Student learning emotion recognition system based on natural language processing
Through the student learning emotion recognition system based on natural language processing, we deeply analyze the subtle emotional characteristics in students' language expressions, construct emotion transfer paths and calculate the emotion fluctuation index, which solves the problem of difficulty in identifying the emotions of students in language subjects in existing technologies, and realizes dynamic monitoring and personalized intervention of emotional changes in the learning process.
Patent Information
- Application Number
- CN202510858552.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-25
- Publication Date
- 2025-10-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing learning process monitoring technologies find it difficult to deeply identify students' subtle emotional characteristics during language learning, especially vague tone, emotional jumps, and hesitant expression. They are unable to accurately identify potential risk states such as continuous frustration and cognitive fluctuations, and lack targeted intelligent guidance.
A student learning emotion recognition system based on natural language processing is adopted, including the first dialogue collection module, emotion recognition module, emotion fluctuation index calculation module, emotion coherence judgment module and intervention strategy generation module. Through multiple rounds of natural language interaction data collection, combined with the Transformer emotion classification model and emotion feature dictionary, the emotion transfer path is constructed and the emotion fluctuation index is calculated to generate personalized intervention strategies.
It has achieved precise identification and dynamic monitoring of students' emotional fluctuations, can timely warn of cognitive fluctuations and potential emotional risks, provide scientific quantitative basis, provide support for teaching management and psychological intervention, and enhance learning experience and mental health protection.
Smart Images

Figure CN120744612A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of natural language processing, and in particular to a student learning emotion recognition system based on natural language processing. Background Art
[0002] When studying language subjects, students often acquire knowledge and develop skills through oral practice, conversations, and written writing. This type of learning is characterized by openness, frequent subjective expression, and diverse feedback. This can easily lead to emotional fluctuations, such as confusion about understanding, difficulty expressing oneself, and anxiety and tension. These emotional fluctuations not only affect learning efficiency but can also further impact self-perception and learning confidence.
[0003] However, existing learning process monitoring technologies mostly focus on behavioral analysis, performance tracking, or objective answer evaluation, lacking in-depth identification and tracking of implicit emotional information in natural language expressions during the learning process. Especially in language-related subjects, traditional methods struggle to capture the subtle emotional characteristics displayed by students during real-world communication, such as ambiguous tone, emotional jumps, and hesitant expression. Furthermore, existing systems often overlook the continuous evolution of emotional states, making it impossible to accurately identify whether students are in potential risk states such as "continuous frustration" or "cognitive fluctuations." They also struggle to provide targeted intelligent guidance based on the specific content of the conversation. Summary of the Invention
[0004] In view of the shortcomings of the existing technology, the present invention provides a student learning emotion recognition system based on natural language processing to solve the problems mentioned in the background technology.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a student learning emotion recognition system based on natural language processing, comprising:
[0006] The first dialogue collection module is used to collect multiple rounds of natural language interaction sentences between students and the learning system or teachers during the language learning process to obtain a first language sequence set;
[0007] a first emotion recognition module, configured to perform context modeling analysis on the first language sequence set, identify emotion expression segments in language sentences, construct a student emotion fluctuation trajectory based on temporal features, and extract a first emotion transfer path;
[0008] The second emotion recognition module is used to extract the understanding confusion feature, language expression discomfort feature, active attempt feature and positive cognitive feedback feature based on the first emotion transfer path, and integrate them into the first emotion transfer path to generate the second emotion transfer path;
[0009] The emotion fluctuation index calculation module is used to assign positive, neutral and negative polarity labels to each fusion label based on the second emotion transfer path, obtain the third polarity sequence, and construct the emotion fluctuation index of the i-th student in the T learning cycle. It is used to measure the frequency of positive and negative emotion switching per unit time. If a student falls into a state of cognitive fluctuation or psychological conflict, the first warning instruction will be triggered;
[0010] The emotional consistency judgment module is used to determine whether there is a "continuous frustration mode" based on the third polarity sequence after receiving the first warning instruction, and to construct the emotional abnormality score index of the i-th student in the T learning cycle. The scoring factors include: negative sentence ratio coefficient α, cognitive contradiction density coefficient β and vocabulary ambiguity rate γ;
[0011] Intervention strategy generation module, based on the abnormal emotion score index of the i-th student in the T learning cycle Conduct assessments to obtain corresponding emotional risk states and generate corresponding intervention strategies.
[0012] Preferably, the first dialogue collection module includes an interactive access unit and a language sequence construction unit;
[0013] The interactive access unit is used to establish a communication connection with the learning platform used by students or the teacher's terminal, access multiple rounds of natural language interactive input channels, collect and obtain natural language input data between the i-th student and the learning system or language subject teacher, and obtain the first language data;
[0014] The language sequence construction unit is used to convert the text input and voice data in the first language data into text form, set a session identifier, perform session recognition on each student's consecutive input sentences, and determine the session boundaries according to the set time interval or chapter switching rules to form a multi-round sentence sequence; pre-process the collected natural language sentences, remove redundant modal particles and unstructured tag information, and screen out sentence content with language subject relevance;
[0015] Removal of redundant modal particles and unstructured tag information includes:
[0016] Remove extra spaces, HTML tags and special symbols;
[0017] Special symbols include: "#", "@" and "$";
[0018] Convert full-width characters to half-width characters, unify Chinese and English punctuation forms, and then unify the encoding format. Add timestamps, student identity information, course and chapter number metadata to each sentence, and generate the first language sequence set of the i-th student containing timestamps and learning metadata.
[0019] Preferably, the first emotion recognition module includes an interjection feature extraction unit and an emotion expression segment recognition unit;
[0020] The modal particle feature extraction unit is used to scan the modal particles in the first language sequence set and use the modal particles as the first emotion amplification feature marker; the modal particle list includes: "ah", "er", "ma", "ya", "ge", "ouch", "wow" and "my god";
[0021] The emotion expression segment recognition unit is used to perform semantic understanding and emotion classification on the sentences in the first language sequence set based on the context information and the Transformer emotion classification model; automatically divide the language segments containing obvious emotion expressions and mark the corresponding emotion categories;
[0022] The steps for the Transformer sentiment classification model to perform semantic understanding and sentiment classification are:
[0023] A context window is formed to extract a single sentence or multiple consecutive sentences from the first language sequence set. The text is segmented and encoded into a token sequence and converted into a token ID.
[0024] The pre-trained Transformer sentiment classification model is used to encode the input token sequence to obtain the contextual semantics of each token. After the Transformer encoding and fully connected layer of the Transformer sentiment classification model, a probability vector P is obtained through softmax.
[0025] P={p1, p2, ..., p n};
[0026] Where n is the number of predefined emotion categories, p j is the predicted probability of the jth category, satisfying:
[0027]
[0028] Find the emotion category j corresponding to the maximum probability * and the maximum probability value p max :
[0029]
[0030] p max =p j*
[0031] Set the confidence threshold R = 0.6, if p max ≥R, then class j * as the final emotion label for the statement;
[0032] If p max <R, it means that the emotion of the sentence is unclear and is labeled as “no obvious emotion”.
[0033] Preferably, the first emotion recognition module further comprises an emotion path generating unit;
[0034] The emotion path generating unit is used to arrange each of the first language sequence combinations in the order of timestamps to form a time-ordered emotion tag sequence S:
[0035] S={(t1, e1), (t2, e2),..., (t m , e m )};
[0036] Among them, t m represents the timestamp or interaction sequence corresponding to the mth statement, e m The final emotion label is obtained by the emotion expression segment recognition unit;
[0037] Assign standardized codes to each emotion category; including:
[0038] Confused C1, anxious C2, happy C3, angry C4, stable C0, and no obvious emotion C5;
[0039] Based on the emotional state group of the sentences at time t and time t+1 in the time series, the emotional state transition frequency matrix is constructed:
[0040]
[0041] in, represents the frequency of emotional state transition between the i-th student at time t and the adjacent time t+1;
[0042] e t+1 Indicates the Cnth emotional state e Cn To the Cbth emotional state e Cb The probability of; P(·) represents probability; the inside = represents logical judgment, which means that at time A at moment t+1, the student’s emotion is e Cn , Condition B At time t, the student's emotion is e Cb ; Cnth emotional state e Cn To the Cbth emotional state e Cb Corresponding to e m It is the final emotion label obtained by the emotion expression fragment recognition unit.
[0043] Preferably, the second emotion recognition module includes a learning emotion feature dictionary construction unit, a language segment screening unit and a second label classification unit;
[0044] The learning emotion feature dictionary construction unit is used to establish a feature dictionary of common emotion representations in language learning scenarios based on the first emotion transfer path, and the dictionary includes the following four categories of emotion feature keyword sets:
[0045] Comprehension confusion characteristics include: “I don’t know,” “What’s going on,” “I don’t understand,” and “I’m confused”;
[0046] Characteristics of language expression discomfort include: "can't say it", "can't pronounce it correctly", "difficult to pronounce" and "this is too complicated";
[0047] Active attempt characteristics include: "I'll try", "I guess so", and "Let me tell you about it";
[0048] Positive cognitive feedback characteristics include: "I see," "I understand," and "It's very simple."
[0049] The language segment screening unit is used to extract the comprehension confusion feature, language expression discomfort feature, active attempt feature, and positive cognitive feedback feature from the sentences in the first language sequence set, and perform a second emotion label discrimination through the second label classification unit in combination with the context information to obtain the second emotion label;
[0050] The second emotional label includes: understanding confusion feature F1, language expression discomfort feature F2, active attempt feature F3 and positive cognitive feedback feature F4.
[0051] Preferably, the second emotion recognition module further includes a fusion unit;
[0052] The fusion unit is used to combine the emotional state corresponding to each sentence in the first emotional transfer path with the second emotional label identified by the second label classification unit to form a second emotional transfer path with attributes; in the second emotional transfer path, each time point has not only an emotional state label, but also behavioral cognitive characteristics as an additional dimension.
[0053] Preferably, the mood fluctuation index calculation module includes a mood polarity determination unit and a mood switching monitoring unit;
[0054] The emotion polarity determination unit is configured to scan the second emotion transfer path, assign positive, neutral, and negative polarity labels to each fusion label, and obtain a third polarity sequence;
[0055] Positive, neutral, and negative polarity markers include:
[0056] Positive polarity markers include: happiness C3 and positive cognitive feedback feature F4;
[0057] Neutral polarity markers include: stable C0 and active attempt feature F3;
[0058] Negative polarity markers include: confusion C1, anxiety C2, understanding confusion feature F1, and language expression discomfort feature F2;
[0059] The emotion switching monitoring unit is used to scan the third polarity sequence and count the number of The number of switching times W, ignoring the switching between neutrals, includes: "negative → neutral → positive" is counted as one switching, that is, W = 1. When "negative → negative" or "positive → neutral → positive" occurs, the number of switching times is not counted, that is, W = 0.
[0060] Preferably, the mood fluctuation index calculation module further includes the mood fluctuation index generation unit and the evaluation unit.
[0061] The mood fluctuation index generating unit is used to extract the mood fluctuation index of the i-th student within a unit time. The number of switching times W is calculated by the following formula to obtain the emotional fluctuation index of the i-th student in the T learning cycle
[0062]
[0063] Among them, W i Indicates the number of students i in the learning cycle The total number of switches between emotional polarities; T g represents the total number of students' words and sentences in T learning cycle; A larger value indicates a more unstable mood;
[0064] The evaluation unit presets the emotion fluctuation threshold B and calculates the emotion fluctuation index of the i-th student in the T learning cycle Comparing with the emotion fluctuation threshold B, the first emotion assessment result is obtained, including:
[0065] when Determining that the student is in a state of cognitive fluctuation or psychological conflict triggers the first warning instruction;
[0066] when It is judged that the student's emotions are relatively stable and the learning state is normal, and the first qualified label is generated.
[0067] Preferably, the emotion coherence judgment module includes a multi-factor score construction unit and an emotion abnormality score generation unit;
[0068] The multi-factor scoring construction unit is used to combine the emotion fluctuation index of the i-th student in the T learning cycle after receiving the first warning instruction. The corresponding students' first language data are extracted to determine whether there are multiple consecutive negative emotion labels. If there are three or more consecutive negative emotion labels, it is determined to be a "continuous frustration mode";
[0069] And calculate the negative sentence ratio coefficient α:
[0070]
[0071] Among them, C neg is the total number of negative emotion sentences in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle;
[0072] To measure whether there are "inconsistent" cognitive expressions, including contradictory sentences such as "I understand" followed by "I still don't understand", or "self-negating" language fragments, the cognitive contradiction density coefficient β is calculated:
[0073]
[0074] Among them, C confloct is the total number of words in cognitively contradictory language segments in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle;
[0075] The proportion of cognitive expressions containing “fuzzy words” is measured. “Fuzzy words” include “almost”, “maybe”, “okay” and “not very clear”. The lexical fuzziness rate γ is calculated as follows:
[0076]
[0077] Among them, C vague is the number of “fuzzy expression words” used by students in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle;
[0078] The emotion abnormality score generation unit is used to extract the negative sentence ratio coefficient α, cognitive contradiction density coefficient β, and vocabulary fuzziness rate γ. After dimensionless processing, the emotion abnormality score index of the i-th student in the T learning cycle is obtained through the weighted formula.
[0079] Preferably, the intervention strategy generation module is used to preset the emotion abnormality threshold Y and calculate the emotion abnormality score index of the i-th student in the T learning cycle. Compare with the emotion abnormality threshold Y to obtain a second assessment result, including:
[0080] when If the student is in a "continuous frustration pattern," the student is deemed to be in a "first emotional risk state," triggering the second warning instruction and initiating a key intervention strategy, including: immediately generating and pushing personalized guidance information, with a push frequency of once per interaction round;
[0081] when If the student exhibits a "continuous frustration pattern," the student is deemed to be in a "second emotional risk state," which is lower risk than the "first emotional risk state." A moderate intervention strategy is initiated, including: immediately generating and pushing personalized guidance information at a frequency of once every 3-5 interaction rounds.
[0082] when And when there is no "continuous frustration mode", the student is determined to be in the "third emotional risk state". The "third emotional risk state" has a lower risk level than the "second emotional risk state", and a low-level intervention strategy is initiated, including: immediately generating and pushing personalized guidance information, with a push frequency of once every 7-9 interaction rounds.
[0083] The present invention provides a student learning emotion recognition system based on natural language processing. It has the following beneficial effects:
[0084] This paper addresses the technical challenges of identifying students' emotions, expressing them implicitly, and experiencing complex dynamic changes in their emotional states during language learning. By providing a student learning emotion recognition system based on natural language processing, this system leverages multiple modules to collaborate, from collecting data from multiple rounds of natural language interactions, to identifying contextual emotional expressions, constructing emotional transition paths, and integrating multi-dimensional emotional features, comprehensively capturing students' complex emotional dynamics.
[0085] First, the system can deeply analyze the subtle emotional fluctuations in students' language expressions, especially the emotional changes at key learning nodes such as "understanding confusion," "expression barriers," and "active attempts," thereby improving the granularity and accuracy of emotion recognition. Secondly, by constructing an emotional fluctuation index to quantify the stability and switching frequency of students' emotions, it achieves early warning of cognitive fluctuations and psychological conflict states, effectively making up for the shortcomings of traditional monitoring that relies solely on behavior or grades. Thirdly, based on multi-factor emotional anomaly scoring, the system accurately identifies emotional risk states such as "continuous frustration mode," providing a scientific quantitative basis for teaching management and psychological intervention.
[0086] Overall, the present invention significantly improves the dynamic tracking ability and risk warning level of students' emotion recognition in language subject learning, helps to achieve targeted personalized emotion intervention and intelligent teaching support, and enhances students' learning experience and mental health protection. BRIEF DESCRIPTION OF THE DRAWINGS
[0087] Figure 1 This is a flow chart of a student learning emotion recognition system based on natural language processing according to the present invention. DETAILED DESCRIPTION
[0088] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0089] Example 1
[0090] See also Figure 1 The present invention provides a student learning emotion recognition system based on natural language processing, comprising:
[0091] The first dialogue collection module is used to collect multiple rounds of natural language interaction sentences between students and the learning system or teachers during the language learning process to obtain a first language sequence set;
[0092] a first emotion recognition module, configured to perform context modeling analysis on the first language sequence set, identify emotion expression segments in language sentences, construct a student emotion fluctuation trajectory based on temporal features, and extract a first emotion transfer path;
[0093] The second emotion recognition module is used to extract the understanding confusion feature, language expression discomfort feature, active attempt feature and positive cognitive feedback feature based on the first emotion transfer path, and integrate them into the first emotion transfer path to generate the second emotion transfer path;
[0094] The emotion fluctuation index calculation module is used to assign positive, neutral and negative polarity labels to each fusion label based on the second emotion transfer path, obtain the third polarity sequence, and construct the emotion fluctuation index of the i-th student in the T learning cycle. It is used to measure the frequency of positive and negative emotion switching per unit time. If a student falls into a state of cognitive fluctuation or psychological conflict, the first warning instruction will be triggered;
[0095] The emotional consistency judgment module is used to determine whether there is a "continuous frustration mode" based on the third polarity sequence after receiving the first warning instruction, and to construct the emotional abnormality score index of the i-th student in the T learning cycle. The scoring factors include: negative sentence ratio coefficient α, cognitive contradiction density coefficient β and vocabulary ambiguity rate γ;
[0096] Intervention strategy generation module, based on the abnormal emotion score index of the i-th student in the T learning cycle Conduct assessments to obtain corresponding emotional risk states and generate corresponding intervention strategies.
[0097] This embodiment, based on natural language processing technology, can accurately and in real time capture the emotions expressed by students through natural language communication during language learning. This overcomes the bottleneck of traditional emotion recognition technologies, which struggle to effectively handle continuous conversational contexts and implicit emotions. By constructing student emotion fluctuation trajectories and emotional transition pathways, the system dynamically monitors emotional trends during the learning process. This system can effectively identify potential emotional risk states such as cognitive fluctuations and repeated frustration, issuing timely warnings and helping teachers and learning platforms more accurately grasp students' psychological conditions. This system integrates multidimensional emotional features, including confusion in understanding, language expression discomfort, proactive attempts, and positive cognitive feedback, making emotion recognition more detailed and comprehensive, and improving the accuracy and practicality of emotion discrimination. By calculating an emotion fluctuation index and anomaly scores, the system can quantify the stability and degree of anomalies of students' emotions, providing a scientific basis for subsequent interventions, facilitating the implementation of personalized and tiered teaching intervention strategies, and improving the relevance and effectiveness of teaching. This system can also help promptly identify students' psychological stress and emotional distress during language learning, promote emotional regulation and mental health, enhance learning motivation and self-efficacy, and effectively promote language learning outcomes.
[0098] Example 2
[0099] This example is explained in Example 1. Figure 1 ,Specifically, the first dialogue acquisition module includes an interactive access unit and a language sequence ,construction unit;
[0100] The interactive access unit is used to establish a communication connection with the learning platform used by students or the teacher's terminal, access multiple rounds of natural language interactive input channels, collect and obtain natural language input data between the i-th student and the learning system or language subject teacher, and obtain the first language data;
[0101] The language sequence construction unit is used to convert the text input and voice data in the first language data into text form, set a session identifier, perform session recognition on each student's consecutive input sentences, and determine the session boundaries according to the set time interval or chapter switching rules to form a multi-round sentence sequence; pre-process the collected natural language sentences, remove redundant modal particles and unstructured tag information, and screen out sentence content with language subject relevance;
[0102] Removal of redundant modal particles and unstructured tag information includes:
[0103] Remove extra spaces, HTML tags and special symbols;
[0104] Special symbols include: "#", "@" and "$";
[0105] Convert full-width characters to half-width characters, unify Chinese and English punctuation forms, and then unify the encoding format. Add timestamps, student identity information, course and chapter number metadata to each sentence, and generate the first language sequence set of the i-th student containing timestamps and learning metadata.
[0106] In this embodiment, the first dialogue collection module can fully and standardizedly obtain the multi-dimensional corpus data of students in the process of language subject learning through efficient access and precise construction of multiple rounds of natural language interaction, ensuring the authenticity and reliability of the data basis for subsequent emotion recognition analysis; at the same time, by removing redundant modal particles, unstructured tags and unified coding formats, the quality and consistency of language data are significantly improved, noise interference is effectively reduced, and the system's recognition accuracy and stability of the emotional characteristics of students' language expression are enhanced, thereby providing a solid guarantee for the in-depth analysis and emotional state monitoring of the emotion recognition module, and promoting intelligent management of the learning process and personalized emotional intervention.
[0107] Example 3
[0108] This example is explained in Example 2. Figure 1 ,Specifically, the first emotion recognition module includes an ,interjection feature extraction unit and an emotion expression ,segment recognition unit;
[0109] The modal particle feature extraction unit is used to scan the modal particles in the first language sequence set and use the modal particles as the first emotion amplification feature marker; the modal particle list includes: "ah", "er", "ma", "ya", "ge", "ouch", "wow" and "my god";
[0110] The emotion expression segment recognition unit is used to perform semantic understanding and emotion classification on the sentences in the first language sequence set based on the context information and the Transformer emotion classification model; automatically divide the language segments containing obvious emotion expressions and mark the corresponding emotion categories;
[0111] The steps for the Transformer sentiment classification model to perform semantic understanding and sentiment classification are:
[0112] A context window is formed to extract a single sentence or multiple consecutive sentences from the first language sequence set. The text is segmented and encoded into a token sequence and converted into a token ID.
[0113] The pre-trained Transformer sentiment classification model is used to encode the input token sequence to obtain the contextual semantics of each token. After the Transformer encoding and fully connected layer of the Transformer sentiment classification model, a probability vector P is obtained through softmax.
[0114] P={p1, p2, ..., p n};
[0115] Where n is the number of predefined emotion categories, p j is the predicted probability of the jth category, satisfying:
[0116]
[0117] Find the emotion category j corresponding to the maximum probability * and the maximum probability value p max :
[0118]
[0119] p max =p j*
[0120] Set the confidence threshold R = 0.6, if p max ≥R, then class j * as the final emotion label for the statement;
[0121] If p max <R, it means that the emotion of the sentence is unclear and is labeled as “no obvious emotion”.
[0122] During the model training and validation phases, prediction tests were conducted on a large amount of labeled data, observing the changes in the model's accuracy, recall, and F1 score at different thresholds. A comprehensive balance was then established between model precision and coverage, and a threshold was selected that ensured high accuracy without excessively discarding valid emotion judgments. A threshold of 0.6 was a commonly used and effective cutoff point in experience. Setting a threshold of 0.6 effectively avoids incorrect emotion classifications (false positives) due to low confidence levels, while also avoiding excessively strict judgments of "no obvious emotion" (missed positives), ensuring the system's sensitivity and robustness in emotion recognition.
[0123] The first emotion recognition module further includes an emotion path generation unit;
[0124] The emotion path generating unit is used to arrange each of the first language sequence combinations in the order of timestamps (or dialogue turns) to form a time-ordered emotion tag sequence S:
[0125] S={(t1, e1), (t2, e2),..., (t m , e m )};
[0126] Among them, t m represents the timestamp or interaction sequence corresponding to the mth statement, e m The final emotion label is obtained by the emotion expression segment recognition unit;
[0127] Assign standardized codes to each emotion category; including:
[0128] Confused C1, anxious C2, happy C3, angry C4, stable C0, and no obvious emotion C5;
[0129] Based on the emotional state group of the sentences at time t and time t+1 in the time series, the emotional state transition frequency matrix is constructed:
[0130]
[0131] in, represents the frequency of emotional state transition between the i-th student at time t and the adjacent time t+1;
[0132] e t+1 Indicates the Cnth emotional state e Cn To the Cbth emotional state e Cb The probability of; P(·) represents probability; the inside = represents logical judgment, which means that at time A at moment t+1, the student’s emotion is e Cn , Condition B At time t, the student's emotion is e Cb ; Cnth emotional state e Cn To the Cbth emotional state e Cb Corresponding to e m It is the final emotion label obtained by the emotion expression fragment recognition unit.
[0133] Examples of emotional paths in some student conversations are shown in Table 1:
[0134] time Student statements Emotional labels t1 This grammar is too difficult...ah... Puzzled t2 Did I get it wrong somewhere? anxiety t3 Oh! So that's how it is! happy t4 I get it! That’s awesome! happy
[0135] The first emotional transfer path: confusion → anxiety → happiness → happiness
[0136] In this embodiment, the first emotion recognition module achieves in-depth capture and precise classification of emotional information in students' language expressions by combining modal particle feature extraction with Transformer-based emotion expression segment recognition. The modal particle feature extraction unit effectively utilizes modal particles with strong emotional connotations as emotion amplification markers, enhancing the sensitivity of emotion recognition, especially with a significant advantage in identifying subtle emotional changes.
[0137] The context-window-based Transformer sentiment classification model, combined with pre-trained deep semantic coding technology, can understand the semantic associations and emotional expressions in the context of multiple rounds of dialogue, automatically divide emotional segments and perform high-confidence emotion classification, greatly improving the accuracy and robustness of recognition, and avoiding the misjudgment and missed judgment problems that are prone to occur in traditional single-sentence analysis.
[0138] The confidence threshold screening mechanism effectively ensures the reliability of the final emotion labels and reduces the interference of noise and invalid labels on subsequent analysis. Furthermore, the emotion path generation unit organizes the emotion label sequence in chronological order, constructing the temporal evolution trajectory of students' emotional states and revealing the transition patterns between emotional states.
[0139] The construction of an emotional state transition frequency matrix not only quantifies the dynamic changes in students' emotions over the learning process but also provides a scientific basis for identifying abnormal emotional fluctuations and predicting emotional risks. This feature enables the learning platform to monitor students' emotional states in real time, assisting teachers in providing precise intervention and support tailored to students' individual emotional changes.
[0140] Example 4
[0141] This example is explained in Example 1. Figure 1 ,Specifically, the second emotion recognition module includes a learning ,emotion feature dictionary construction unit, a language segment screening unit, and a ,second label classification unit;
[0142] The learning emotion feature dictionary construction unit is used to establish a feature dictionary of common emotion representations in language learning scenarios based on the first emotion transfer path, and the dictionary includes the following four categories of emotion feature keyword sets:
[0143] Comprehension confusion characteristics include: “I don’t know,” “What’s going on,” “I don’t understand,” and “I’m confused”;
[0144] Characteristics of language expression discomfort include: "can't say it", "can't pronounce it correctly", "difficult to pronounce" and "this is too complicated";
[0145] Active attempt characteristics include: "I'll try", "I guess so", and "Let me tell you about it";
[0146] Positive cognitive feedback characteristics include: "I see," "I understand," and "It's very simple."
[0147] The language segment screening unit is used to extract the comprehension confusion feature, language expression discomfort feature, active attempt feature, and positive cognitive feedback feature from the sentences in the first language sequence set, and perform a second emotion label discrimination through the second label classification unit in combination with the context information to obtain the second emotion label;
[0148] The second emotional label includes: understanding confusion feature F1, language expression discomfort feature F2, active attempt feature F3 and positive cognitive feedback feature F4.
[0149] The second emotion recognition module further includes a fusion unit;
[0150] The fusion unit is used to combine the emotional state corresponding to each sentence in the first emotional transfer path with the second emotional label identified by the second label classification unit to form a second emotional transfer path with attributes; in the second emotional transfer path, each time point has not only an emotional state label, but also behavioral cognitive characteristics as an additional dimension.
[0151] In this embodiment, the second emotion recognition module, by constructing an exclusive emotion feature dictionary for language learning scenarios, accurately captures the multi-dimensional emotional expressions of students in the learning process, especially key emotional features such as understanding confusion, expression barriers, active attempts and positive feedback, thereby achieving deep semantic analysis of students' emotional states and comprehensive recognition of behavioral cognitive features. Through multi-label classification and discrimination combining language fragment screening with context, it can accurately distinguish and label students' different emotional features, effectively improving the granularity and diversity of emotion recognition. The fusion unit further synchronizes the emotional state with the behavioral cognitive features to construct a multi-dimensional second emotion transfer path, enriching the descriptive dimension of the emotional state, and providing more comprehensive and accurate data support for subsequent emotional fluctuation analysis and personalized intervention strategies. It significantly enhances the system's ability to capture the complexity of students' learning emotions and the monitoring effect of the dynamic evolution of emotions, and promotes the implementation of targeted teaching support and psychological counseling.
[0152] Example 5
[0153] This example is explained in Example 4. Figure 1 ,Specifically, the mood fluctuation index calculation module includes a mood polarity determination unit and a mood switching monitoring unit;
[0154] The emotion polarity determination unit is configured to scan the second emotion transfer path, assign positive, neutral, and negative polarity labels to each fusion label, and obtain a third polarity sequence;
[0155] Positive, neutral, and negative polarity markers include:
[0156] Positive polarity markers include: happiness C3 and positive cognitive feedback feature F4;
[0157] Neutral polarity markers include: stable C0 and active attempt feature F3;
[0158] Negative polarity markers include: confusion C1, anxiety C2, understanding confusion feature F1, and language expression discomfort feature F2;
[0159] The emotion switching monitoring unit is used to scan the third polarity sequence and count the number of The number of switching times W, ignoring the switching between neutrals, includes: "negative → neutral → positive" is counted as one switching, that is, W = 1. When "negative → negative" or "positive → neutral → positive" occurs, the number of switching times is not counted, that is, W = 0.
[0160] The mood fluctuation index calculation module also includes the mood fluctuation index generation unit and the evaluation unit.
[0161] The mood fluctuation index generating unit is used to extract the mood fluctuation index of the i-th student within a unit time. The number of switching times W is calculated by the following formula to obtain the emotional fluctuation index of the i-th student in the T learning cycle
[0162]
[0163] Among them, W i Indicates the number of students i in the learning cycle The total number of switches between emotional polarities; T g represents the total number of students' words and sentences in T learning cycle; A larger value indicates a more unstable mood;
[0164] The evaluation unit presets the emotion fluctuation threshold B and calculates the emotion fluctuation index of the i-th student in the T learning cycle Comparing with the emotion fluctuation threshold B, the first emotion assessment result is obtained, including:
[0165] when Determining that the student is in a state of cognitive fluctuation or psychological conflict triggers the first warning instruction;
[0166] when It is judged that the student's emotions are relatively stable and the learning state is normal, and the first qualified label is generated.
[0167] In this embodiment, the mood fluctuation index calculation module accurately quantifies the frequency and intensity of student emotional fluctuations through a detailed classification of emotional polarity and dynamic monitoring of switching behaviors during learning. The mood polarity determination unit combines specific emotion categories with learning behavior characteristics to scientifically classify positive, neutral, and negative polarities, achieving precise assignment of multi-dimensional emotion polarity labels. The mood switching monitoring unit, based on polarity sequences, rationally distinguishes effective emotion switching paths, eliminates interference from neutral fluctuations, and ensures the objectivity and practicality of mood fluctuation statistics.
[0168] The Mood Fluctuation Index Generation Unit uses the ratio of the number of switching times to the total number of sentences learned to construct a standardized mood fluctuation index, which quantitatively reflects students' emotional stability. A higher value indicates greater emotional instability. The Assessment Unit uses a preset threshold to determine this index, enabling timely identification of students experiencing cognitive fluctuations or psychological conflicts, triggering appropriate warning mechanisms and enabling early intervention.
[0169] Example 6
[0170] This example is explained in Example 5. Figure 1 ,Specifically, the emotion coherence judgment module includes a multi-factor scoring ,construction unit and an emotion abnormality scoring generation unit;
[0171] The multi-factor scoring construction unit is used to combine the emotion fluctuation index of the i-th student in the T learning cycle after receiving the first warning instruction. The corresponding students' first language data are extracted to determine whether there are multiple consecutive negative emotion labels. If there are three or more consecutive negative emotion labels, it is determined to be a "continuous frustration mode";
[0172] And calculate the negative sentence ratio coefficient α:
[0173]
[0174] Among them, C neg is the total number of negative emotion sentences in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle;
[0175] To measure whether there are "inconsistent" cognitive expressions, including contradictory sentences such as "I understand" followed by "I still don't understand", or "self-negating" language fragments, the cognitive contradiction density coefficient β is calculated:
[0176]
[0177] Among them, C confloct is the total number of words in cognitively contradictory language segments in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle;
[0178] The proportion of cognitive expressions containing “fuzzy words” is measured. “Fuzzy words” include “almost”, “maybe”, “okay” and “not very clear”. The lexical fuzziness rate γ is calculated as follows:
[0179]
[0180] Among them, C vague is the number of “fuzzy expression words” used by students in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle;
[0181] The emotion abnormality score generation unit is used to extract the negative sentence ratio coefficient α, cognitive contradiction density coefficient β, and vocabulary fuzziness rate γ. After dimensionless processing, the emotion abnormality score index of the i-th student in the T learning cycle is calculated using the following formula:
[0182]
[0183] Where w1, w2, and w3 represent the weights of the negative sentence ratio coefficient α, the cognitive contradiction density coefficient β, and the lexical ambiguity rate γ, respectively. 0<w1<1, 0<w2<1, 0<w3<1, and w1+w2+w3=1.
[0184] In this embodiment, the emotional coherence assessment module not only uses the emotional fluctuation index but also incorporates a "continuous frustration pattern" recognition mechanism. This module comprehensively analyzes students' continuous negative emotional states through a multi-factor scoring system (including the proportion of negative sentences, the density of cognitive contradictions, and the rate of lexical ambiguity), enabling accurate identification of emotional anomalies. This module dynamically captures the continuous accumulation of negative emotions during language learning, promptly identifying recurring cognitive frustration and psychological stress, and avoiding misjudgments caused by single emotional fluctuations.
[0185] By effectively identifying "continuous frustration patterns," the system can provide early warnings of potential learning bottlenecks and emotional risks facing students, facilitating the timely implementation of targeted intervention strategies and significantly improving the scientific and effective management of emotional risks. Furthermore, the emotional abnormality scoring index provides a quantitative basis for teaching management and psychological counseling, helping to provide personalized emotional care and teaching support, and promoting the restoration of students' positive emotions and increased learning motivation.
[0186] In summary, this module enhances the system's ability to understand the continuity and coherence of students' emotional states, improves the warning accuracy and response speed of the overall learning emotion recognition system, and greatly assists in emotional health management and optimization of learning effects in language subject learning environments.
[0187] Example 7
[0188] This example is explained in Example 6. Figure 1 Specifically, the intervention strategy generation module is used to preset the emotion abnormality threshold Y and calculate the emotion abnormality score index of the i-th student in the T learning cycle. Compare with the emotion abnormality threshold Y to obtain a second assessment result, including:
[0189] when If the student is in a "continuous frustration pattern," the student is deemed to be in a "first emotional risk state," triggering the second warning instruction and initiating a key intervention strategy, including: immediately generating and pushing personalized guidance information, with a push frequency of once per interaction round;
[0190] when If the student exhibits a "continuous frustration pattern," the student is deemed to be in a "second emotional risk state," which is lower risk than the "first emotional risk state." A moderate intervention strategy is initiated, including: immediately generating and pushing personalized guidance information at a frequency of once every 3-5 interaction rounds.
[0191] when And when there is no "continuous frustration mode", the student is determined to be in the "third emotional risk state". The "third emotional risk state" has a lower risk level than the "second emotional risk state", and a low-level intervention strategy is initiated, including: immediately generating and pushing personalized guidance information, with a push frequency of once every 7-9 interaction rounds.
[0192] In this embodiment, the intervention strategy generation module scientifically divides students' emotional risk levels based on the emotional abnormality score index and the "continuous frustration mode" identification results, and realizes hierarchical and personalized emotional intervention. By setting differentiated push frequencies and key intervention measures, it is possible to flexibly adjust the guidance intensity and response speed for students in different risk states, avoiding student disgust caused by excessive intervention and ensuring that students with high emotional risks can receive timely and effective help. This module effectively improves the pertinence and practicality of emotional intervention, promotes students' emotional recovery and the establishment of a positive learning attitude, and enhances the intelligent management capabilities and teaching support effects of the overall learning emotion recognition system.
[0193] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by technicians in this field for each set of sample data; as long as it does not affect the proportional relationship between the parameter and the quantized value.
[0194] The above formulas are obtained by collecting a large amount of data and performing software simulation, and a formula close to the actual value is selected. The coefficients in the formula are set by those skilled in the art according to actual conditions. The above is only a preferred specific implementation method of the present invention, but the protection scope of the present invention is not limited to this. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, can make equivalent replacements or changes based on the technical solution and inventive concept of the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A student learning emotion recognition system based on natural language processing, characterized by: include: The first dialogue collection module is used to collect multiple rounds of natural language interaction sentences between students and the learning system or teachers during the language learning process to obtain a first language sequence set; a first emotion recognition module, configured to perform context modeling analysis on the first language sequence set, identify emotion expression segments in language sentences, construct a student emotion fluctuation trajectory based on temporal features, and extract a first emotion transfer path; The second emotion recognition module is used to extract the understanding confusion feature, language expression discomfort feature, active attempt feature and positive cognitive feedback feature based on the first emotion transfer path, and integrate them into the first emotion transfer path to generate the second emotion transfer path; The emotion fluctuation index calculation module is used to assign positive, neutral and negative polarity labels to each fusion label based on the second emotion transfer path, obtain the third polarity sequence, and construct the emotion fluctuation index of the i-th student in the T learning cycle. It is used to measure the frequency of positive and negative emotion switching per unit time. If a student falls into a state of cognitive fluctuation or psychological conflict, the first warning instruction will be triggered; The emotional consistency judgment module is used to determine whether there is a "continuous frustration mode" based on the third polarity sequence after receiving the first warning instruction, and to construct the emotional abnormality score index of the i-th student in the T learning cycle. The scoring factors include: negative sentence ratio coefficient α, cognitive contradiction density coefficient β and vocabulary ambiguity rate γ; Intervention strategy generation module, based on the abnormal emotion score index of the i-th student in the T learning cycle Conduct assessments to obtain corresponding emotional risk states and generate corresponding intervention strategies.
2. A student learning emotion recognition system based on natural language processing according to claim 1, characterized in that: The first dialogue collection module includes an interaction access unit and a language sequence construction unit; The interactive access unit is used to establish a communication connection with the learning platform used by students or the teacher's terminal, access multiple rounds of natural language interactive input channels, collect and obtain natural language input data between the i-th student and the learning system or language subject teacher, and obtain the first language data; The language sequence construction unit is used to convert the text input and voice data in the first language data into text form, set a session identifier, perform session recognition on each student's continuous input sentences, and determine the session boundary according to the set time interval or chapter switching rules to form a multi-round sentence sequence; Preprocess the collected natural language sentences, remove redundant modal particles and unstructured tag information, and filter out sentences with language-related content; Removal of redundant modal particles and unstructured tag information includes: Remove extra spaces, HTML tags and special symbols; Special symbols include: "#", "@" and "$"; Convert full-width characters to half-width characters, unify Chinese and English punctuation forms, and then unify the encoding format. Add timestamps, student identity information, course and chapter number metadata to each sentence, and generate the first language sequence set of the i-th student containing timestamps and learning metadata.
3. The student learning emotion recognition system based on natural language processing according to claim 2 is characterized in that: The first emotion recognition module includes an interjection feature extraction unit and an emotion expression segment recognition unit; The modal particle feature extraction unit is used to scan the modal particles in the first language sequence set and use the modal particles as the first emotion amplification feature marker; the modal particle list includes: "ah", "er", "ma", "ya", "ge", "ao ya", "wow" and "my god"; The emotion expression segment recognition unit is used to perform semantic understanding and emotion classification on the sentences in the first language sequence set based on the context information and the Transformer emotion classification model; automatically divide the language segments containing obvious emotion expressions and mark the corresponding emotion categories; The steps for the Transformer sentiment classification model to perform semantic understanding and sentiment classification are: A context window is formed to extract a single sentence or multiple consecutive sentences from the first language sequence set. The text is segmented and encoded into a token sequence and converted into a token ID. The pre-trained Transformer sentiment classification model is used to encode the input token sequence to obtain the contextual semantics of each token. After the Transformer encoding and fully connected layer of the Transformer sentiment classification model, a probability vector P is obtained through softmax. P={p1,p2,...,p n }; Where n is the number of predefined emotion categories, p j is the predicted probability of the jth category, satisfying: Find the emotion category j corresponding to the maximum probability * and the maximum probability value p max : Set the confidence threshold R = 0.6, if p max ≥R, then class j * as the final emotion label for the statement; If p max <R, it means that the emotion of the sentence is unclear and is labeled as "no obvious emotion".
4. The student learning emotion recognition system based on natural language processing according to claim 3 is characterized in that: The first emotion recognition module further includes an emotion path generation unit; The emotion path generating unit is used to arrange each of the first language sequence combinations in the order of timestamps to form a time-ordered emotion tag sequence S: S={(t1,e1),(t2,e2),...,(t m ,e m )}; Among them, t m represents the timestamp or interaction sequence corresponding to the mth statement, e m The final emotion label is obtained by the emotion expression segment recognition unit; Assign standardized codes to each emotion category; including: Confused C1, anxious C2, happy C3, angry C4, stable C0, and no obvious emotion C5; Based on the emotional state group of the sentences at time t and time t+1 in the time series, the emotional state transition frequency matrix is constructed: in, represents the frequency of emotional state transition between the i-th student at time t and the adjacent time t+1; e t+1 Indicates the Cnth emotional state e Cn To the Cbth emotional state e Cb The probability of; P(·) represents probability; the inside = represents logical judgment, which means that at time A at moment t+1, the student’s emotion is e Cn , Condition B At time t, the student's emotion is e Cb ; Cnth emotional state e Cn To the Cbth emotional state e Cb Corresponding to e m It is the final emotion label obtained by the emotion expression fragment recognition unit.
5. The student learning emotion recognition system based on natural language processing according to claim 4 is characterized in that: The second emotion recognition module includes a learning emotion feature dictionary construction unit, a language segment screening unit, and a second label classification unit; The learning emotion feature dictionary construction unit is used to establish a feature dictionary of common emotion representations in language learning scenarios based on the first emotion transfer path, and the dictionary includes the following four categories of emotion feature keyword sets: Comprehension confusion characteristics include: "I don't know", "What's going on", "I don't understand" and "I'm confused"; Characteristics of language discomfort include: "can't say it", "can't pronounce it correctly", "difficult to pronounce", and "this is too complicated"; Active attempt features include: "I'll try", "I guess so", and "Let me tell you about it"; Positive cognitive feedback characteristics include: "I see," "I understand," and "It's very simple." The language segment screening unit is used to extract the comprehension confusion feature, language expression discomfort feature, active attempt feature, and positive cognitive feedback feature from the sentences in the first language sequence set, and perform a second emotion label discrimination through the second label classification unit in combination with the context information to obtain the second emotion label; The second emotional label includes: understanding confusion feature F1, language expression discomfort feature F2, active attempt feature F3 and positive cognitive feedback feature F4.
6. The student learning emotion recognition system based on natural language processing according to claim 5 is characterized in that: The second emotion recognition module further includes a fusion unit; The fusion unit is used to combine the emotional state corresponding to each sentence in the first emotional transfer path with the second emotional label identified by the second label classification unit to form a second emotional transfer path with attributes; in the second emotional transfer path, each time point has not only an emotional state label, but also behavioral cognitive characteristics as an additional dimension.
7. The student learning emotion recognition system based on natural language processing according to claim 6 is characterized in that: The mood fluctuation index calculation module includes a mood polarity determination unit and a mood switching monitoring unit; The emotion polarity determination unit is configured to scan the second emotion transfer path, assign positive, neutral, and negative polarity labels to each fusion label, and obtain a third polarity sequence; Positive, neutral, and negative polarity markers include: Positive polarity markers include: happiness C3 and positive cognitive feedback feature F4; Neutral polarity markers include: stable C0 and active attempt feature F3; Negative polarity markers include: confusion C1, anxiety C2, understanding confusion feature F1, and language expression discomfort feature F2; The emotion switching monitoring unit is used to scan the third polarity sequence and count the number of "positive The number of "negative" switching is W, ignoring the switching between neutrals, including: "negative → neutral → positive" is recorded as one switching, that is, W = 1. When "negative → negative" or "positive → neutral → positive" occurs, the number of switching is not counted, that is, W = 0.
8. The student learning emotion recognition system based on natural language processing according to claim 7 is characterized in that: The mood fluctuation index calculation module also includes a mood fluctuation index generation unit and an evaluation unit; The mood fluctuation index generating unit is used to extract the "positive" The number of negative switching times W is calculated by the following formula to obtain the emotional fluctuation index of the i-th student in the T learning cycle Among them, W i Indicates that the i-th student is "positive" in the learning cycle The total number of switches between negative and positive emotion polarities; T g represents the total number of students' words and sentences in T learning cycle; A larger value indicates a more unstable mood; The evaluation unit presets the emotion fluctuation threshold B and calculates the emotion fluctuation index of the i-th student in the T learning cycle Comparing with the emotion fluctuation threshold B, the first emotion assessment result is obtained, including: when Determining that the student is in a state of cognitive fluctuation or psychological conflict triggers the first warning instruction; when It is judged that the student's emotions are relatively stable and the learning state is normal, and the first qualified label is generated.
9. The student learning emotion recognition system based on natural language processing according to claim 8 is characterized in that: The emotion coherence judgment module includes a multi-factor scoring construction unit and an emotion abnormality score generation unit; The multi-factor scoring construction unit is used to combine the emotion fluctuation index of the i-th student in the T learning cycle after receiving the first warning instruction. The corresponding student's first language data is extracted to determine whether there are multiple consecutive negative emotion labels. If there are three or more consecutive negative emotion labels, it is determined to be a "continuous frustration mode"; And calculate the negative sentence ratio coefficient α: Among them, C neg is the total number of negative emotion sentences in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle; To measure whether there are "inconsistent" cognitive expressions, including contradictory sentences such as "I understand" followed by "I still don't understand", or "self-negating" language fragments, the cognitive contradiction density coefficient β is calculated: Among them, C confloct is the total number of words in cognitively contradictory language segments in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle; The proportion of cognitive expressions that contain "fuzzy words" is measured. "Fuzzy words" include vague sentences such as "almost", "maybe", "okay", and "not very clear". The lexical fuzziness rate γ is calculated as: Among them, C vague is the number of "fuzzy expression words" used by students in a learning cycle; T cs It indicates the total number of words spoken by students in this cycle; The emotion abnormality score generation unit is used to extract the negative sentence ratio coefficient α, cognitive contradiction density coefficient β, and vocabulary fuzziness rate γ. After dimensionless processing, the emotion abnormality score index of the i-th student in the T learning cycle is obtained through the weighted formula.
10. The student learning emotion recognition system based on natural language processing according to claim 9 is characterized in that: The intervention strategy generation module is used to preset the abnormal emotion threshold Y and calculate the abnormal emotion score index of the i-th student in the T learning cycle. Compare with the emotion abnormality threshold Y to obtain a second assessment result, including: when If the student exhibits a "continuous frustration pattern," the student is deemed to be in a "first emotional risk state," triggering a second warning instruction and initiating a key intervention strategy, including immediately generating and pushing personalized guidance information once per interaction round. when If the student exhibits a "continuous frustration pattern," the student is deemed to be in a "secondary emotional risk state," which is lower risk than the "first emotional risk state." A moderate intervention strategy is initiated, including: immediately generating and delivering personalized guidance information, with a frequency of once every 3-5 interaction rounds. when If there is no "continuous frustration pattern", the student is determined to be in the "third emotional risk state", which is lower than the "second emotional risk state" and a low-level intervention strategy is initiated, including: Personalized guidance information is generated and pushed immediately, with a push frequency of once every 7-9 interaction rounds.
