1. Introduction

The profound integration of artificial intelligence and public discourse is reshaping the forms of dissemination and mobilization mechanisms of social movements in the digital space. The core of this transformation lies in the breakthrough in emotion computing technology: through facial expression recognition, voice emotion analysis, and text emotion mapping, AI systems can capture and quantify the emotional fluctuations of the crowd in real time. This type of technology was first applied in the fields of corporate customer service and entertainment, and is now rapidly penetrating into political communication and social resistance scenarios. Its core value lies in optimizing communication strategies through real-time feedback of emotional data to achieve accurate information delivery and amplify emotional resonance.

This technological shift is particularly notable in the process of forming the "emotional community." Social media platforms are not only barometers of public sentiment, but also shape the direction of collective emotions through algorithmic recommendation mechanisms. Protest organizers are using emotion recognition technology to decode the online sentiment map and design more provocative communication content accordingly. Communication plans designed around targeted emotions such as anger and fear can increase the effectiveness of online mobilization by more than 60%, but technological empowerment also leads to profound ethical paradoxes [1].

When emotions become measurable and controllable data variables, questions of authenticity and the right to know emerge. Communication strategies that rely too heavily on emotional optimization can lead to performative protests, while algorithmic black boxes threaten the democratic foundation of public discussions. This study, by building a multimodal emotion recognition model and combining simulated communication experiments with ethical framework analysis, reveals how emotional intelligence tools are reshaping the logic of digital resistance mobilization. The experimental data show that the multimodal model integrating facial expressions and textual emotions achieves an F1 value of 0.89 in the emotion classification task, which is 21% higher than the single-modal scheme. The research also highlights that the unbridled application of affective computing can exacerbate the tendency towards digital authoritarianism, and calls for the establishment of a governance framework that takes into account both technological utility and social responsibility.

2. Literature review

2.1. AI in emotion recognition

Figure 1: Architecture of a convolutional neural network for facial emotion recognition (source: techxplore.com)

Emotion recognition technology has evolved from basic emotion annotation to systems capable of analyzing subtle multimodal emotions. Current mainstream solutions include facial expression recognition based on convolutional neural networks. As shown in Figure 1, the system feeds facial images into the feature extraction layer and captures key features such as eyebrow movements and eye shapes layer by layer through standardized processing and nonlinear activation functions. After several layers of abstraction, the output end of the network presents the classification results of emotions such as joy and anger based on probability distribution.

Text sentiment analysis relies on the transformer architecture to capture semantic changes and sentiment fluctuations in online texts. Technological developments show that multimodal systems integrating voice, image, and text are becoming the dominant solution for emotion analysis. This type of intelligent system forms the technical basis for public opinion monitoring. Its core value lies in optimizing communication strategies through emotion map analysis, such as identifying group anger index in real time during public events to provide decision support for crisis public relations [2]. Experimental data show that the emotion classification accuracy of the multimodal system is 21% higher than that of the single-modal system, demonstrating unique advantages in protest discourse analysis.

2.2. Affective publics and strategic communication

The concept of “emotional community” reveals a new form of group aggregation in the digital age: dynamic online communities formed based on shared emotional experiences. This type of group is not a fixed organization but a temporary community that dynamically coalesces around hot events. The bond that sustains it is emotional energy flowing in real time. Like and share mechanisms and algorithmic recommendation systems on social platforms act as emotional amplifiers, allowing emotions such as anger and hope to spread quickly among groups.

In this context, communication strategies focus on accurately capturing the pulse of traditional emotions and designing visual symbols and narrative frames that can evoke empathy. The emotional appeal of protest labels, individualized narrative strategies, and symbolic visual elements have become the primary tools for building temporary alliances [3]. The communication effect no longer relies solely on rational argumentation but rather on the power of emotional communication. Experimental data shows that the volume of retweets for tweets containing angry content is 3.2 times higher than that for neutral content. This emotionally motivated communication strategy has become the underlying logic of modern social movements. Not only is it reshaping the form of public discourse, but it is also profoundly influencing individual participation in decision-making, marking a paradigm shift in the model of political participation.

2.3. Algorithmic governance and emotional labor

Emotion recognition technology is evolving from a passive monitoring tool to an active intervention mechanism. When these systems are integrated into the platform's algorithm, they can analyze users' emotional fluctuations in real time, allowing for dynamic adjustments to the communication strategy. This technological intervention has given rise to a new model of behavioral regulation: content visibility and user interaction depend on whether emotional expression conforms to algorithmic preferences.

In a context of resistance, this has given rise to the phenomenon of "algorithmic emotional labor." Participants actively or passively adjust their methods of emotional expression to conform to the algorithm's expectations in order to maintain communication effectiveness. For example, the use of specific angry emojis in protest tweets can increase the post's engagement volume by 2.3 times. This type of pressure creates an emotional echo effect, gradually transforming online emotional expression into a performance behavior dominated by the platform's rules [4]. Algorithmic governance not only reconstructs how emotions are presented but also changes their social value attributes. While technological empowerment can maintain the popularity of sports, there is a risk of instrumentalizing human emotions—rich emotional experiences are compressed into quantifiable interaction indicators. In the context of social movements, the tension between authentic emotions and algorithmic performances triggers profound ethical dilemmas. For example, a case from an environmental protection movement shows that a communication strategy that relies too heavily on algorithmic optimization led to the dissipation of substantive demands, evolving into a symbolic carnival fueled by traffic. This forces us to reexamine the limits and boundaries of technological intervention in the public domain [5].

3. Experimental methods and process

3.1. Data sources and sampling

This study adopts a stepwise experimental design to explore how emotional intelligence tools influence digital resistance communication strategies. The data collection covers three aspects: 1) obtaining images of high-profile protest activities in North America, Southeast Asia, and other places from public video platforms to provide real-scene training materials for the expression recognition model; 2) capturing social media tweets on issues such as environmental protection and racial equality, and filtering out text data containing strong emotional expressions; 3) integrating multimodal benchmark datasets (including facial expressions, tone of voice, and text sentiment annotations) to pre-train and validate the model's performance [6]. The experiment pays special attention to high-frequency emotions in six types of protest scenes such as anger and fear. By building a multimodal fusion model (F1 = 0.89), the emotional resonance law between facial micro-expressions and slogan texts in protest images is analyzed. At the data preprocessing stage, key frames were extracted from 480 hours of video material, and social media texts were de-identified to ensure compliance with ethical standards.

3.2. Preprocessing and annotation

The research data undergoes a systematic preprocessing process: key frames of video material are intercepted at 5-second intervals for grayscale processing and unified lighting conditions. Social media text data should eliminate non-standard characters such as emojis and hyperlinks. Emotions were manually annotated by professional annotators, and the annotation consistency coefficient exceeded 0.85. Disputed samples were eliminated after a double-blind review. In the preprocessing stage, the focus is on the treatment of backlit images and blurred images in protest videos, and the de-identification processing of social media texts is completed [7]. The last constructed dataset covers multiple scenarios such as environmental protest images dominated by anger (accounting for 37%) and tweets on labor rights with a fearful tone (29%), ensuring the generalization ability of the model training.

3.3. AI tools and model configuration

Facial emotion recognition adopts the open-source tools OpenFace and DeepFace, and these two models perform well in complex scenarios. For text sentiment analysis, a dedicated model is built based on the RoBERTa architecture and optimized through a specific sentiment corpus. The multimodal fusion scheme builds a unified prediction model by combining facial expression features with text sentiment vectors. The technical implementation is based on the TensorFlow and PyTorch frameworks and is deployed in the NVIDIA GPU server cluster. In the parameter tuning phase, the grid search method is adopted, with the F1 values of each emotion category as the optimization objective, and the optimal model configuration is finally determined. The experiment revealed that the system integrating facial micro-expressions and emotional signals from text improved the accuracy rate by 19% in the emotion classification task in protest scenes [8].

4. Results and discussion

4.1. Model performance and emotion detection accuracy

Experimental data show that the facial emotion recognition model performs well in classifying basic emotions (see Table 1), with an average F1 value of 0.81. Among them, joy had the highest recognition accuracy (F1 = 0.87), followed by sadness (F1 = 0.85). However, the model has limitations in discriminating complex emotions such as contempt and confusion (with F1 values ​​of 0.62 and 0.59, respectively), mainly due to environmental disturbances such as mask occlusion and rapid movement in protest images.

The text sentiment model is remarkable in detecting dimensions such as irony and complex emotions (F1 = 0.84), and the Transformer-based architecture is particularly good at analyzing metaphorical expressions in political slogans. The multimodal scheme integrating facial and textual features has the best overall efficiency (F1 = 0.89), which confirms that the cross-modal synergy effect can effectively solve the problem of context ambiguity [9]. In labor rights cases, the multimodal model reduced the misjudgment rate of anger emotions from 18% in the single-modal model to 7%, significantly improving the accuracy of emotion analysis.

Table 1: F1 scores of emotion recognition models by modality and emotion type

4.2. Emotional resonance and audience mobilization

The communication plan designed based on the emotion recognition system (see Table 2) shows consistent advantages in terms of audience interaction volume and engagement effectiveness. Retweets of traditional emotion-driven content such as anger and fear increased by 56% compared to the baseline, while tweets conveying hope and unity are more likely to trigger in-depth interactions such as comments, reflections, and financial donations.

Feedback from user interviews shows that content with a high degree of emotional adaptation can strengthen identity and make participants feel authentic and emotionally resonant. However, long-term reliance on communication strategies that arouse highly negative emotions such as anger and despair will lead to emotional fatigue and decreased participation [10]. In one environmental protection movement, the interaction volume of related tweets decreased by 42% in four weeks. Research has confirmed that effective digital protest communication must strike a balance between emotional intensity and narrative authenticity, avoiding the alienation of technological optimization towards emotional exploitation.

Table 2: Platform engagement metrics by emotion type (relative to baseline)

5. Conclusion

This study reveals the dynamic correlation between emotion recognition technology and the spread of digital manifestations. By integrating multimodal analyses such as facial expressions and textual emotions, the positive role of algorithmic emotion analysis in improving communication effectiveness was verified. Experimental data shows that multimodal models have significant advantages in identifying subtle differences in emotions. Emotional anchors such as anger and hope increase content transmission volume by 56% and call-to-action participation by 34%. However, the flip side of technological empowerment is the risk of emotional marketing—algorithmic filtering mechanisms can erode the authenticity of public discussions, and the misuse of users' emotional data has triggered conflicts over the right to know. A case from the labor movement shows that an anger dissemination strategy that relies excessively on algorithmic optimization saw a 42% decrease in interaction over four weeks, while the volume of participant interrogations increased by 27%. To achieve this, the application of technology must establish a transparent governance framework and integrate users' emotional autonomy into system design considerations. Future research must expand cross-cultural datasets and explore the development of community-based participatory AI. Only by balancing technological efficiency and digital ethics can the positive value of artificial intelligence be realized in the process of social transformation.