the adolescent is constructed using the policy planner. This belongs to S3 above. Finally, the motivational state vector output by the motivational diagnosis neural network and the language generation chain output by the policy planner are used as input data for the large language model. The large language model outputs adaptive language content. This belongs to S4 above.

[0052] Compared to the problems described in the background art, the embodiments of the present invention, by integrating three complementary information sources—facial expressions, voice, and text—overcome the limitations of single-modal perception, thereby more accurately capturing the learner's true and complex emotional state. Furthermore, by mapping surface emotions to the core, interveneable psychological dimension of learning motivation, the embodiments of the present invention essentially provide a psychological diagnosis, moving beyond simply responding to emotions to understanding the psychological causes behind them. Further, by planning multiple diagnosed motivational issues into an intervention path map with a clear sequence, the embodiments of the present invention ensure that the system output is not a collection of scattered, potentially conflicting comforting statements, but a strategic communication sequence with a logical progression. Furthermore, by transforming abstract intervention strategies and specific motivational states into natural, fluent, and warm dialogues in the learner's ears, the embodiments of the present invention ensure that the final output strictly adheres to the preset psychological intervention logic while possessing high linguistic flexibility and approachability, avoiding a mechanical and template-like feel. Therefore, the present invention can improve the ability to identify motivations behind emotions and solve the shortcomings of superficial comforting responses that fail to accurately alleviate deep-seated learning resistance.

[0053] like Figure 3 The diagram shown is a functional module diagram of a system for recognizing emotions and generating adaptive language content for adolescent learning, based on the present invention.

[0054] The adolescent learning emotion recognition and adaptive language content generation system 300 described in this invention can be installed in a computer device. Depending on the functions implemented, the adolescent learning emotion recognition and adaptive language content generation system may include an emotion recognition module 301, a motivation analysis module 302, a language chain construction module 303, and a content generation module 304. The module described in this invention can also be called a unit, referring to a series of computer program segments that can be processed by the computer device's processor and perform a fixed function, stored in the computer device's memory.

[0055] In this embodiment of the invention, the functions of each module / unit are as follows: The emotion recognition module 301 is used to identify the emotion category label and emotion category intensity of the adolescent based on the adolescent's facial expression information, voice information and text information when the adolescent is learning a language. The motivation analysis module 302 is used to analyze the motivation state vector corresponding to the emotion category label and the emotion category intensity. The motivation state vector is a vector of motivation states, which include self-efficacy, achievement motivation, goal orientation, expected value, self-determination, sensitivity to external rewards, attribution tendency, learned helplessness, growth mindset, and appropriate value of motivation intensity. The language chain construction module 303 is used to construct the language generation chain of the adolescent based on the motivation state vector, wherein the language generation chain includes various intervention nodes and node order, and each intervention node includes a primary intervention node, a secondary intervention node and a closing intervention node. The content generation module 304 is used to generate adaptive language content corresponding to the motivation state vector at each intervention node of the language generation chain.

[0056] In detail, the modules in the adolescent learning emotion recognition and adaptive language content generation system 300 described in this embodiment of the invention employ the same methods as described above. Figure 1 The techniques used in the adolescent learning emotion recognition and adaptive language content generation methods described herein are the same and can produce the same technical effects, so they will not be elaborated here.

[0057] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0058] Finally, it should be noted that in the above embodiments, each embodiment can be combined with each other or independent. Deleting any one of them will not affect the technical implementation of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not 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 technical solutions of the present invention.

Claims

1. A method for emotion recognition and adaptive language content generation in adolescent learning, characterized in that, The method includes: When adolescents are learning a language, based on their facial expressions, speech, and text information, the system identifies the adolescents' emotion category labels and emotion category intensity. Analyze the motivational state vectors corresponding to the emotion category labels and the intensity of the emotion categories. The motivational state vectors are vectors of motivational states, which include self-efficacy, achievement motivation, goal orientation, expected value, self-determination, sensitivity to external rewards, attribution tendency, learned helplessness, growth mindset, and appropriate value of motivational intensity. Based on the motivational state vector, the language generation chain of the adolescent is constructed, wherein the language generation chain includes various intervention nodes and node order, and each intervention node includes a primary intervention node, a secondary intervention node, and a closing intervention node. Adaptive language content corresponding to the motivational state vector is generated at each intervention node of the language generation chain.

2. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 1, characterized in that, The analysis of the motivational state vector corresponding to the emotion category label and the emotion category intensity includes: The emotion category label and the emotion category intensity are combined to form an emotion input vector, wherein the number of dimensions of the emotion input vector is equal to the number of emotion categories in the emotion category label, and each dimension value in the emotion input vector corresponds to the emotion category intensity. The emotion input vector is input into a motivational diagnostic neural network, wherein the number of neurons in the input layer of the motivational diagnostic neural network is equal to the number of dimensions of the emotion input vector, and the number of neurons in the output layer of the motivational diagnostic neural network is equal to the number of motivational states. The motivational state vector corresponding to the emotion input vector is obtained through the motivational diagnostic neural network, wherein each dimension value of the motivational state vector represents the activation intensity of each motivational state.

3. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 2, characterized in that, The step of obtaining the motivational state vector corresponding to the emotional input vector through the motivational diagnostic neural network includes: In the hidden layer of the motivation diagnosis neural network, the emotion input vector is nonlinearly transformed by a leaky rectified linear unit to obtain a hidden representation vector. The hidden representation vector is mapped to the output layer of the motivational diagnostic neural network; The hidden representation vector is converted into an inactive vector by using the linear weighting unit and the bias addition unit of the output layer. A range pruning operation is performed on each dimension value of the inactive vector to obtain the motivation state vector.

4. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 3, characterized in that, The step of performing a range pruning operation on each dimension value of the inactive vector to obtain the motivational state vector includes: Identify the current dimension name for each dimension value of the inactive vector; Query the upper and lower threshold values ​​corresponding to the current dimension name in the dimension-threshold lookup table, wherein the dimension-threshold lookup table is determined by the acceptable fluctuation range of each motivation state; If the dimension value of the inactive vector is greater than the upper limit threshold, the dimension value of the inactive vector is set to the upper limit threshold to determine the first value range clipping dimension value; If the dimension value of the inactive vector is less than the lower threshold, the dimension value of the inactive vector is set to the lower threshold to determine the second value range clipping dimension value. The first domain clipping dimension value and the second domain clipping dimension value are recombined to obtain the motivation state vector.

5. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 1, characterized in that, The construction of the adolescent's language generation chain based on the motivational state vector includes: The motivational state vector is input into the strategy planner; In the strategy planner, the three dimensions with the highest activation intensity in the motivational state vector are queried; The three intervention nodes corresponding to the three dimension values ​​are matched sequentially. The three intervention nodes are arranged in descending order of activation intensity to determine the node order among the intervention nodes; The language generation chain is determined by the various intervention nodes and their order; wherein, the intervention nodes include strategy keywords and target intents.

6. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 1, characterized in that, Before constructing the adolescent's language generation chain based on the motivational state vector, the process also includes: Obtain the strategy description text; extract keywords from the strategy description text to obtain the core strategy corpus; The core corpus of the strategy is vectorized to obtain vectorized corpus; The vectorized corpus is stored in the policy embedding pool to generate policy-vector key-value pairs; The policy-vector key-value pairs are classified and mapped using the 1-nearest neighbor assignment algorithm to obtain the policy-motivation mapping table.

7. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 1, characterized in that, The step of generating adaptive language content corresponding to the motivational state vector at each intervention node of the language generation chain includes: The language generation chain is combined with the motivational state vector to form system prompt words; The system prompts are input into the large language model; In the large language model, adaptive language content covering each intervention node is generated based on the system prompts.

8. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 7, characterized in that, In the large language model, based on the system prompts, adaptive language content covering each intervention node is generated, including: In the large language model, initial language content covering each intervention node is generated based on the system prompt words; Identify the coverage of strategy keywords in the initial language content; When the coverage rate of the strategy keywords is lower than the preset coverage rate, the missing keywords are supplemented in the initial language content through the large language model to obtain adaptive language content.

9. The method for adolescent learning emotion recognition and adaptive language content generation as described in claim 1, characterized in that, The process of identifying the emotional category label and intensity of the adolescent based on their facial expression information, voice information, and text information includes: The facial expression information, the voice information, and the text information are respectively input into the corresponding emotion recognition sub-model to output the preliminary emotion category and preliminary emotion intensity; The three sets of preliminary emotion categories and preliminary emotion intensities corresponding to the facial expression information, the voice information, and the text information are weighted and fused to determine the emotion category label and the emotion category intensity.

10. A system for recognizing emotions and generating adaptive language content in adolescent learning, characterized in that, The system includes: An emotion recognition module is used to identify the emotion category label and emotion category intensity of teenagers when they are learning a language, based on their facial expression information, voice information and text information. The motivation analysis module is used to analyze the motivational state vectors corresponding to the emotion category labels and the intensity of the emotion categories. The motivational state vectors are vectors of motivational states, which include self-efficacy, achievement motivation, goal orientation, expected value, self-determination, sensitivity to external rewards, attribution tendency, learned helplessness, growth mindset, and appropriate value of motivation intensity. The language chain construction module is used to construct the language generation chain of the adolescent based on the motivation state vector. The language generation chain includes various intervention nodes and node order. Each intervention node includes a primary intervention node, a secondary intervention node, and a closing intervention node. The content generation module is used to generate adaptive language content corresponding to the motivational state vector at each intervention node of the language generation chain.

Images