Abnormal psychological detection method based on multi-dimensional psychological feature modeling

Abstract

The invention relates to the technical field of social media platform analysis, and discloses an abnormal psychology detection method based on multi-dimensional psychological feature modeling, which comprises the following steps: setting an emotion large category label and a behavior sub category label according to a contradictory dispute text on a social media; collecting a data set of contradictory dispute texts on the social media; cue words are designed, and an emotion classification task and a behavior recognition classification task of the contradictory dispute text are combined into one task; selecting a large language model and a parameter scale of the large language model, and performing training fine tuning on the large language model by using LLaMA-Factory based on cue words and training samples in a training set; and inputting the contradictory dispute text and the prompt word in the test set into the trained large language model to obtain an abnormal psychological detection result. According to the method, the emotion classification spanning single dimension can be realized, and the comprehensive assessment of the individual psychological state is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of social media platform analysis, and in particular to an abnormal psychology detection method based on multi-dimensional psychological feature modeling. Background Art

[0002] In the modern society of information explosion, social media has become an important platform for people to express their opinions and share their lives. However, negative emotions on social media can also have many adverse effects. In order to effectively deal with this challenge, modeling and predicting the psychological characteristics of social media users has become an important research direction. By constructing psychological characteristic prediction methods and models, potential risk users can be identified in advance, and intervention measures can be taken in time.

[0003] With the rapid development of artificial intelligence technology, especially the emergence of large language models such as Qwen, new possibilities have been provided for the prediction of psychological characteristics. With the advantages of deep learning and natural language processing technology, these models can deeply analyze the emotional tendencies and psychological states in users' speech, providing strong support for the modeling of psychological characteristics. However, although existing studies have achieved certain results in the field of psychology, most of them focus on the perspective of victims and ignore the analysis and prediction of psychological dynamics from the perspective of the actors themselves.

[0004] This invention aims to expand this boundary, using Qwen's deep learning and natural language processing capabilities to build a comprehensive abnormal psychological detection system that focuses on identifying emotional states such as anxiety, depression, impulsivity and anger. Compared with previous technical solutions, the core value of this invention lies in its forward-looking nature, which not only focuses on the sequelae of conflicts that have already occurred, but also strives to warn of potential dispute risks through subtle changes in psychological characteristics before the problem escalates.

[0005] The challenge faced by the present invention is that, compared with the relatively mature assessment of depression and suicidal tendencies in the field of mental health, there is a lack of a unified assessment framework and standard for psychological abnormalities that cause conflicts and disputes. In addition, obtaining high-quality, labeled datasets of abnormal psychological behaviors is also a major problem, especially under the premise of protecting personal privacy and complying with data protection regulations. In view of this, it is necessary to innovatively develop a set of analysis models that can not only sensitively capture abnormal psychological signals, but also have high accuracy and versatility. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides an abnormal psychology detection method based on multi-dimensional psychological feature modeling. By integrating social media data, constructing a multi-dimensional psychological feature scale and annotated data set, and performing data fine-tuning modeling on a large language model, a more accurate multi-dimensional psychological emotion feature modeling and detection method is created, thereby realizing the efficient construction of psychological portraits of individuals or groups on social media.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] An abnormal psychology detection method based on multi-dimensional psychological feature modeling is used to classify emotions and identify behaviors of conflict and dispute texts, specifically including the following steps:

[0009] Step 1: Based on the conflict and dispute texts on social media, select the major emotion categories and the behavioral subcategories under each major emotion category, and set the major emotion category labels and the behavioral subcategory labels;

[0010] Step 2: Collect a dataset of conflict and dispute texts on social media, annotate and divide the dataset into training and test sets;

[0011] Step 3: Design prompt words to combine the emotion classification task and behavior recognition classification task of conflict and dispute text into one task;

[0012] Step 4: Select the large language model and the parameter scale of the large language model, and use LLaMA-Factory to train and fine-tune the large language model based on the prompt words and training samples in the training set; the training samples include the conflict and dispute texts used for training, as well as the corresponding emotion category labels and behavior subcategory labels;

[0013] Step 5: Input the prompt words and the conflict texts in the test set into the trained large language model to obtain abnormal psychological detection results, including the emotion category labels and behavior subcategories of the conflict texts.

[0014] Furthermore, step one specifically includes:

[0015] The broad categories of emotions included anxiety, depression, anger, fear, low self-esteem, fatigue, and impulsivity;

[0016] Behavioral subcategories of anxiety include: fidgeting, nervousness, and excessive worrying;

[0017] Behavioral subcategories of depression include: emotional suppression, loss of pleasure, and lack of aspiration;

[0018] Angry behavioral subcategories included: getting angry for no reason, being verbally abusive, and behaving recklessly;

[0019] Behavioral subcategories of fear include: fearful trembling and overt avoidance;

[0020] Behavioral subcategories of low self-esteem include: self-blame, shame, and low self-esteem;

[0021] Behavioral subcategories of fatigue included: distractibility, emotional lability, and lack of energy;

[0022] Impulsive behavior subcategories include: rapid unplanned reactions, lack of control over behavior, and lack of consideration for consequences.

[0023] Furthermore, step three specifically includes the following steps:

[0024] S31, preset roles and context:

[0025] In the prompt words, the identities of psychology professors and data processing experts are set for the large language model, and it is made clear that the task of the large language model is to classify emotions and identify behaviors of conflicting texts in social media according to the given large emotion category labels and behavior subcategory labels, so as to guide the model to generate more accurate, coherent and relevant texts;

[0026] S32, write all set emotion category labels and behavior subcategory labels into prompt words in a fixed json format and transmit them to the large language model, stipulate that the conflict and dispute text input to the large language model is in text format, and the output of the large language model is in json format including emotion category labels and behavior subcategory labels;

[0027] S33, in the prompt word, requires the large language model to first classify the emotions of the conflict text, and then detect the behavioral subcategories contained in the conflict text according to the identified large emotion categories.

[0028] Compared with the prior art, the beneficial technical effects of the present invention are:

[0029] 1. The present invention selects seven major emotion category labels - depression, anger, fatigue, fear, inferiority, anxiety, and impulse, as the macro-framework for the emotional classification of social media content. Under each major emotion label, it is further refined into multiple behavioral sub-category labels, which are customized according to the specific external behavioral patterns and internal psychological characteristics of each emotion. Through the application of this double-layer label system (i.e., the combination of emotion category labels and behavioral sub-category labels), it is possible to decode and classify the speech related to conflicts and disputes on social media in a more refined manner.

[0030] 2. With the help of the complex context understanding and generation capabilities of the large language model, the present invention can deeply analyze the abnormal psychological signs hidden in multi-source data such as social media, forums, and online comments. Through delicate emotion capture and semantic analysis, it can identify the psychological tendencies related to specific conflicts and disputes, and provide data support for early intervention. As a pre-trained large language model, Qwen can accurately map complex human psychological expressions with its learning experience on large-scale text data, and can accurately infer emotional states even in the face of ambiguous or obscure texts. This enables the present invention to transcend the emotion classification of a single dimension and achieve a comprehensive assessment of the individual's psychological state.

[0031] 3. Data collection focuses on major domestic social media platforms, aiming to capture subtle but critical emotional signals directly from the communication context of social media. This strategy not only ensures the localization and real-world relevance of the data, but also improves the accuracy and sensitivity of the model in identifying conflicts and potential conflict points. By deeply mining the massive amount of information on these platforms, the present invention can more accurately map the emotional context and provide a scientific basis and early warning signals for preventing and alleviating various conflicts. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The present invention is an overall flow chart. DETAILED DESCRIPTION

[0033] In view of the current problems of difficulty in analyzing abnormal psychological characteristics caused by conflicts and disputes on social platforms, lack of data acquisition and model benchmarks, the present invention uses a pre-trained large language model Qwen to capture the complex semantic information between abnormal psychological related texts under the social media conflict and dispute system, and fine-tunes the large language model through complex prompts.

[0034] In summary, the present invention includes four steps: (1) Based on relevant psychology research and speeches on social media, select seven types of emotional category labels for conflict and dispute texts on social media, and construct behavioral subcategory labels respectively. (2) Collect a dataset of conflict and dispute texts on social media, and annotate and divide the dataset. (3) Design Prompt to combine the emotion classification task and the behavior recognition task into one task. (4) Select the base model and parameters, and use LLaMA-Factory to train and fine-tune the large language model.

[0035] Below, the four steps are introduced in detail.

[0036] Step 1: Selection of emotions related to conflicts and disputes on social media and construction of labels.

[0037] The relevant speeches on conflicts and disputes on various social media platforms are organized into text form to obtain conflict and dispute texts. Based on relevant psychology research, the present invention collects, evaluates and classifies conflict and dispute texts on various domestic social media platforms, selects depression, anger, fatigue, fear, inferiority, anxiety, and impulse as the emotion category labels for social media conflict and dispute texts, and then determines the behavior subcategory labels of each emotion category label according to the behavioral performance and psychological characteristics corresponding to each emotion, as shown in Table 1. The emotion category labels and behavior subcategory labels are used in combination to identify and classify social media conflict and dispute texts. The present invention uses various domestic social media platforms as data sources for collection and analysis, which can better reflect various potential conflicts and disputes, making the subsequent model output more accurate.

[0038] Table 1 Selection of major categories of emotions and behavioral subcategories related to conflicts and disputes on social media

[0039] Emotional categories Behavior Subcategories anxiety Restlessness, nervousness, excessive worry Depression Depression, loss of pleasure, lack of will anger Unexplained anger, verbal abuse, and reckless behavior fear Fear, trembling, obvious avoidance self-abasement Self-blame, shame, low self-esteem tired Distraction, mood swings, lack of energy impulse Reacting quickly without planning, behaving out of control, and not considering the consequences

[0040] Step 2: Collection and classification of abnormal psychological data under the contradiction and dispute system.

[0041] Collection process: The present invention innovatively relies on the extensive domestic social media ecosystem as the source of data collection, aiming to timely identify individuals with potential abnormal psychological characteristics through detailed analysis of users' online speech. In view of the wide coverage of social media discussions, the present invention specially designs a data screening strategy for specific conflict and dispute topics, focusing on collecting speeches that are directly related to various conflicts and disputes or indirectly imply the user's psychological state. To this end, the present invention independently developed an intelligent crawler tool, which can accurately locate pages containing preset conflict and dispute tags or corresponding discussion sections, and efficiently capture relevant user comments and posts in compliance with relevant laws and regulations to ensure the timeliness and pertinence of data collection.

[0042] Although the collected raw data is large in quantity and wide in scope, it is disorganized and needs to undergo a strict screening and purification process. The first step is to preliminarily classify the data, eliminate irrelevant or redundant information, and only retain the content that directly reflects the user's psychological state in a specific conflict situation, such as conflict and dispute texts expressing anxiety, depression, impulse or anger. Subsequently, according to the data structure defined by the research requirements (such as user identifier, specific content, source link, etc.), the screened data is formatted to construct an unlabeled conflict and dispute text data set dedicated to abnormal psychological feature analysis. This process not only ensures the purity and relevance of the data, but also lays a solid foundation for subsequent deep text analysis and psychological feature mining, and is expected to reveal the deep connection between abnormal psychology and conflict and dispute. More specifically, the present invention crawls multiple emotional category label data such as anxiety, anger, etc. based on some conflict and dispute keywords. Subsequently, some high-quality data are manually screened out for use, and the seven emotional category label data are annotated with behavioral subcategory labels by psychology experts as sample data. Subsequently, the original conflict and dispute text is rewritten using the powerful generation ability of GPT4. Given the original sentence, the model can generate synonymous sentences, near-meaning sentences, or sentences with similar structures but different expressions. Multiple enhanced data are obtained through context insertion, sentence combination, noise injection, and sentiment transfer, plus another 100 pieces of data irrelevant to emotions (these 100 pieces of data can be used for negative sample training of the model to help the model learn to distinguish different emotions).

[0043] Division process: According to common practice, the present invention uses 80% of the data as a training set for model parameter learning and optimization; the remaining 20% ​​of the data is used as a test set to evaluate the performance of the model on unseen data. When allocating data, pay attention to maintaining a balance between the training set and the test set in terms of sentiment categories, text length, or other important dimensions to avoid bias and ensure that the evaluation results of the model are more representative. In this way, combined with GPT's text generation capabilities for data enhancement, not only can the amount of training data be effectively expanded, but also the performance and generalization ability of the model in tasks such as sentiment analysis and text classification can be significantly improved.

[0044] Step 3: Design Prompt to combine the emotion classification task and behavior recognition task into one task.

[0045] The present invention aims to model and detect multi-dimensional psychological and emotional characteristics. There are two main tasks in the detection stage. One is to classify the seven emotions of social media conflict and dispute texts, and the other is to detect the specific behaviors and psychological characteristics of social media conflict and dispute texts. In order to streamline the training process, optimize the training logic, and ensure the training effect, the present invention merges the two tasks into one task and designs Prompt.

[0046] In training large language models, the term "prompt" is often used to refer to a piece of input text used to guide model generation or learning. A prompt can be a question, a sentence, a description, or a complete chapter, depending on the specific task and goal. When using prompts for model training, it can be used in the following aspects: 1. Generation tasks. For generation tasks, prompts can provide an initial partial sentence or question to guide the model to generate relevant answers, summaries, translations, and other content. The design of prompts can affect the style, content, and accuracy of model generation. 2. Question-answering tasks. In question-answering tasks, prompts are usually questions, and the model needs to understand the question based on the prompts and generate corresponding answers. Prompts can provide background, context, or other prompt information for the question, which helps the model better understand and answer the question. 3. Dialogue tasks. In dialogue tasks, prompts can be the context or initial dialogue of a dialogue, which is used for dialogue generation or response of the model. The setting of prompts can affect the fluency, coherence, and content of the dialogue. 4. Text classification tasks. In text classification tasks, prompts can be a description or sentence to guide the model to classify text. Prompts can provide key information, category examples, or illustrations to help the model classify more accurately. When using prompts, you need to design appropriate prompts based on the needs of the specific task and model. Well-designed prompts can provide clear guidance and context, helping the model better understand the task and generate accurate output. At the same time, the choice of prompts can also affect the behavior and performance of the model, so experiments and adjustments are needed to obtain the best results.

[0047] Specifically, step three includes the following steps: the first step is to preset roles and contextual background. In Prompt, a psychology professor and data processing expert are preset for the large language model, and it is clear that the task of the large language model is to classify the emotions of social media conflict and dispute texts according to given labels, so as to guide the model to generate more accurate, coherent and relevant texts. There are several advantages to doing this. First, the generation scope of the large language model is limited: by presetting roles and contexts, the output of the large language model can be limited to the subject scenarios and task requirements of the emotion classification of speech related to conflicts and disputes, avoiding the generation of irrelevant content and improving the quality and readability of the output content. Second, improve logical coherence: through role definition and background setting, the large language model can better understand the task scenario and content requirements, making the generated text logic more coherent, the content more reasonable, and the context more fitting. Third, increase the efficiency of model generation: by limiting the large language model to specific roles, it can understand the context faster, reduce the search space required for the model to generate text, and improve the efficiency of model generation. The example of Prompt in this embodiment is shown in Table 2.

[0048] Table 2

[0049]

[0050]

[0051]

[0052] The second step is to give the set label data in json format and specify the input and output format. The present invention writes all the set emotion category labels and behavior subcategory labels into Prompt in a fixed json format and passes them to the model, specifies the input social media conflict and dispute text as a specific text content format, and outputs it in a json format containing emotion category labels and behavior subcategory labels. Structuring data in json format can make data organization and access clearer and more convenient, and can also represent more complex data relationships, while improving text readability. By limiting the format and content of model input and output, the error rate of model training can be reduced and the relevance of model output content can be improved.

[0053] The third step is to determine the reasoning process of the large language model. The present invention combines emotion classification and behavior recognition into one task and designs Prompt. In Prompt, the large language model is required to first classify the social media conflict text into seven categories of emotions in a sequential order, and then detect which corresponding behavior subcategory labels are contained in the speech according to the recognized emotions. Completing two classification tasks with one Prompt avoids tedious work such as intent distribution or switching between multiple models, simplifies the model training process, and improves model recognition efficiency.

[0054] Step 4: Select the base model and parameters, and train and fine-tune the large language model. Write the labeled data in batches to Prompt, select the base model, and train and fine-tune it.

[0055] Step 4 specifically includes the following steps: Step 1, select the base model Qwen1.5-base. Tongyi Qianwen 1.5 (qwen1.5) is an open source large language model in the Tongyi Qianwen series developed by Alibaba Cloud. The series includes multi-version and multi-scale open source models such as Base and Chat to meet different computing needs. There are several advantages to choosing qwen1.5: First, the multi-language capability is improved. Qwen1.5 has been significantly optimized in multi-language processing capabilities, supporting a wider range of language types and more complex language scenarios. Second, human preference alignment. By adopting technologies such as direct policy optimization (DPO) and proximal policy optimization (PPO), the alignment of the model with human preferences is enhanced. Third, long sequence support. All scales of qwen1.5 models support context lengths of up to 32,768 tokens, which greatly improves the ability to process long texts. In version Qwen1.5, six sizes of Base and Chat models including 0.5B, 1.8B, 4B, 7B, 14B and 72B are open source, and a quantitative model is also provided. Choosing the qwen model will be easier to use.

[0056] The second step is to use LLaMA-Factory to train and fine-tune the large language model. LLaMA-Factory is an easy-to-use large language model fine-tuning framework. This framework simplifies the process of model fine-tuning and is easy to access. It provides a very complete development interface, including model fine-tuning methods, parameter configuration, dataset settings, model saving, model merging, model testing, and model trials, etc. It can efficiently and cost-effectively support fine-tuning and secondary training of more than 100 models.

[0057] The implementation of this invention will involve several key steps: first, integrate existing psychology literature and clinical knowledge to design a psychological abnormality indicator system that is applicable to a wide range of conflict and dispute situations; second, take advantage of Qwen's model architecture to learn the subtle patterns of abnormal psychological expressions through massive text data training; at the same time, take strict data desensitization and privacy protection measures to ensure that the analysis process is legal and compliant; finally, through continuous iteration and verification, establish a prediction model that is both sensitive and robust.

[0058] In summary, the present invention aims to fill the gap in the current field of psychological anomaly detection in preventing conflicts and disputes, and to use the advanced analysis capabilities of large language models such as Qwen to open up new paths for early warning and effective intervention.

[0059] Example

[0060] ① Selection of emotions related to conflicts and disputes in social media and construction of labels. The present invention examined a number of mainstream domestic open social media platforms, mainly including Weibo super topics, Zhihu questions and answers, Douban mutual aid groups and Xiaohongshu and other related content. At the same time, it also communicated with psychological experts to judge and select social media platforms and judge and classify social media conflict and dispute texts. After research, it was decided to use seven emotions, namely depression, anger, fatigue, fear, inferiority, anxiety, and impulse, as the major emotion category labels for model training, and then judge each emotion separately, select 2 to 3 representative behavioral manifestations and psychological characteristics, and further classify them to obtain behavioral subcategory labels.

[0061] ② Manual data annotation and data set construction: After investigation, it was found that the three platforms of Zhihu, Douban and Xiaohongshu are not suitable for comprehensive data capture due to their structural limitations and content nature. Finally, we chose to crawl topics and user comments in super chats on Weibo with keywords related to various conflicts and disputes, so as to reflect the emotional status of users on social media platforms as realistically as possible. The present invention uses a custom web crawler to crawl nearly 30,000 comments from Weibo, including the expression of users' real emotions, knowledge sharing of emotional psychology, subjective guidance of emotional bloggers, and objective description of emotional phenomena. In view of the fact that the research focus is on emotions related to conflicts and disputes, the present invention focuses on retaining data on users' real emotional expressions. After manual screening, 100 pieces of data for each type of emotion are retained, totaling 700 pieces of data, and data cleaning and format conversion are performed for further processing. Subsequently, psychology experts were contacted to select 70 pieces of data from 700 pieces as sample data and manually annotated. The annotated data is fed to GPT4 for data enhancement. Given the original sentence, the original text is rewritten to allow the model to generate synonymous sentences, near-meaning sentences, or sentences with similar structures but different expressions. 700 pieces of enhanced data were obtained through context insertion, sentence combination, noise injection, and emotion transfer. Finally, 100 pieces of data with no emotional content were selected from the original data (for negative sample training of the model to help the model learn to distinguish different emotions), totaling 800 pieces. Finally, the training set and test set were divided into 8:2 and trained together.

[0062] ③ Prompt design: First, set the roles of psychologists and data processing experts for the large language model and the context for psychological and emotional analysis. Give label samples in json format and specify the input and output formats. Integrate the manually labeled data in ② into Prompt to obtain training samples.

[0063] ④ Model training: Use the LLaMA-Factory framework to select appropriate parameter training and fine-tune various models such as Baichuan, Qwen, and GPT, and then compare the training and test results. Finally, select Qwen-1.5base base model 1.8B, select 9.00E-05 for parameter learning rate, select 9 for Epoch, and keep other parameters as default. After training and fine-tuning, a model with better effect is obtained. The test results are F1=0.08375 for the emotion category label, F1=0.7109 for the behavior subcategory label, and 2.08s for inference time.

[0064] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0065] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. An abnormal psychology detection method based on multi-dimensional psychological feature modeling, characterized in that: It is used to classify emotions and identify behaviors of conflicting texts, including the following steps: Step 1: Based on the conflict and dispute texts on social media, select the major emotion categories and the behavioral subcategories under each major emotion category, and set the major emotion category labels and the behavioral subcategory labels; Step 2: Collect a dataset of conflict and dispute texts on social media, annotate and divide the dataset into training and test sets; Step 3: Design prompt words to combine the emotion classification task and behavior recognition classification task of conflict and dispute text into one task; Step 4: Select the large language model and the parameter scale of the large language model, and use LLaMA-Factory to train and fine-tune the large language model based on the prompt words and training samples in the training set; the training samples include the conflict and dispute texts used for training, as well as the corresponding emotion category labels and behavior subcategory labels; Step 5: Input the prompt words and the conflict texts in the test set into the trained large language model to obtain abnormal psychological detection results, including the emotion category labels and behavior subcategories of the conflict texts.

2. The abnormal psychology detection method based on multi-dimensional psychological feature modeling according to claim 1 is characterized in that: Step 1 specifically includes: The broad categories of emotions included anxiety, depression, anger, fear, low self-esteem, fatigue, and impulsivity; Behavioral subcategories of anxiety include: fidgeting, nervousness, and excessive worrying; Behavioral subcategories of depression include: emotional suppression, loss of pleasure, and lack of aspiration; Angry behavioral subcategories included: getting angry for no reason, being verbally abusive, and behaving recklessly; Behavioral subcategories of fear include: fearful trembling and overt avoidance; Behavioral subcategories of low self-esteem include: self-blame, shame, and low self-esteem; Behavioral subcategories of fatigue included: distractibility, emotional lability, and lack of energy; Impulsive behavior subcategories include: rapid unplanned reactions, lack of control over behavior, and lack of consideration for consequences.

3. The abnormal psychology detection method based on multi-dimensional psychological feature modeling according to claim 1 is characterized in that: Step three specifically includes the following steps: S31, preset roles and context: In the prompt words, the identities of psychology professors and data processing experts are set for the large language model, and it is made clear that the task of the large language model is to classify emotions and identify behaviors of conflicting texts in social media according to the given large emotion category labels and behavior subcategory labels, so as to guide the model to generate more accurate, coherent and relevant texts; S32, write all set emotion category labels and behavior subcategory labels into prompt words in a fixed json format and transmit them to the large language model, stipulate that the conflict and dispute text input to the large language model is in text format, and the output of the large language model is in json format including emotion category labels and behavior subcategory labels; S33, in the prompt word, requires the large language model to first classify the emotions of the conflict text, and then detect the behavioral subcategories contained in the conflict text according to the identified emotion categories.

Images