Personality recognition method based on emotion information

Abstract

The invention provides a personality recognition method based on emotion information, and the method comprises the steps: 1, defining personality recognition in a session, and formalizing a personality recognition problem in the session into a natural language reasoning I problem; step 2, constructing emotional dialogue content; step 3, constructing personality description; sorting and summarizing positive and negative descriptions of each personality traits, and providing detailed information about behavior characteristics and emotion expression tendencies as assumptions in the NLI problem; and 4, training and reasoning natural language reasoning. According to the invention, the interaction quality and personalized experience of man-machine interaction are improved.

Description

Technical Field

[0001] The present application relates to the technical field of large language model recognition, and in particular to a personality recognition method based on emotion information. Background Art

[0002] In modern education scenarios, large language models and human-computer interaction technologies are revolutionizing the way learning and teaching are done, including personalized learning based on artificial intelligence, emotion-aware learning systems, and adaptive learning paths. It is crucial to quickly and accurately identify the user's personality from the conversation content, better understand the user's personalized needs and requirements, and ultimately be able to provide high-quality services, as shown in the following examples:

[0003] Scenario: A user is reviewing for an upcoming math test and is using an emotion-aware learning system to help with preparation.

[0004] User: Oh, I feel frustrated every time I think about these math problems. I feel like I can never figure them out.

[0005] AI System (General Comfort Response): I understand that math problems can sometimes test our patience. Let's look at the specific problem you're having and maybe we can find a way to solve it.

[0006] User: I feel like whenever I have a problem, I feel better if I discuss it with someone. I often feel helpless when I'm alone.

[0007] AI system (response after observing an Extraverted Feeling personality): It sounds like you feel more comfortable and motivated when communicating with people. Let's try this approach: I will walk you through some problems, and you can try to explain the solution process to a friend, which will not only help you understand better, but also make the learning process more interesting.

[0008] User: That's a great idea! I like studying with my friends.

[0009] AI System: Great! I’m here to provide support when you are ready to discuss these topics. Also, if you would like, we can arrange some online group study sessions so you can get more insights and support in your discussions.

[0010] User: I am looking forward to this way of learning, thank you for your suggestions!

[0011] Although important, personality recognition in conversation (PRC) also faces significant challenges. PRC strives to classify the personality of a specified speaker into predefined personality categories (as shown in Table 1) by analyzing the conversation content. However, personality is a relatively permanent trait that describes the consistency and individuality of a person's behavior. Accurate personality assessment usually requires access to detailed personal information or long-term historical data, such as self-report papers, questionnaires, or longitudinal records from social media, which are not always available in daily conversation scenarios.

[0012] Recent studies have addressed the PRC problem by leveraging the advanced semantic understanding capabilities of deep neural networks to analyze conversation content. Specifically, researchers design conversation encoding models or fine-tune pre-trained language models for PRC. However, these approaches focus primarily on explicit conversation content, while ignoring two major issues that hinder their recognition performance. First, the semantic information contained in short utterances in a conversation is inherently limited. These approaches ignore key implicit factors contained in the conversation, such as emotions and attitudes that reflect the speaker's personality, which hinders recognition accuracy. Second, focusing only on the input conversation content ignores the semantic understanding of personality category labels, thereby ignoring additional discriminative cues and reducing the interpretability of the results.

[0013] Specifically, the current model still has the following problems to be solved:

[0014] Technical Problem 1: How can a large language model (LLM) be able to capture and understand implicit factors such as emotions and attitudes in conversation analysis as sensitively as humans? In conversation content analysis, previous studies have often focused only on explicit text content (such as the literal meaning of the discourse), while ignoring implicit factors that can reflect the speaker's personality, especially emotions and attitudes. Emotions are an important factor in expressing individual personality traits. For example, an individual who tends to be neurotic may show more negative emotions in a conversation. If these implicit information is not captured and analyzed, the accuracy of personality recognition will be limited.

[0015] Technical Problem 2: How can a large language model (LLM) deeply understand the semantics of personality category labels to improve the accuracy and interpretability of personality recognition? Traditional personality recognition methods mainly focus on the input conversation content, which has rich behavioral and emotional connotations. If these connotations are not fully understood by the model, it is difficult to effectively use these labels to improve the accuracy of recognition and the interpretability of the results. In addition, it is difficult to fully capture and utilize the subtle differences in these personality descriptions based solely on the literal information of the conversation content, which affects the final personality recognition performance.

[0016] Technical Problem 3: How can a large language model (LLM) accurately identify emotional changes in a conversation in real time to support real-time personality recognition? In practical applications, it is a technical challenge to accurately obtain the speaker's emotional state in real time during a conversation. Changes in emotional state reflect the personality traits of an individual, so accurate recognition of emotions is crucial for personality recognition. However, existing technologies often have difficulty capturing and analyzing emotional changes in a conversation in real time, especially when the conversation content is complex or the emotional expression is not obvious. Summary of the invention

[0017] This application provides a personality recognition method based on emotional information, which can be used to solve the technical problem of difficulty in real-time capturing and analyzing emotional changes in conversations.

[0018] The present application provides a personality recognition method based on emotional information, the method comprising:

[0019] Step 1: Define the personality recognition in conversation and formalize the personality recognition problem in conversation into a natural language inference (NLI) problem.

[0020] Step 2: Construct emotional dialogue content;

[0021] Step 3: Construct a personality description; organize and summarize the positive H of each personality trait. pos and negative H neg Description: Descriptions are used as hypotheses in NLI questions to provide detailed information about behavioral characteristics and emotional expression tendencies;

[0022] Step 4: Train and reason about natural language inference (NLI).

[0023] Furthermore, in step 1, personality recognition in conversation is defined, and the personality recognition problem in conversation is formalized as a natural language inference (NLI) problem; including:

[0024] Step 11, define the problem of personality identification in conversation;

[0025] Step 12, the goal is to identify speaker s’ personality traits into predefined personality categories using Jung’s eight personality theories, using an eight-dimensional binary vector to represent each trait;

[0026] Step 13, the PRC problem is formalized into a natural language inference (NLI) problem to determine the relationship between the conversation content, i.e., premise, and the personality label description, i.e., hypothesis, in order to identify each specific personality feature y of the speaker s given the input conversation content p , sentiment NLI maximizes two probabilities P pos and P neg The method is as follows:

[0027] Ppos +P neg =P(y p |X af ;H pos (p))+P(1-y p |X af ;H neg (p)) (1)

[0028] X af ={(u1,a1),(u2,a2),...,(u m ,a m )} By adding the sentiment annotation a i Attach to each session u i The emotional dialogue content obtained, H pos (p) and H neg (p) are natural language descriptions of people with and without a specific personality trait p, respectively.

[0029] Furthermore, step 2, constructing emotional dialogue content, includes:

[0030] Step 21, use the pre-trained language model Llama-7b (M ERC ) to perform emotion recognition in the conversation, for each utterance u i Assign emotional labels i (i=1,2,…,m), the method is as follows:

[0031] e1,e2,...,e m =M ERC (u1,u2,...,u m ) (2)

[0032] Step 22, combine each utterance with the sentiment annotation, given a set of dialogue content X = {u1,u2,...,u m}, interlocutor s and other interlocutors s other , and a set of sentiment labels {e1,e2,...,e m}, use the following manual template to encapsulate the emotional description and generate a dialogue content X with emotion af :

[0033] Step 221, initialize X af is empty;

[0034] Step 222, for each dialogue u in X i , i=1,2,…,m: if i=1, i.e. the first sentence of the dialogue: then execute step 223, otherwise execute step 224;

[0035] Step 223, if ui From s, then a i Set to "the initial value of s is e i "; otherwise, a i Set to "Initial S other for e i "; then execute step 225;

[0036] Step 224, if s is not 1: If u i From s, then a i Set to "s to e i Response"; otherwise, a i Set to "Then S other becomes e i ”;

[0037] Step 225, u i and a i Append the combination to X af ;

[0038] Furthermore, in step 4, training and reasoning natural language inference (NLI) includes:

[0039] Step 41: By constructing NLI prompt samples, combined with the emotional dialogue content, i.e. premise X af and personality description, i.e. hypothesis H, to construct NLI prompt samples; the method is as follows:

[0040]

[0041] in, is the encapsulated sample, p is the individual feature, and [MASK] is the mask used for filling;

[0042] Step 42: During training, adjust the positive and negative labels of the samples to optimize the model according to the actual values ​​of the personality traits, as follows:

[0043]

[0044] Step 43: “yes” and “no” are the basic facts of the NLI sample; according to the NLI sample, four probabilities are expressed as follows:

[0045]

[0046] Step 44: Use the P-Tuning fine-tuning method to fill [MASK] with the last word generated by Llama; the training goal is to fine-tune the parameters θ in the model M to maximize the sum of probabilities as follows:

[0047] argmax(P pos (yes)+Pneg (no))*y p +P neg (yes)+P pos (no))*(1-y p ) (6)

[0048] Step 45: During inference, compare the probabilities of the positive and negative descriptions to determine the final personality recognition result; the method is as follows:

[0049] argmax[P neg (yes)+P pos (no),P pos (yes)+P neg (no)] (7)

[0050] In order to solve the problem of enabling LLM to sensitively capture and understand implicit factors in conversation analysis, this application integrates emotion recognition and natural language reasoning to enable LLM to use the description of personality traits to infer whether the traits can be supported by given conversation content, just like human analysis, thereby sensitively and accurately interpreting personality traits from conversation clues.

[0051] In order to enable LLM to deeply understand the semantics of personality category labels, personality category labels are introduced to enable LLM to understand and interpret the deep semantics of personality traits. By framing personality recognition as an NLI problem, the model can evaluate whether the conversation content supports a specific personality description. In addition, by injecting emotion annotations into the conversation, the model's understanding of the relationship between emotion and personality is further enhanced, thereby improving the accuracy and explanatory power of personality recognition.

[0052] To address the issue of how LLM can accurately identify emotional changes in conversations in real time to support real-time personality recognition, LLM can capture emotional changes in conversations in real time through specific emotion recognition fine-tuning; combine each conversation sentence with the corresponding emotion annotation to form emotional conversation content, enhancing the ability to understand and analyze emotions; and construct personality recognition as a natural language inference (NLI) problem. The model can instantly and accurately infer personality traits during the conversation, improving the interactive quality and personalized experience of human-computer interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A flowchart provided for an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0055] To fill the current research gap, this application proposes sentiment natural language inference, which aims to accurately and interpretably identify personality in conversations through sentiment annotations of conversation content (i.e., discrete emotion category labels) and natural language descriptions of personality labels. There is a strong correlation between personality and emotional expression in conversation, however, accurately obtaining real-time sentiment annotations of conversation content for personality identification is not a simple task. This application fine-tunes a pre-trained language model specifically for emotion identification in conversations, facilitating real-time sentiment annotation of utterances. In addition, the semantics behind personality labels contain descriptions of individual behavioral characteristics and emotional expression tendencies, providing additional basis for personality identification. In order to leverage them to solve our problem, this application adopts textual descriptions of personality labels from psychological findings and formulates personality identification as a natural language inference (NLI) problem.

[0056] Specifically, the goal is to determine whether the speaker's personality description (as a hypothesis) is correct given the content of the conversation (as a premise). This NLI problem is solved by constructing a natural language prompt (including the above premise and hypothesis) to fine-tune the pre-trained language model.

[0057] Compared to existing methods, Emotion NLI uses auxiliary emotion information to enhance the limited semantics in the conversation content. Emotion NLI then links the language and behavior of the speaker in the conversation with the description of personality traits, producing more interpretable recognition results. In addition, emotion natural language inference mainly operates on the input dialog content and is compatible with large pre-trained language models in various architectures.

[0058] Emotional NLI can be effectively used for personality identification in real-world human-computer interaction applications. However, personality is a multifaceted concept that has been studied in various disciplines including physiology, linguistics, and cognitive science. Therefore, rigorous personality assessment still relies on professional methods. The purpose and advantage of emotional NLI is to be able to make a preliminary assessment of the user's personality in a conversational environment, thereby promoting a better conversational experience.

[0059] The tool learning method for large language models provided in this application is based on the following definitions:

[0060] Definition 1: A large language model is an artificial intelligence system trained with a large amount of text data that can understand and generate natural language and is used to perform various language-related tasks such as translation, text generation, and summarization.

[0061] Definition 2: The problem of personality recognition in conversation (PRC) is to identify the speaker's personality traits by analyzing the text conversation content. This is a particularly important task in human-computer interaction (HCI) because it can help the system better understand the user's personalized needs and preferences, thereby providing more personalized services.

[0062] Definition 3: Natural Language Inference (NLI), a branch of computational linguistics, is a system that determines whether a given "hypothesis" is true (implication), false (contradiction), or undetermined (neutral) based on a given "premise";

[0063] Definition 4: Affective Natural Language Inference, which is an accurate and interpretable personality identification method proposed. It uses the sentimentality in the conversation content and frames personality identification as an NLI problem, where the textual description of personality traits is checked against the conversation content;

[0064] Definition 5: Jung's eight personality types theory is based on four psychological functions of thinking, feeling, intuition, and emotion, as well as two psychological attitudes of introversion and extroversion. The combination of these functions and attitudes forms eight different personality types, and each person's behavior pattern and decision-making method are significantly influenced by their dominant psychological functions and attitudes;

[0065] Definition 6: Pre-trained language models are deep learning models pre-trained on large-scale corpora for understanding and generating natural language. These models learn the semantic and syntactic structures of words, phrases, and long articles by capturing the statistical properties of language;

[0066] Definition 7: Sentiment annotations are sentiment labels assigned to parts of a conversation, helping to infer the speaker’s emotional state and personality traits. These annotations are derived from the emotions expressed in the conversation.

[0067] Definition 8: Llama-7b is a multimodal large-scale language model based on a generative adversarial network with approximately 7 billion parameters.

[0068] Definition 9: P-tuning is a parameter adjustment technique in natural language processing that only fine-tunes a small number of additional parameters while freezing most of the parameters of the pre-trained language model.

[0069] The method provided in this application specifically comprises the following steps:

[0070] Step 1: Define personality recognition in conversation. First, the problem that this patent wants to solve is clarified, which specifically includes the following steps:

[0071] Step 11, define the problem of personality recognition in conversation, the PRC problem is as described in Definition 2;

[0072] Step 12, the goal is to identify the personality traits of speaker s as predefined personality categories using Jung's eight personality theories, using an eight-dimensional binary vector to represent each trait. For details, please refer to Table 1; Jung's eight personality theories are as described in Definition 5.

[0073] Step 13, the PRC problem is formalized into a natural language inference (NLI) problem to determine the relationship between the conversation content, i.e., premise, and the personality label description, i.e., hypothesis, in order to identify each specific personality feature y of the speaker s given the input conversation content p , sentiment-NLI maximizes two probabilities P pos and P neg The method is as follows:

[0074] P pos +P neg =P(y p |X af ;H pos (p))+P(1-y p |X af ;H neg (p)) (1)

[0075] Among them, the NLI problem is as shown in Definition 3, X af ={(u1,a1),(u2,a2),...,(u m ,a m )} By adding the sentiment annotation a i Attach to each session u i The emotional dialogue content obtained, H pos (p) and H neg (p) are natural language descriptions of people with and without a specific personality trait p, respectively.

[0076] Table 1: Positive and negative label descriptions of Jung's eight personality theories

[0077]

[0078]

[0079] Step 2: Construct emotional dialogue content, which includes the following sub-steps:

[0080] Step 21: Use the pre-trained language model Llama-7b (M ERC ) to perform emotion recognition in the conversation, for each utterance u i Assign emotional labels i (i=1,2,…,m), the method is as follows:

[0081] e1,e2,...,e m =M ERC (u1,u2,...,u m ) (2)

[0082] Wherein, the pre-trained language model is as described in Definition 6; Llama is as described in Definition 8;

[0083] Step 22: Combine each utterance with the sentiment annotation. Given a set of dialogue content X = {u1,u2,...,u m}, interlocutor s and other interlocutors s other , and a set of sentiment labels {e1,e2,...,e m}, use the following manual template to encapsulate the emotional description and generate a dialogue content X with emotion af :

[0084] Step 221, initialize X af is empty;

[0085] Step 222, for each dialogue u in X i , i=1,2,…,m: if i=1, i.e. the first sentence of the dialogue: then execute step 223, otherwise execute step 224;

[0086] Step 223, if u i From s, then a i Set to "the initial value of s is e i "; otherwise, a i Set to "Initial other for e i "; then execute step 225;

[0087] Step 224, if s is not 1: If u i From s, then a i Set to "s to e i Response"; otherwise, a i Set to "Then s other becomes e i ”;

[0088] Step 225, u i and a i Append the combination to X af ;

[0089] After the algorithm is finished, X af Contains complete emotional dialogue content.

[0090] Among them, the sentiment annotation is as described in Definition 7;

[0091] These emotion descriptions add additional emotional context to each utterance, thus providing additional semantic information when analyzing personality traits;

[0092] Step 3: Construct personality description, which includes the following sub-steps:

[0093] Organize and summarize the positive H of each personality trait posand negative H neg Description: Descriptions are used as hypotheses in NLI problems to provide detailed information about behavioral characteristics and emotional expression tendencies, allowing the model to understand complex academic terms and personality descriptions using everyday language;

[0094] Step 4: Train and reason about the natural language inference (NLI) part. The specific steps are as follows:

[0095] Step 41: By constructing NLI prompt samples, combined with the emotional dialogue content, i.e. premise X af and personality description, i.e. hypothesis H, to construct NLI prompt samples; the method is as follows:

[0096]

[0097] in, is the encapsulated sample, p is the individual feature, and [MASK] is the mask used for filling;

[0098] Step 42: During training, adjust the positive and negative labels of the samples to optimize the model according to the actual values ​​of the personality traits, as follows:

[0099]

[0100] Step 43: “yes” and “no” are the basic facts of the NLI sample; according to the NLI sample, four probabilities are expressed as follows:

[0101]

[0102] Step 44: Use the P-Tuning fine-tuning method. The P-Tuning fine-tuning method is as shown in Definition 3. Specifically for sentiment natural language inference, the last word generated by Llama is used to fill [MASK]. The training goal is to fine-tune the parameter θ in the model M to maximize the sum of probabilities. The method is as follows:

[0103] argmax(P pos (yes)+P neg (no))*y p +(P neg (yes)+P pos (no))*(1-y p )#(6)

[0104] Step 45: During inference, compare the probabilities of the positive and negative descriptions to determine the final personality recognition result; the method is as follows:

[0105] argmax[P neg (yes)+P pos (no),P pos(yes)+P neg (no)]#(7).

[0106] The present application is further described below in conjunction with specific embodiments:

[0107] Example:

[0108] Examples of users using this patent to conduct Q&A:

[0109] 1. The user describes their mental state. For example, "Every time I practice this speech, I feel nervous and my heart beats faster. I have tried practicing alone, but I find that I feel more comfortable and better able to control my emotions when I practice in front of others."

[0110] 2. Identify the user's personality traits as predefined personality categories, use an eight-dimensional binary vector to represent each trait in Jung's eight personality theories, and determine the relationship between the conversation content (premise) and the personality label description (hypothesis).

[0111] 3. Use the specially adjusted pre-trained language model Llama-7b for PRC, tag each utterance with emotion, combine the utterance with its emotional annotation, and encapsulate its emotional description through manual templates to form emotionally rich conversation content, provide additional emotional context for the utterance, and provide more semantic information to assist in analyzing personality traits.

[0112] 4. Based on the user's conversation, summarize the positive (H pos ) and negative (H neg ) description, which is used for assumptions in natural language reasoning problems and details behavioral characteristics and emotional expression tendencies.

[0113] Construct NLI prompt samples, combine emotional conversation content and personality trait descriptions, fine-tune the pre-trained language model, adjust the positive and negative labels of the samples according to the actual personality trait values ​​to optimize the model performance, and determine the final personality recognition result by comparing the probability and negative descriptions.

[0114] This application proposes a natural language inference method based on emotion. By combining emotion annotation and natural language inference, the model's deep understanding of the content of the conversation and the sensitivity of emotional changes are enhanced, thereby improving the accuracy and interpretability of personality recognition. In addition, the real-time solution of the method provided by this application enables it to adapt to the rapidly changing conversation situation through special adjustments to the pre-trained model, and realizes real-time and accurate annotation of emotions. This application innovatively introduces emotional annotations into the content of the conversation, which helps to understand the potential emotions and attitudes reflected in the speech and improves the accuracy of personality assessment. Innovatively proposes natural language inference (NLI) for personality recognition, and by treating personality recognition as an NLI problem, uses the descriptive power of natural language to determine whether the content of the conversation supports the description of personality traits. Innovatively use the advanced language understanding ability of the pre-trained language model to interpret the personality characteristics in the conversation, and fine-tune the pre-trained language model to perform emotion recognition in the conversation. Innovatively proposes a real-time interpretability system, which aims to provide real-time insights into personality characteristics and enhances its applicability in interactive scenarios such as human-computer interaction, so that understanding the user's personality can effectively customize the response.

[0115] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A personality recognition method based on emotional information, characterized in that: The method comprises: Step 1: Define the personality recognition in conversation and formalize the personality recognition problem in conversation into a natural language inference (NLI) problem. Step 2: Construct emotional dialogue content; Step 3: Construct a personality description; organize and summarize the positive H of each personality trait. pos and negative H neg Description: Descriptions are used as hypotheses in NLI questions to provide detailed information about behavioral characteristics and emotional expression tendencies; Step 4: Train and reason about natural language inference (NLI).

2. The method according to claim 1, characterized in that Step 1: Define the personality recognition in conversation and formalize the personality recognition problem in conversation into a natural language inference (NLI) problem; including: Step 11, define the problem of personality identification in conversation; Step 12, the goal is to identify speaker s’ personality traits into predefined personality categories using Jung’s eight personality theories, using an eight-dimensional binary vector to represent each trait; Step 13, the PRC problem is formalized into a natural language inference (NLI) problem to determine the relationship between the conversation content, i.e., premise, and the personality label description, i.e., hypothesis, in order to identify each specific personality feature y of the speaker s given the input conversation content p , sentiment NLI maximizes two probabilities P pos and P neg The method is as follows: P pos +P neg =P(y p |X af ;H pos (p))+P(1-y p |X af ;H neg (p)) (1) X af ={(u1,a1),(u2,a2),...,(u m ,a m )} By adding the sentiment annotation a i Attach to each session u i The emotional dialogue content obtained, H pos (p) and H neg (p) are natural language descriptions of people with and without a specific personality trait p, respectively.

3. The method according to claim 1, characterized in that Step 2: Construct emotional dialogue content, including: Step 21, use the pre-trained language model Llama-7b (M ERC ) to perform emotion recognition in the conversation, for each utterance u i Assign emotional labels i (i=1,2,…,m), the method is as follows: e1,e2,...,e m =M ERC (u1,u2,...,u m ) (2) Step 22, combine each utterance with the sentiment annotation, given a set of dialogue content X = {u1,u2,...,u m }, interlocutor s and other interlocutors s other , and a set of sentiment labels {e1,e2,...,e m }, use the following manual template to encapsulate the emotional description and generate a dialogue content X with emotion af : Step 221, initialize X af is empty; Step 222, for each dialogue u in X i , i=1,2,…,m: if i=1, i.e. the first sentence of the dialogue: then execute step 223, otherwise execute step 224; Step 223, if u i From s, then a i Set to "s's initial value is e i "; otherwise, a i Set to "Initial other for e i "; then execute step 225; Step 224, if s is not 1: If u i From s, then a i Set to "s to e i Response"; otherwise, a i Set to "Then s other becomes e i ”; Step 225, u i and a i Append the combination to X af。 4. The method according to claim 1, characterized in that Step 4: Train and reason about natural language inference (NLI), including: Step 41: By constructing NLI prompt samples, combined with the emotional dialogue content, i.e. premise X af and personality description, i.e. hypothesis H, to construct NLI prompt samples; the method is as follows: in, is the encapsulated sample, p is the individual feature, and [MASK] is the mask used for filling; Step 42: During training, adjust the positive and negative labels of the samples to optimize the model according to the actual values ​​of the personality traits, as follows: Step 43: "yes" and "no" are the basic facts of the NLI sample; according to the NLI sample, four probabilities are expressed as follows: Step 44: Use the P-Tuning fine-tuning method to fill [MASK] with the last word generated by Llama; the training goal is to fine-tune the parameters θ in the model M to maximize the sum of probabilities as follows: argmax(P pos (tes)+P neg (no))*y p +(P neg (yes)+P pos (no))*(1-y p )(6) Step 45: During inference, compare the probabilities of the positive and negative descriptions to determine the final personality recognition result; the method is as follows: argmax[P neg (yes)+P pos (no),P pos (yes)+P neg (no)](7)。

Images