A personalized metacognitive enhancement heuristic emotional support dialogue method

By combining user personality and emotional state with a deep learning model to generate heuristic questions, this approach addresses the problem that existing emotional support methods do not consider metacognitive abilities, thereby enhancing users' emotional control and experience.

CN116702794BActive Publication Date: 2026-11-17EAST CHINA NORMAL UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310704220.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-06-14
Publication Date
2026-11-17
Estimated Expiration
2043-06-14

AI Technical Summary

Technical Problem

Existing emotional support methods fail to effectively consider users' individual metacognitive abilities, resulting in inflexible responses and poor user experience.

Method used

By using a deep learning model that combines user personality and current emotional state, heuristic questions are generated, providing personalized emotional support by enhancing users' ability to manage their emotions.

Benefits of technology

By combining users' metacognitive abilities, questions tailored to their characteristics can be generated, thereby enhancing users' emotional metacognitive abilities, effectively reducing negative emotions, and improving user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116702794B_ABST
    Figure CN116702794B_ABST
Patent Text Reader

Abstract

This invention discloses a personalized metacognitive enhancement-based heuristic emotional support method. Its key features include employing a personality-driven metacognitive ability extraction method, utilizing a language model to generate a set of heuristic questions, and using an evaluation algorithm based on user personality and current emotional state to select the most appropriate question as a response. This method enhances emotional support by improving the user's emotional control. Specifically, the method includes: acquiring the user's current emotional state, current metacognitive ability characteristics, and current state representation; selecting the most appropriate question to ask; and feeding the user's feedback and their metacognitive characteristics back into the language model for the next iteration until the user's mood improves. Compared to existing technologies, this invention addresses emotional issues from a personalized perspective, leverages the powerful generative capabilities of the language model, designs language generation constraints for metacognitive enhancement, and obtains responses that better meet the user's emotional needs and are more diverse and vivid.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of emotion support task technology in dialogue systems, specifically a heuristic emotion support dialogue method based on personalized metacognitive enhancement. Background Technology

[0002] Emotional support is an important subtask in the field of dialogue systems. Its purpose is to provide comfort, understanding, encouragement, and attention to people through machine-to-human dialogue, thereby alleviating negative emotions. This task focuses on how to identify, understand, simulate, respond to, and process human emotions and feelings through question-and-answer systems, and to ask appropriate questions at the right time to provide effective emotional support. Early emotional support methods mostly identified emotional information in text based on predefined rules, patterns, or features, and then selected corresponding responses from multiple pre-designed templates based on the identified emotional polarity. However, limited response templates lack flexibility, and templated answers may not meet the complex emotional needs of users.

[0003] With the continuous maturation of various generative language modeling technologies, the encoder-decoder paradigm has gradually become the mainstream of emotion support methods. Most existing emotion support methods first use an encoder to encode features such as the user's emotional state, behavioral intentions, and historical dialogue information, then select an appropriate response strategy, and finally generate the final response through a decoder. However, existing research often only judges the user's current state and provides simple comfort based on the text content input during historical interactions or combined with common sense information, failing to consider the user's individual characteristics, such as metacognitive abilities. Metacognition is an individual's awareness and monitoring of their own thinking and learning activities. From an emotional perspective, a user's metacognitive ability is reflected in their ability to recognize and control their own emotions. Research shows that guiding learners to reflect on and introspect on their negative emotions through heuristic questions can effectively improve users' metacognitive abilities in terms of emotions, that is, enhance their ability to control their own emotions. Therefore, for emotion support methods, if users' metacognitive abilities can be incorporated as an important personalized feature into the model, and further improved through heuristic questioning, negative emotions can be effectively mitigated, and their negative impact avoided.

[0004] Current emotional support technologies rely solely on the text content entered during historical interactions or common-sense information to determine the user's current state and offer simple comfort. They fail to consider the user's individual characteristics, cannot effectively resolve emotional issues, and result in a poor user experience. Summary of the Invention

[0005] The purpose of this invention is to provide a heuristic emotional support dialogue method based on personalized metacognitive enhancement, addressing the shortcomings of existing technologies. It employs a deep learning model to extract personality-driven metacognitive abilities, using extracted metacognitive abilities as indirect information. This is combined with user emotions extracted from text, and a large-scale generative language model is used to generate a set of heuristic questions. Finally, an evaluation algorithm based on user personality and current emotional state selects the most appropriate question as the response. By enhancing the user's emotional control, this method provides emotional support, effectively solving the problems of traditional emotional support methods lacking metacognitive guidance in responses and failing to consider the user's personalized metacognitive abilities.

[0006] The specific technical solution to achieve the purpose of this invention is: a heuristic emotional support method based on personalized metacognitive enhancement. Its characteristics include using user personality to extract their metacognitive abilities as indirect information, combining this with user emotions extracted from text, generating a set of heuristic questions using a large-scale generative language model, and finally selecting the most appropriate question as a response based on an evaluation algorithm of user personality and current emotional state. This method helps provide emotional support by enhancing the user's ability to manage their emotions. The method includes the following specific steps:

[0007] Step 1: Obtain the user's current emotional state based on the current dialogue information.

[0008] Assuming the current round of historical dialogue information t represents the current round of the dialogue, and s represents the user's statements in each round. First, all historical dialogue information is concatenated to form a long sequence of text. Here, w represents each character in the text, and n represents the total length of the text. The text sequence is then input into a deep neural network (DNN) to obtain the user's current emotional state C. g .

[0009] Step 2: Obtain the user's current metacognitive ability characteristics based on the user's personality and current dialogue information.

[0010] The long sequence text generated in step one by obtaining the current dialogue information A new text sequence representation C is generated through encoder E. t Specifically, as shown in equation (a) below:

[0011] C t =E(w1,w2,...,w n (a).

[0012] Next, personality information of the current user is obtained using scales or questionnaires, assuming an m-dimensional vector P = {p1, p2, ..., p...}. mThe value of m is taken as a representation of the user's personality, where m ranges from 100 to 200. Then, the personality representation vector P is re-encoded by encoder E to obtain a new code. Specifically, as shown in equation (b) below:

[0013]

[0014] Then, by combining the specific text representations in the current dialogue information, a metacognitive feature representation H of the user's current state is generated. s Specifically, as shown in equation (c) below:

[0015]

[0016] Where Z is the metacognitive attention parameter represented by the following equation (d):

[0017]

[0018] in, C represents a text sequence t The transpose of .

[0019] Step 3: Integrate the user's current emotions into personalized metacognitive ability features to obtain a representation of the user's current state.

[0020] First, the long sequence of text C formed by obtaining the current dialogue information in step two is represented as C. t The hidden layer representation H is obtained through a bidirectional long short-term memory artificial neural network Bi-LSTM. t Specifically, as follows (e -1 ) and (e -2 As shown in the formula:

[0021] H t =BiLSTM(C t ) (e -1 );

[0022] H t =[h1,h2,...,h n ] (e -2 ).

[0023] Next, through the attention mechanism, combined with the user's current emotion C obtained in steps one and two, g Metacognitive ability representation H s The final overall user current state F is obtained, as shown in equation (f) below:

[0024] F = W f d+b f (f).

[0025] Among them, Wf For trainable parameters; b f Let be the trainable parameters; d is as shown in equation (g) below:

[0026]

[0027] Among them, h i H represents the hidden layer. t The i-th term; β i The attention weights are expressed by the following equation (h):

[0028]

[0029] Among them, h k H represents the hidden layer. t The kth term; A is the emotional metacognitive representation expressed by the following equation (i); e(h i A) is represented by the following equation (j):

[0030] A = [C g H s ] (i);

[0031] e(h i A) = v T tanh(W h h i +W a A+b) (j).

[0032] Where v is the weight vector; v T W is the transpose of v; h W a b and b are the parameters that can be trained.

[0033] Step 4: Construct language generation conditions and input them into the language model to generate several heuristic emotion support questions.

[0034] First, a semantic alignment module is designed to align the fused user current state F with the natural language in the dialogue information in terms of dimensions, thereby obtaining the aligned user current state representation. Based on this, the user's current state is used as a concatenation and fusion condition for language generation, and then guiding prompts are designed as language generation prompts, which are then combined with the original current dialogue information C. t They are fed together into a large generative language model to generate multiple heuristic sentiment support questions, as shown in equation (k) below:

[0035]

[0036]

[0037] Wherein, Input represents the content that is ultimately input into the large generative language model, Q represents manually designed guiding prompts that guide the model to generate heuristic questions, LLM represents a large generative language model, and Output represents the e heuristic sentiment support questions that are ultimately generated.

[0038] Step 5: Use a matching algorithm based on user metacognitive features to select the most appropriate question to ask.

[0039] Using the multiple heuristic emotional support questions obtained in step four as candidates, each of these questions is encoded to obtain a representation O for each question, as shown in equation (l) below:

[0040] [O1,O2,...,O e ] = E(Output) (l).

[0041] Next, a high-dimensional metacognitive space is constructed using cognitive psychology theory. The user's metacognitive features Hs and the sequence representation O obtained from encoding each candidate question are mapped into this space, as shown in equations (m) to (n) below:

[0042]

[0043]

[0044] Where W and W' are trainable mapping matrices, and then, based on this space, the distance between the high-dimensional representation of each candidate question and the user's metacognitive high-dimensional representation is calculated, i.e., the distances expressed by the following equations (n-1) and (n-2):

[0045]

[0046]

[0047] By selecting the questions that have the smallest distance, i.e., the best match with the user's current metacognitive characteristics, we can help them reflect on the formation of negative emotions and regulate themselves to get out of negative emotions, thereby improving their metacognitive abilities in terms of emotions and helping them to better control their emotions.

[0048] Step 6: Feed the user feedback and personalized metacognitive ability feature representation Hs back into the language model to enter the next iteration.

[0049] After receiving new feedback from the user, the new question, learner feedback, and historical dialogue information are concatenated to obtain new dialogue information. Then, the method described above is repeated to recalculate the new user's current state representation. The new dialogue information is then re-input into the language model. By stimulating the language model's own hidden language capabilities, the strategy for generating questions in the next round is changed, and new emotional support questions are generated. This process is repeated iteratively until the learner's mood improves.

[0050] Compared with existing technologies, this invention combines the user's metacognitive abilities during the user dialogue process to propose emotional support questions that are more tailored to the user's characteristics. It also guides users to reflect on and introspect on their negative emotions through heuristic questions, thereby enhancing the user's emotional metacognition, that is, their ability to control their own emotions. From the perspective of improving emotional metacognition, it can more effectively solve their emotional problems and further improve the user experience. Attached Figure Description

[0051] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the specific embodiments and accompanying drawings. Except for the contents specifically mentioned below, the processes, conditions, and experimental methods for implementing the present invention are all common knowledge and general knowledge in the art, and the present invention does not have any particular limitations.

[0053] Example 1

[0054] See Figure 1 Follow these steps to provide personalized metacognitive enhancement and heuristic emotional support:

[0055] Step 1: Obtain the user's current emotional state based on the current dialogue information.

[0056] The current round of historical dialogue information can be obtained in real time during the conversation with the user.

[0057] Assuming the current round of historical dialogue information t represents the current round of the dialogue, and s represents the user's statements in each round. First, all historical dialogue information is concatenated to form a long sequence of text. Here, w represents each character in the text, and n represents the total length of the text. The text sequence is then input into a deep neural network (DNN) to obtain the user's current emotional state C. g .

[0058] Step 2: Obtain the user's current metacognitive ability characteristics based on the user's personality and current dialogue information.

[0059] The user's personality information was obtained using the professional psychological scale BFI, and the results were presented in the form of the Big Five personality traits.

[0060] In this embodiment, the user's personality is represented in 100 dimensions.

[0061] First, the long sequence of text generated in step one by obtaining the current dialogue information is processed. A new text sequence representation C is generated through encoder E. t Specifically, as shown in equation (a) below:

[0062] C t =E(w1,w2,...,w n (a).

[0063] Next, the Big Five personality traits of the current user are obtained using methods such as questionnaires, and then projected to one hundred dimensions using an upgraded approach, P = {p1, p2, ..., p...} m The personality representation vector P is used as the user's personality representation, where m = 100. Then, the personality representation vector P is also re-encoded by encoder E to obtain a new code. Specifically, as shown in equation (b) below:

[0064]

[0065] Then, by combining the specific text representations in the current dialogue information, the user's current metacognitive feature representation Hs is generated, as shown in equation (c) below:

[0066]

[0067] Where Z is represented by the following equation (d):

[0068]

[0069] in, C represents a text sequence t The transpose of.

[0070] Step 3: Integrate the user's current emotions into personalized metacognitive ability features to obtain a representation of the user's current state;

[0071] First, the long sequence of text C formed by obtaining the current dialogue information in step two is represented as C. t The hidden layer representation H is obtained through a bidirectional long short-term memory artificial neural network Bi-LSTM. t Specifically, as follows (e -1 ) and (e -2 As shown in the formula:

[0072] H t =BiLSTM(C t ) (e -1 );

[0073] H t =[h1,h2,...,h n ] (e-2 ).

[0074] Next, through the attention mechanism, combined with the user's current emotion C obtained in steps one and two, g Metacognitive ability representation H s The final overall user current state F is obtained, as shown in equation (f) below:

[0075] F = W f d+b f (f).

[0076] Among them, W f For trainable parameters; b f Let be the trainable parameters; d is as shown in equation (g) below:

[0077]

[0078] Among them, h i H represents the hidden layer. t The i-th term; β i The attention weights are expressed by the following equation (h):

[0079]

[0080] Among them, h k H represents the hidden layer. t The kth term; A is the emotional metacognitive representation expressed by the following equation (i); e(h i A) is represented by the following equation (j):

[0081] A = [C g H s ] (i);

[0082] e(h i A) = v T tanh(W h h i +W a A+b) (j).

[0083] Where v is the weight vector; v T W is the transpose of v; h W a b and b are the parameters that can be trained.

[0084] Step 4: Construct language generation conditions and input them into the language model to generate several heuristic emotion support questions.

[0085] In this embodiment, the large-scale generative language model is GPT4.0.

[0086] First, a semantic alignment module is designed to align the fused user current state F with the natural language in the dialogue information in terms of dimensions, thereby obtaining the aligned user current state representation. Based on this, the user's current state is used as a language generation condition through concatenation and fusion. Guiding prompts are then designed as language generation prompts and fed into a large-scale generative language model along with the original current dialogue information Ct to generate multiple heuristic emotion support questions. The specific implementation is shown in equation (k) below:

[0087]

[0088]

[0089] Wherein, Input represents the content that is ultimately input into the large generative language model, Q represents manually designed guiding prompts that guide the model to generate heuristic questions, LLM represents a large generative language model, and Output represents the e heuristic sentiment support questions that are ultimately generated.

[0090] Step 5: Use a matching algorithm based on user metacognitive features to select the most appropriate question to ask.

[0091] Using the multiple heuristic emotional support questions obtained in step four as candidates, each of these questions is encoded to obtain a representation O for each question, as shown in equation (l) below:

[0092] [O1,O2,...,O e ] = E(Output) (l).

[0093] Next, a high-dimensional metacognitive space is constructed, and the user's metacognitive features H are respectively... s The sequence representation O obtained by encoding each alternative problem is mapped to this space, as shown in equations (m) to (n) below:

[0094]

[0095]

[0096] Where W1 and W2 are trainable mapping matrices, and then, based on this space, the distance between the high-dimensional representation of each candidate question and the user's metacognitive high-dimensional representation is calculated, i.e., the following (n -1 ) and (n -2 Distance expressed by the formula:

[0097]

[0098]

[0099] By selecting the questions that have the smallest distance, i.e., the best match with the user's current metacognitive characteristics, we can help them reflect on the formation of negative emotions and regulate themselves to get out of negative emotions, thereby improving their metacognitive abilities in terms of emotions and helping them to better control their emotions.

[0100] Step Six: Feed user feedback and the user's own metacognitive features back into the language model to proceed to the next iteration.

[0101] After receiving new feedback from the user, the new question, learner feedback, and historical dialogue information are concatenated to obtain new dialogue information. Then, the method described above is repeated to recalculate the new user's current state representation. The new dialogue information is then re-input into the language model. By stimulating the language model's own hidden language capabilities, the strategy for generating questions in the next round is changed, and new emotional support questions are generated. This process is repeated iteratively until the user's mood improves.

[0102] This invention can combine the user's metacognitive abilities during the user dialogue process to propose emotional support questions that are more tailored to the user's characteristics; and guide the user to reflect on and introspect on their own negative emotions through heuristic questions, thereby improving the user's emotional metacognition, that is, the ability to control their own emotions; at the same time, compared with the existing technology, starting from the source of the user's low mood, it can more effectively solve their emotional problems and further improve the user experience.

[0103] The scope of protection of this invention is not limited to the above embodiments. Any variations and advantages that can be conceived by those skilled in the art without departing from the spirit and scope of the inventive concept are included in this invention and are protected by the appended claims.

Claims

1. A personalized metacognition-enhanced heuristic emotional support dialogue method, characterized by, This method employs a personality-driven metacognitive ability extraction approach, utilizing a language model to generate a set of heuristic questions. Based on an evaluation algorithm that considers user personality and current emotional state, the most appropriate question is selected as the response. This approach aims to enhance emotional support by improving users' emotional control. The specific steps of this method are as follows: Step one, obtaining the current emotional state of the user according to current dialogue information ; Step 2: Based on the user's personality and current dialogue information, obtain the current metacognitive ability characteristics; Step 3: Integrate the user's current emotions into personalized metacognitive ability features to obtain a representation of the user's current state; Step 4: Construct language generation conditions and input them into the language model to generate several heuristic emotion support questions; Step 5: Use a matching algorithm based on user metacognitive features to select the most appropriate question to ask; Step 6: Feed user feedback and your own metacognitive features back into the language model and proceed to the next iteration until the user's mood improves. In step two, the process of acquiring the current metacognitive ability characteristics is as follows: 2-1: Acquire long-sequence text generated by current dialogue information by the encoder E, a text sequence represented by the following (a) formula : (a); 2-2: Let m-dimensional vector For the user's character representation, re-encode it through the encoder E to obtain an encoding represented by the following (b) formula : (b); 2-3: Encoding The data is fed into a softmax layer and a layernorm layer to obtain a personalized metacognitive ability feature representation as shown in equation (c) below. : (c); wherein, The meta-cognitive attention parameter is represented by the following (d) formula: (d); wherein is the transpose of the text sequence ​ In step three, the user's current state is obtained as follows: 3-1: Text sequence representation formed from current dialogue information The hidden layer representation represented by the following (e) formula is obtained by a bidirectional long short-term memory artificial neural network Bi-LSTM : (e); 3-2: Representing emotions and personalized metacognition feature representation by the attention mechanism to obtain a user current state representation represented by the following (f) formula : (f); wherein is represented by the following formula (g): (g); wherein, is an attention weight represented by the following (h): (h); wherein is an emotional meta-cognition representation expressed by the following (i) formula; is expressed by the following (j) formula: (i); (j); wherein, is a weight vector; is the transpose of , and are parameters that can be trained; In step five, the most appropriate question is selected and asked as follows: 5-1: Based on the multiple heuristic emotional support questions obtained in step four as candidates, the questions are encoded to obtain the sequence representation O of the following equation (l): (l); wherein, O for each question's representation; 5-2: Construct a high-dimensional metacognitive space based on cognitive psychology to represent the personalized metacognitive ability characteristics of users. The sequence representation O is mapped to the metacognitive high-dimensional space, resulting in the personalized metacognitive representations after mapping, represented by the following equations (m) and (n). and sequence characterization : (m); (n); wherein, is the trained mapping matrix; 5-3: Based on the metacognitive high-dimensional space, calculate the distance between the high-dimensional representation of each candidate question and the user's metacognitive high-dimensional representation. The question with the smallest distance is the one that best matches the user's current metacognitive features.

2. The personalized metacognitive enhancement heuristic emotional support dialogue method according to claim 1, characterized in that, In the step one, the current emotional state The acquisition process of the current emotional state is specifically as follows: 1-1: Form a long sequence text by splicing all the historical conversation information of the user where w represents each character in the text, and n represents the total length of the text. 1-2: inputting the text sequence into a deep neural network DNN to obtain the current emotional state of the user .

3. The personalized metacognition-augmented heuristic emotional support dialogue method of claim 1, wherein, In step four, the generation of heuristic emotional support questions is as follows: 4-1: Representing emotions and personalized metacognitive features Fusing; user current state representation aligned in dimension with natural language in dialogue information user current state representation aligned in dimension with natural language in dialogue information , constructing a language generation condition represented by the following (k) formula: (k); in, This is the content that is ultimately input into the language model; Human-designed guiding prompts guide the language model to generate thought-provoking questions; For language models; The final e heuristic emotional support questions are generated.

4. The personalized metacognitive enhancement heuristic emotional support dialogue method according to claim 1, characterized in that, In step six, user feedback and personalized metacognitive ability features are represented. The language model is then sent back for the next iteration, as detailed below: 6-1: Combine the user's feedback questions, learner feedback, and historical dialogue information to obtain new dialogue information, and repeat the iterations of steps one through five to recalculate the new user's current state representation; 6-2: Input new dialogue information and user metacognitive state into the language model. By stimulating the language model's own hidden language capabilities, change the strategy for generating questions in the next round and generate new emotional support questions. Repeat this process until the user's mood improves.

Citation Information

Patent Citations

  • Psychological counseling (dialogue) system and method based on artificial intelligence

    CN111667926A

  • Personality-enhanced dialogue emotion prediction model and construction method thereof

    CN114970561A