Context in Sentiment and Emotion Analysis by Ben Bos A Thesis in the Field of Computer Science for the Degree of Master of Science in Computer Science of MASTER OF SCIENCE in Computer Science Approved: Dr. Robert Ball Advisor/Committee Chair Dr. Nicole Anderson Committee Member Dr. Sarah Herrmann Committee Member WEBER STATE UNIVERSITY 2025 Abstract This thesis explores the development, distribution, and analysis of a research project aimed at examining the emotional and cognitive responses of participants exposed to different stimuli in the form of sentences and accompanying images. The thesis involved a survey distributed to 120 participants, gathering demographic data and assessing responses to 35 sentence-based stimuli, each paired with images representing positive, neutral, or negative emotional states. The research process included a design phase, where I formulated hypotheses, developed the survey structure, and implemented the collection of both qualitative and quantitative data. Following distribution, I employed statistical methods, including the Chisquare test, to analyze participant responses, with particular attention to variations in emotional interpretation and valence based on the stimuli. The findings from this analysis contributed to a deeper understanding of the interaction between visual stimuli and emotional processing in individuals with varying backgrounds in education and field of study. The results underscore the significance of emotional context in shaping cognitive interpretations, revealing distinct patterns related to educational background and exposure to emotional imagery. The conclusions drawn from this research provide valuable insights into the psychological mechanisms underlying emotional perception and contribute to the broader discourse on emotion research. Table of Contents Introduction 4 Related Work 9 Continued Research and Project Ideas 13 Methods 15 Building the Website 17 Distributing the Survey 19 Analyzing the Results 20 Conclusion 28 Appendix 1 32 Appendix 2 66 References 69 Introduction Understanding emotions is crucial to the way we function as human beings. It helps us to communicate better. Having high communication skills can significantly reduce arguments and other conflicts that might occur from someone misinterpreting something that one has said or written. Two ways that we currently analyze text are sentiment analysis and emotion analysis. “Sentiment analysis … is the field of study that analyzes people's opinions, sentiments, evaluations, appraisals, attitudes, and emotions towards entities such as products, services, organizations, individuals, issues, events, topics, and their attributes.” [1] Emotion analysis is “the method of defining and evaluating the emotions conveyed in textual data.” [2] My research shows how context can affect a person’s reported emotion analysis of a given text. Context is “circumstances that form the setting for an event, statement, or idea, and in terms of which it can be fully understood and assessed” [3]. Why does context matter to sentiment and emotion analysis? There are certainly many instances where this might not matter (e.g. a description of a service in an invoice, doctors note on a diagnosis, etc.). However, in general day-to-day life this type of analysis is constantly occurring and is incredibly important. During casual conversation, whether verbal or written, it is important to be able to have context for what people are saying to you. How someone conveys his or her message can significantly change how someone else interprets what is being said. This non-direct communication is what helps give context to what someone says. When this is removed, significant amounts of context can be missed, and miscommunication can occur. When any form of communication happens through some sort of text, by nature a lot of context can be missed. An example of why context matters would be a simple text conversation. If two people are not careful with how they word their messages, it is easy to misinterpret the emotion or sentiment behind a message. This is an example on a smaller, more personal scale. How might this translate to a larger, more commercial context? When companies more easily and quickly analyze reviews on their products and services, they can better learn what is working well for them and what is not. Reading reviews can take a lot of time, especially for large companies. Understanding how context affects the emotion and sentiment tied to a given text can help improve interpretation of text and potentially add to the development of software for recognizing specific emotion and sentiment from text. I distributed a survey to 120 participants who reported an emotion and valence they interpreted to be associated with a given sentence. For example, John Doe read a sentence. After John Doe read the sentence, he reported what emotion he associated with the sentence. He repeated this process for all provided sentences. There were a total of 35 sentences that each participant read through. I measured context by providing participants with a specific picture for the point of priming the individual. These pictures had either positive or negative valence associated with them. Valence refers to the intrinsic positivity or negativity of an emotion or experience. In psychology, it describes whether a stimulus evokes a positive (e.g., happiness) or negative (e.g., sadness) emotional response. Sentence 1 My old cat has died and gone to the next life. Figure 1. Positive Cat Figure 1 depicts a cat that has died and gone to the next life. This image displays a positive event because of the bright colors and happy looking cat. Figure 2. Negative Cat Figure 2 depicts a cat that has died and gone to the next life. This image displays a negative event because of the dark colors and not being able to see the cat's face. There was one picture for each valence for each sentence with a total of 70 pictures. This served as an independent variable. For a control group there was also the possibility of no picture to be displayed throughout the survey. The intent of the pictures was to see if it gave the participants some sort of context to what they read. The participants were not given any information about the picture itself, it simply appeared along with the sentence. The participants read a sentence and reported the specific emotion and valence that they believed to be associated with the sentence. This reported emotion and valence was the dependent variable. For example, a participant might have opened the survey to find a picture of a spooky pond. The sentence to go along with it could say something like “The fox jumped over the log, dashed between the trees, and disappeared into the forest.” The participant would select from a dropdown menu an emotion they perceive to be associated with the sentence. The emotions they picked from were from a selection of the Universal Emotions (Enjoyment, Sadness) [4]. This was a between-subject experiment because each participant got a different experience for the survey that they completed. Each participant conducted an analysis of 35 sentences, with or without a picture. The pictures displayed were not the same between separate surveys. My research question was the following: If given a form of context with positive or negative valence while reading a given sentence, will the reader shift the way they emotionally analyze the words of the given sentence? My hypotheses were the following: 1. Given a context with positive valence along with a sentence, a study subject’s emotional interpretation will be influenced to be positive. 2. Given a context with negative valence along with a sentence, a study subject’s emotional interpretation will be influenced to be negative. 3. Given a sentence with no context, the emotional interpretation will be evenly split between positive and negative among all study subjects 4. Female participants will report more positive emotions to all sentences than male participants will. 5. Male participants will report more negative emotions to all sentences than female participants will. 6. Younger participants will report more positive emotions to all sentences than older participants will. I think that context to any sort of message is important to conveying a desired meaning, whether it be a written or oral message. This will show that even without directly trying to influence emotion one way or the other, having a context will influence to a degree how people interpret things, in this case its words in a sentence. Related Work It is common for people to assume that certain emotions are opposite. While this can often be true, emotions are not necessarily binary, and they can coexist within an individual. This is particularly important when analyzing the emotional content of text, as different individuals may interpret the same sentence in varying ways. Some words have widely agreedupon emotional associations, while others may evoke multiple emotions depending on the context. The lexicon used in text analysis significantly impacts these interpretations [5]. Merriam-Webster defines a lexicon as "the vocabulary of a language, an individual speaker or group of speakers, or a subject." Notably, a lexicon is not confined to a single language but instead to specific groups or situations, making it a context-aware lexicon [5]. This reinforces the importance of context in determining emotional and sentiment-based interpretations of text. As noted, "text can capture but only a portion of emotions expressed by a human" [6], further highlighting the limitations of textual analysis when broader contextual elements are not considered. Sentiment analysis typically determines whether a text conveys a positive, neutral, or negative tone, while emotion analysis assigns specific emotions, such as happiness or sadness [7]. The context in which a text appears directly influences the lexicon used, which in turn affects the emotions and sentiments derived from it. This underscores the fact that emotions are not solely derived from the words themselves but are shaped by the surrounding context. Additionally, the granularity of emotional classification plays a crucial role—individuals who classify emotions at a more detailed level may perceive different emotional tones in the same text. Tools like the Feelings Wheel, developed by Dr. Gloria Willcox, help break down primary emotions (such as happiness, sadness, or anger) into secondary and tertiary emotions, offering a more nuanced view [8]. This tool helps individuals identify their emotions with greater specificity, which is essential because "emotions are not binary" [5]. Just because someone does not feel happiness, it does not necessarily mean they are experiencing fear or anger—there is a vast spectrum of emotions in between. This idea of nuance in emotion perception extends beyond text alone. Visual Sentiment Analysis (VSA) explores how images evoke emotions and influence perception. Research in this field suggests that sentiment derived from images can shape how individuals interpret accompanying text. Just like in text-based sentiment analysis, VSA systems rely on emotional models, dataset definitions, and feature designs to determine how different components of an image contribute to sentiment perception [9]. Both VSA and textual sentiment analysis highlight the essential role that contextual influences—whether visual, cultural, or social—play in sentiment interpretation. Cultural and individual differences further complicate emotion perception. Research has shown that individuals from Asian cultures tend to rely more heavily on contextual cues when assessing emotions compared to those from Western cultures, though this effect diminishes with age. Younger individuals also tend to recall more negative emotional stimuli than older individuals, whose emotional recall is generally more balanced [10]. These findings suggest that emotion and sentiment analysis models must account for not only textual and visual context but also cultural and demographic factors that influence perception. By considering these differences, sentiment analysis can better capture the emotional tone of individuals from diverse backgrounds, providing insights that are more accurate. Furthermore, emotion perception is inherently subjective. Different individuals may associate different emotions with the same stimulus. While many sentiment analysis models focus on identifying the dominant emotion in a given text, research suggests that personalized emotion perceptions are shaped by multiple factors, including visual content, social context, temporal evolution, and location [11]. This highlights the complexity of emotion classification and suggests that a one-size-fits-all approach may be insufficient. Instead, models that account for personalized and contextual differences in emotion perception are likely to offer more meaningful and accurate results. Another advanced form of sentiment analysis is Aspect-Based Sentiment Analysis (ABSA), which categorizes text by specific aspects and identifies the sentiment attributed to each one. This approach allows for more granular sentiment analysis, particularly in contexts like customer feedback, where different aspects of a product or service may elicit different emotional reactions. ABSA enables businesses to make data-driven decisions by automating sentiment analysis processes, ensuring consistency across multiple texts, and providing real-time insights into customer satisfaction [12]. Finally, sentiment analysis is also applied in dialogue systems, where it is used to assess the emotional tone of conversations. This has practical implications for improving humancomputer interactions. For instance, if a system detects negative sentiment during a conversation, it can escalate the issue to a human operator for better resolution. Since dialogues consist of a series of interactions, context plays a crucial role in analyzing sentiment accurately. Studies show that context is particularly important in human-computer interactions and in short human-human conversations, where client sentiment must be closely monitored. Research using customer-support dialogues has demonstrated that models incorporating context—especially those based on Bidirectional Encoder Representations from Transformers (BERT)-based classifiers—achieve superior sentiment analysis performance [13]. This emphasizes the necessity of considering prior utterances and dialogue structure to effectively evaluate sentiment in conversational data. Continued Research and Project Ideas Once I had looked at a sufficient amount of related research for this project, I began formulating my research questions and hypotheses. Determining when I had gathered enough background information was not a straightforward process. I initially started by reviewing existing literature related to emotion recognition, survey methodologies, and the impact of visual stimuli on perception. This included analyzing prior studies that explored how individuals interpret emotional content, as well as looking at different methods researchers have used to assess these interpretations. While reviewing the literature, I identified gaps that my study could address, particularly in how emotional valence is influenced by contextual imagery. These gaps guided the development of my research questions and hypotheses. As I refined my focus, I considered different methodologies for data collection. I ultimately chose a survey format because it allowed for a controlled and systematic collection of participant responses while reaching a broader audience compared to in-person studies. However, the decision to create a web-based survey stemmed from both practical and technical considerations. Given my background in computer science, I saw an opportunity to integrate my technical skills into the research process. This approach also provided greater accessibility for participants, as they could complete the survey remotely rather than in a lab setting. Initially, I explored alternative options, such as using pre-existing survey platforms, but I found that they lacked the flexibility I desired to implement specific experimental conditions, such as controlling the sequence of images shown to participants. Developing the website posed several challenges, as I had limited prior experience in web programming beyond one or two courses in my academic curriculum. This required me to learn new frameworks and programming languages while also understanding how to configure and deploy servers to host the survey. I faced several technical hurdles, such as managing session storage to preserve participant responses across multiple pages and ensuring that the website functioned correctly across different devices and browsers. To overcome these challenges, I relied heavily on online documentation, tutorials, and trial-and-error debugging. Additionally, I sought guidance from my brother, who works in web development, and his insights were invaluable in helping me navigate unfamiliar technologies. Once I had a solid grasp of my research objectives and the technical requirements of the survey, I compiled my findings, outlined my initial design concepts, and began drafting my thesis proposal for committee review. The proposal stage was critical, as it required me to clearly articulate my research questions, justify my chosen methodology, and outline the expected contributions of my study. During the proposal phase, I also had to consider ethical considerations, such as ensuring informed consent from participants and protecting their data privacy. Throughout the entire process, I continued refining my research approach, revisiting literature, and adjusting my methodology based on new insights. Research, I found, is rarely a linear process—it involves constant iteration, reflection, and adaptation. Methods For my research, I wanted to distribute the survey to a small range of demographics. This included male and female for the reported sex of the participants, their age, and their educational field (which was grouped into either STEM or Non-STEM). I decided to use the educational field as a tested demographic ad hoc. I made this decision because my educational background is in STEM and I feel that I have noticed a stereotype where those in the STEM background tend to have very different emotional interpretations than those of other backgrounds. With this range of demographics, I decided to distribute the survey to 120 participants where 60 were male and 60 were female. Among the male and female demographics, each had 30 that had a STEM educational background, and 30 had a Non-STEM educational background. With this identified number of participants, I ensured that I got an equal amount of survey context types for each demographic combination. For example, among the 30 males, who had a STEM educational background, 10 received a survey with no context throughout the survey, 10 received a survey with a positive context throughout the survey, and 10 received a survey with a negative context throughout the survey. This same pattern of 10 surveys of each context type was repeated for each of the tested demographic combinations. Originally, I had wanted to distribute the survey with only 10 questions each. The purpose was to reduce the amount of participant fatigue. Due to feedback from my committee, I decided to expand the number of questions to 35 so that I would have more reliable and more robust results. We decided that this expansion would not be significant enough to fatigue a participant to the point where it affected their responses to the survey questions. In addition to this, they also suggested that I make sure the webpages have visual responsiveness so that the participant would be assured that their interactions with the webpage were working, as well as making sure that the implementation of the survey would be compatible with different devices because each user would be using devices of different types and sizes. Although I did create a few of the sentences for the survey myself, I generated most of the sentences with ChatGPT. This process was not as simple as telling the AI tool to give me 35 sentences. It mostly gave me sentences that would make sense to a reader, but I wanted to be sure that everything a participant read would comprehensible so I ask the tool several times for various sentences. Once I finally had my list of sentences picked out, I utilized a couple different free AI tools to generate images with both positive and negative contexts using the sentences as prompts. Building the Website This was by far the most difficult milestone. I started the process of building the website to host the survey by creating a Git repository to ensure I did not unintentionally lose any code. Initially, I wanted to host the website on AWS, but this proved to be more complex than I anticipated. Due to this newfound complexity, I decided to host the website on GitHub Pages and use Heroku for my database. To test the functionality of my database, I created some sample queries but quickly discovered that GitHub Pages did not support this capability. Fortunately, Heroku also supports server hosting, so I moved the backend server code there, which allowed me to connect seamlessly to the MySQL database. Once I had these technical details resolved, I began creating the code for the demographics page, which served as the first step in the user flow of the survey. Next, I focused on the survey page itself. My initial goal was to create a dynamic survey page that could generate the necessary sentences and pictures on the same page. However, due to the limitations of GitHub Pages, this approach was not feasible. Instead, I opted to create an individual page for each question in the survey. Each survey page was identical in structure, differing only in the sentences presented. This approach allowed me to reuse much of the code and streamline the development process by consolidating common functionality into a shared code file. This not only simplified the coding process but also ensured consistency across all survey pages. With the survey questions finalized, I shifted my attention to enhancing the user experience. I developed a landing page that featured the informed consent document, giving participants a clear understanding of the survey’s purpose and their rights. Additionally, I created a final page that marked the end of the survey and informed participants they could close the tab. Despite the challenges and setbacks, completing this milestone was immensely rewarding. It required me to overcome significant technical hurdles, adapt to unforeseen limitations, and refine my approach to web development. Through this process, I gained valuable skills in problem solving, coding, and project management, which were integral to the successful implementation of my thesis project. Some example images to show what the web pages looked like can be found in appendix 2. Distributing the Survey After the website was completed and IRB approval was received, I started to distribute the survey to willing participants. Getting the number of responses needed proved to be more challenging than I initially anticipated. I overestimated the responsiveness of people I know, assuming they would promptly take part in the survey. While every person I distributed the survey to agreed to participate, many took more than a week to respond. To increase participation, I enlisted the help of family and friends to reach more individuals and gather additional results. During the distribution phase, a critical issue with the survey design was brought to my attention. Because my code randomly selected the type of picture displayed for each survey, there was no guarantee that I would achieve the required balance of results across all picture types for the identified demographic groups. For instance, before I reached the minimum number of responses in some categories, I already had more than the minimum in others. This imbalance posed a problem for my data analysis and the validity of my research findings. To address this issue, I updated my code to include a query to the database that checked whether the minimum number of responses had been met for a specific demographic and picture combination. If the required number of results were already achieved in one category, the code would automatically select another picture type to display, ensuring a more balanced distribution of responses across all groups. This adjustment not only resolved the imbalance but also demonstrated the importance of adapting to unforeseen challenges in the research process. Ultimately, this milestone emphasized the need for flexibility, problem solving, and the ability to refine my approach to meet the requirements of my study. Analyzing the Results Upon completion of gathering my results, I utilized Python and the scipy.stats, pandas, and matplotlib.pyplot packages for analysis and data display in order to answer my hypotheses. For convenience of the reader, the following are my hypotheses from the introductory chapter: 1. Given a context with positive valence along with a sentence, a study subject’s emotional interpretation will be influenced to be positive. 2. Given a context with negative valence along with a sentence, a study subject’s emotional interpretation will be influenced to be negative. 3. Given a sentence with no context, the emotional interpretation will be evenly split between positive and negative among all study subjects 4. Female participants will report more positive emotions to all sentences than male participants will. 5. Male participants will report more negative emotions to all sentences than female participants will. 6. Younger participants will report more positive emotions to all sentences than older participants will. I began my analysis by comparing the counts of emotion responses for each picture emotion used in the survey. I applied a similar process to examine other comparisons, such as emotional responses between male and female participants, STEM and non-STEM participants, male STEM and female STEM participants, and more. I employed statistical tests such as the Chi-Square test, T-test, Fisher’s Exact Test, and ANOVA as appropriately to analyze the data more thoroughly. To test my first three hypotheses, testing for valence, I ran Chi-Square analysis to determine if there was a statistical significance between the valence of the context given in the survey, and the valence selected by the participant for each question in the survey. The relation between the picture emotion and valence selected was statistically significant, X2 (2, N = 120) = 55.53, p < .0001. This rejects the Null Hypothesis that there is not a difference between the groups. In other words, given text with a positive image context, then it will likely positively influence how they interpret the valence of the sentence they are reading, while if you give them a negative or no context, then they will tend to interpret the sentence more negatively. This can be useful to help marketing departments understand that they can manipulate consumers’ emotions given a positive or negative context with their advertisements. These results were produced from the survey questions where after a participant read the survey sentence, they then selected the valence they felt was best represented in the sentence. Figure 3. Count of each Valence selected for each survey context category. Figure 3 shows the count of each valence selected for each survey’s context category. The n represents a negative valence response, and the p represents a positive valence response. The categories represented in the different surveys were positive, negative, and none. To further analyze my first three hypotheses I tested for emotional interpretation. To do this I ran Chi-Square analysis to determine if there was a statistical significance between the selected emotions and the valence of the context for a given survey. The relation between the picture emotion and emotion selected was statistically significant, X2 (12, N = 120) = 87.87, p < .0001. This rejects the Null Hypothesis that there is not a relation between a context and an individual’s emotional interpretation. In other words, given text with a positive image context, then it will likely positively influence how they emotionally interpret the sentence they are reading, while if you give them a negative or no context, then they will tend to interpret the sentence more negatively. These results were produced from the survey questions where after a participant read the survey sentence, they then selected the emotion they felt was best represented in the sentence. Figure 4. Count of Each Emotion Selected. Figure 4 shows the count of each emotion selected for each survey’s context category. The categories represented in the different surveys were positive, negative, and none. To test my fourth and fifth hypotheses, testing for emotion, I used a Chi-Square analysis to determine if there was a statistical significance between the emotion selected and the participant’s reported sex. The relation between a participant's sex and emotion selected was statistically significant, X2 (6, N = 120) = 71.85, p < .0001. These findings support the notion that females tend to report different emotions than males. This knowledge has practical implications for advertisers seeking to tailor emotional appeals in their content. If an advertisement aims to evoke specific emotions in men, advertisers may need to employ more deliberate strategies to achieve this effect. Conversely, if the goal is to elicit specific emotions in women, additional effort may be required to produce the intended response. Further research is needed to be done to determine which emotions specifically relate to which sex. Figure 5. Percentage of Each Emotion Type by Reported Sex of Participant. Figure 5 shows the percentage of each emotion selected for each participant’s reported sex. The reported sex categories represented in the different surveys were male and female. To further test my fourth and fifth hypotheses, testing for valence, I used the Fisher’s Exact. Using Fisher’s Exact Test, it was determined that there was no statistical significance between male and female participants and the valence they selected (p = .1640). To test my last hypothesis, that younger participants would report more positively, I used the Chi-Square test to determine if there is statistical significance between a participant's age and their reported emotion. The relation between a participant’s age group (binned: 18-29, 30-39, 40-49, etc.) and emotion selected was significant, X2 (36, N = 120) = 200.6, p < .0001. These findings suggest that older individuals may experience different levels of emotion compared to younger participants. An alternative explanation is that younger individuals may be facing different life challenges than their older counterparts. Further research is needed to explore this relationship and determine whether a significant correlation exists. Figure 6. Emotion Percentage by Age. Figure 6 displays the percentage of each emotion selected during the survey separated by age. The age categories displayed are under 30 years old and 30 years old and above. To further analyze my last hypothesis I used a Chi-Square analysis to determine if there is a relationship between the age of a participant and their reported valence. The relation between a participant’s age group (binned: 18-29, 30-39, 40-49, etc.) and valence selected was significant, X2 (6, N = 120) = 26.56, p = .0002. This finding further supports the conclusion that older participants are more likely to interpret situations positively. Building on the previous analysis, additional research could help determine whether this trend reflects a mere correlation or if underlying factors contribute to younger individuals experiencing more negative emotions in contemporary society. Figure 7. Valence Percentage by Age. Figure 7 displays the percentage of each valence selected during the survey separated by age. The age categories displayed are under 30 years old and 30 years old and above. After formulating my initial hypotheses, I considered additional variables for analysis. Upon further reflection, I decided to examine whether differences exist in the interpretation of valence and emotion between individuals with a STEM education and those without. This inquiry was inspired by my own experiences as someone with a STEM background, as I have observed differences in how I perceive the world compared to those without similar educational training. I used a Chi-Square test to determine if there was a relationship between a participant’s educational field and the valence they selected. It was determined that the relation between a participant’s education field and valence selected was not significant, X2 (1, N = 120) = 3.2, p = .0735. To further analyze my last hypothesis, I used a Chi-Square test to determine if there was a relationship between a participant’s educational field and the emotions they selected. It was determined that the relation between a participant’s education field and emotion selected was significant, X2 (6, N = 120) = 42.52, p <.0001. The statistical significance of this comparison suggests that while individuals with different educational backgrounds may perceive the valence of a given context similarly, the specific emotions associated with it may vary. Most emotions were selected in similar proportions across groups, with the primary exception being the emotion of contempt. The reason for this discrepancy remains unclear, highlighting the need for further research to determine whether this pattern reflects an underlying cause or is merely a coincidence. Figure 8. Emotion Percentage by Educational Field. Figure 8 displays the percentage of each emotion selected during the survey separated by educational field. The following summarizes my statistical analysis: ● Given a context with positive valence, participants' emotional interpretations are influenced to be positive. ● Given a context with negative valence, participants' emotional interpretations are influenced to be negative. ● Given a sentence with no context, emotional interpretation is not evenly split between positive and negative and is influenced to be negative instead. ● Given a participants’ reported sex is female, that participant is influenced to report more positively labeled emotions. ● Given a participants’ reported sex is male, that participant is influenced to report more negatively labeled emotions. ● Younger participants’ will not report emotions that are more positive and will report emotions that are more negative instead. This milestone was both challenging and enlightening, as it required me to deepen my understanding of statistical analysis and adapt my approach to ensure meaningful and robust conclusions. The insights gained from this phase have been invaluable in shaping the interpretation and significance of my research findings. Conclusion The central research question of this study was: If given a form of context with positive or negative valence while reading a sentence, will the reader shift the way they emotionally analyze the words in that sentence? The results suggest that yes; emotional interpretation is influenced by context. Evaluation of Hypotheses 1. Given a context with positive valence, participants' emotional interpretations will be influenced to be positive. Supported 2. Given a context with negative valence, participants' emotional interpretations will be influenced to be negative. Supported These findings indicate that the provided contextual images influenced participants' emotional responses. The most noticeable effect was seen in the selection of "happy" emotions, though differences were also observed across other emotions. 3. Given a sentence with no context, emotional interpretation will be evenly split between positive and negative. Not Supported This hypothesis was incorrect. While I initially assumed a uniform distribution, the results instead showed that responses without context tended to be more negative rather than being evenly split. 4. Female participants will report more positively to all sentences than male participants. Inconclusive 5. Male participants will report more negatively to all sentences than female participants. Inconclusive These hypotheses were based on stereotypes regarding gender and emotional expression. The data seemed to support these assumptions for selected emotions, but not for selected valence. I initially predicted that men, due to stereotypical notions of aggression, would respond more negatively, and women more positively. Further research is needed to explore potential factors influencing gender-based emotional reporting. 6. Younger participants will report more positively to all sentences than older participants. Not Supported This hypothesis was also based on a stereotype—that older individuals tend to be more negative or "grouchy." However, the results did not support this assumption. It is possible that societal shifts in stress levels across age groups have influenced emotional perception, with younger individuals experiencing more stress and uncertainty than previous generations. While this is speculative, it presents an interesting area for future research. The statistical analyses showed that participants' emotional responses varied significantly based on the contextual valence provided, confirming that context plays a role in shaping emotional interpretation. However, there was not always a significant shift in how participants within differing demographics assigned valence to sentences. One potential explanation is the absence of a neutral option. Multiple participants noted that a neutral valence option could have provided a more appropriate representation of the sentence they read. Additionally, some participants may not have fully understood the term "valence," which could have led to random or inconsistent responses in this category. If I had taught participants about valence before they took the survey, the results may have been different. Another limitation revealed in the study was that there was limited sample diversity. I had male and female, STEM and Non-STEM, older and younger participants, but they were all from the same geographical area. I think that due to the amount of participants I had, it worked well enough, but to get better and more accurate results it would be good to increase the amount of participants and expand to other geographical areas. This would more than likely provide more diverse perspectives and interpretations of the various survey sentences. Final Thoughts This study provides evidence that context influences emotional interpretation. The findings challenge several preconceived notions about gender, age, and emotion processing. While the statistical analyses yielded meaningful insights, they also highlighted limitations. Future research should address these limitations to refine our understanding of how external context shapes emotional perception. One idea, as mentioned before, would be to include a neutral valence option. It may not be something that is always identified, but on occasion individuals may not feel any particular way about a given stimuli. Another option for future work would be to test for more demographics. This can give a greater insight to how different groups of people think. To be able to do that, you will also need to increase the sample size. The sample size used in this study may have already been too small to give the best results possible, so any further division on demographics will increase the need for a greater sample size. One final idea for future work would be to apply this research to real time systems such as AI marketing review tools or chat bots. An AI marketing review tool could be developed to better analyze customer reviews on products and services to tell them what is working and what is not. Additionally, marketing teams could use one of these tools to review their advertisements preemptively to tell them how various demographics might react to it. Another real time system application would be to chat bots. With the growing use of emoji’s in our written conversations, this is getting to be less of a problem. However, there are still times where even when accompanied by emoji’s it is still hard to understand what message someone is attempting to convey. Additionally, those who have autism may have more struggles with this understanding of various conveyed messages. Some sort of chat assistant could be developed to help people better understand what someone means by what they have written. Appendix 1 Survey Sentences and Accompanying Pictures Sentence 2 Time seemed to stand in that moment. Figure 9. Positive Time. Figure 9 shows a happy moment between a couple on the beach. They embrace each other and are filled with happiness, as time seems to stand still in that moment. Figure 10. Negative Time. Figure 10 shows a room with very little sunlight coming in. In the middle of the room is a chair with a vase of flowers on it, all having dull colors. Sentence 3 Silence enveloped the room. Figure 11. Positive Silence. Figure 11 displays an empty bedroom with a neatly made bed, lots of sunlight, and teddy bears scattered around. The light gives a warm feeling to the room and the neatly made bed gives a comforting feeling. Figure 12. Negative Silence. Figure 12 shows an empty room with some light coming in through the window. A bed mostly in the shadows of the room looks completely un-made and filthy. The ceiling and walls are cracking. The floor has some paper and other garbage scattered around. Sentence 4 The old oak tree stood tall, its branches reaching towards the sky, casting long shadows across the water. Figure 13. Positive Oak Tree. Figure 13 shows a tall tree. The tree’s leaves are green and vibrant with life. The tree is next to a large body of water. Figure 14. Negative Oak Tree. Figure 14 shows a tall tree that appears to be dead. The surrounding forest is foggy and appears to be filled with other dead trees. Sentence 5 Raindrops tapped against the windowpane. Figure 15. Positive Rain. Figure 15 shows a window that has some rain trickling down it. Just outside the window are some brightly colored flowers. Figure 16. Negative Rain. Figure 16 shows an old dirty window frame. It is raining outside and it seems that it is leaking through the window because the ground inside of the window is wet and reflecting the window frame on it. Sentence 6 Footsteps echoed in the alleyway. Figure 17. Positive Alleyway. Figure 17 shows a brightly lit alleyway. There are several brightly colored plants along the sides of the walkways. Happy people are seen walking by. Figure 18 Negative Alleyway. Figure 18 shows a dimly lit alleyway. A single person’s silhouette is seen walking through the alley. Sentence 7 The wind whispered through the leaves. Figure 19 Positive Leaves. Figure 19 shows bright sunlight leaks through the trees. All the plants and trees are vibrant with life. A single dirt path cuts straight through the middle of the trees and other plants. Figure 20. Negative Leaves. Figure 20 shows a foggy forest filled with dead trees. Dried and dead leaves are scattered across the ground. A single path with tree roots crowding it cuts straight through the middle of the trees. Sentence 8 Shadows lengthened as the sun dipped below the horizon, transforming the world with the approach of evening. Figure 21. Positive Horizon. Figure 21 shows a peaceful shallow pond reflecting the surrounding landscape off it. Brightly colored trees are scattered along the shoreline. A mountain range is seen in the distant background and the sunset light dips behind it. Figure 22. Negative Horizon. Figure 22 shows a peaceful but unsettling landscape stretching out towards the horizon. The sun is slowly setting behind the horizon. Very little plant life is seen. Sentence 9 The moon cast a light over the landscape. Figure 23. Positive Moon. Figure 23 shows a peaceful nighttime landscape with some houses scattered around. A bright full moon is seen in the background giving light to the landscape. A small stream is seen along the side of the house at the front of the image. Figure 24 Negative Moon. Figure 24 shows an unsettling nighttime landscape. Very little plant life is seen. A single tree is seen alongside the path cutting through the middle of the image. A full moon is seen in the sky giving an eerie light to the landscape. Sentence 10 Time seemed to slip through my fingers. Figure 25. Positive Time Slip. Figure 25 shows a small family sitting around a table. On the table is a birthday cake with several candles on it. The family is seen enjoying the moment together. Figure 26. Negative Time Slip. Figure 26 shows a single person walking along a sandy landscape. The sand seems to stretch for miles. The person’s shadow seems to be stretching out from them as the sun is setting. Sentence 11 The sound of thunder rumbled. Figure 27. Positive Thunder. Figure 27 shows a puppy trotting through a wet garden. A rainbow is seen in the background in front of dark thundering storm clouds. Figure 28. Negative Thunder. Figure 28 shows a dark but grassy plain stretching out to the horizon. Dark stormy clouds are seen overhead. A single but large and bright lightning strike cuts down the middle of the image going straight into the ground. Sentence 12 The world seemed to be holding. Figure 29. Positive Holding. Figure 29 shows a small family holding hands walking along the beach. Some boats are seen sailing across the water towards the horizon as the sun is setting. Birds are flying overhead. Figure 30. Negative Holding. Figure 30 shows a single person’s silhouette at the front of the image next to a dead tree. The barren landscape stretches out to the horizon. A pillar of smoke is seen in the background going up into the sky. Sentence 13 The scent of rain hung in the air. Figure 31. Positive Scent of Rain. Figure 31 shows a house that is surrounded by brightly colored flowers. The bright sunlight covers the area. A stone pathway leading to the house has water trickling down it. The roof of the house is wet. Figure 32. Negative Scent of Rain. Figure 32 shows a woman walking down the street holding an umbrella over her head. The street is covered with water as rain pours down. The sky is covered with gray storm clouds. Sentence 14 The silence was broken only by my heartbeat. Figure 33. Positive Silence. Figure 33 shows a woman meditating next to a stream. The stream calmly flows through the middle of a forest. The sunlight trickles through the trees. Figure 34. Negative Silence. Figure 34 shows a man meditating on the shoreline of a swampy forest. The trees throughout the forest are all dead. An eerie fog creeps through the forest. Sentence 15 A car rumbled down the dusty highway. Figure 35. Positive Car Rumble. Figure 35 shows a convertible car driving down the road into the sunset. The rolling hills surrounding the road are covered in yellow and green plant life. A mountain is seen in the distance. Figure 36. Negative Car Rumble. Figure 36 shows an old rusty car in the middle of a muddy old road. Dead trees line the road all the way into the distance and fade into the fog. Sentence 16 A person walked on the sidewalk. Figure 37. Positive Walking Person. Figure 37 shows two people walking down the sidewalk of a brightly lit neighborhood. The houses are all close together with beautiful flower gardens in front of them. Some trees line the walkway. Figure 38. Negative Walking Person. Figure 38 shows a person walking alone down the sidewalk. Tall buildings line the sidewalk. A small crowd of people is seen walking in the distance. Sentence 17 Leaves rustled in the wind. Figure 39. Positive Leaves Rustling. Figure 39 shows a brightly lit park covered with trees. The leaves are red and yellow. Some park benches and tables are seen as deer are walking around. Figure 40. Negative Leaves Rustling. Figure 40 shows a single dull yellow leaf that sits right at the front of the image. Behind the leaf are many other dull yellow and orange leaves scattered across the ground. In the distance is a forest of dark and dead trees. Sentence 18 Someone wrote on a piece of paper. Figure 41. Positive Writing. Figure 41 shows a woman sitting at a desk in a home office. The desk sits in front of a window full of sunlight. The woman is writing on a piece of paper. Figure 42. Negative Writing. Figure 42 shows a woman sitting at a desk in a dimly lit room. The only light coming from a small candle on the desk and from the window. The woman is writing on a piece of paper. Sentence 19 A clock ticked on the wall. Figure 43. Positive Clock Tick. Figure 43 shows a big grandfather clock on the side of a room. The room is brightly lit by the sunlight coming through the window. A few chairs and tables are scattered around the room. Figure 44. Negative Clock Tick. Figure 44 shows a small clock sitting on the wall of an old room. The wallpaper seems to be falling apart. Pieces of the wallpaper are scattered around the floor. The light coming through the window dimly lights the room. Sentence 20 A phone rang in the room. Figure 45. Positive Phone Ring. Figure 45 shows an old rotary phone sitting on a desk on the side of a home office. Houseplants are scattered around the room. The room is brightly lit by the sunlight coming through the window. Figure 46. Negative Phone Ring. Figure 46 shows an old rotary phone sitting on a table in the middle of a dark room. The little bit of light that is in the room comes through an old window. Sentence 21 A train passed by on the tracks. Figure 47. Positive Train. Figure 47 shows a steam train driving along some train tracks. The tracks run through a valley full of houses. The houses and the train are full of bright colors. Figure 48. Negative Train. Figure 48 shows a steam train driving along some train tracks. The surrounding landscape is barren. Some telephone poles are seen lining the side of the tracks. Sentence 22 A key turned in a lock. Figure 49. Positive Key Turn. Figure 49 shows a brightly colored front door to a house. Some brightly colored plants are seen on the steps leading up to the door. Figure 50. Negative Key Turn. Figure 50 shows a small amount of light coming through the door into a small room. Other than the light coming through the door, the room is dark. Sentence 23 A baby cried in the next room. Figure 51. Positive Baby Cry. Figure 51 shows a small nursery with two cribs. Bright light entering through the window fills the room. Some stuffed animals are seen lying around on the floor and in the cribs. Figure 52. Negative Baby Cry. Figure 52 shows a sad baby sitting in the middle of the room. The room is dark and filthy. The room has a little bit of light coming in through the window. Sentence 24 A car horn honked outside. Figure 53. Positive Car Honk. Figure 53 shows a woman sitting on a couch in a living room. The room is lit by the evening sun coming through the window and a few lamps around the room. Figure 54. Negative Car Honk. Figure 54 shows a narrow street with tall buildings. Several cars are parked on the side of the road. There is some water on the side of the road and fog in the distance. Sentence 25 The child ran through the park. Figure 55. Positive Child Park. Figure 55 shows a lively park. Several kids are seen running around with colorful balloons. The midday sun brightly lights the park. Figure 56. Negative Child Park. Figure 56 shows a single child running down a sidewalk through a park. Some leafless trees are lining the sidewalk. A small crowd of people is seen in the distance. Sentence 26 The man stood at the edge of the cliff and looked down. Figure 57. Positive Cliff. Figure 57 shows a small beach with a tall cliff surrounding it. A man is seen standing on top of the cliff looking towards the setting sun. On the opposite side of the cliff are some houses. Figure 58. Negative Cliff. Figure 58 shows a rocky shore. A person is seen standing on the edge of a cliff watching as giant waves crash into the shore. Sentence 27 The woman picked up the phone and dialed a number. Figure 59. Positive Phone. Figure 59 shows a woman sitting on a couch talking on the phone. The room she is in is brightly lit by the sunlight coming in through the window. The woman has a smile on her face. Figure 60. Negative Phone. Figure 60 shows a woman on the phone under a small lamp. The woman has a sad look on her face. The room she is in is dark. Sentence 28 The dog barked loudly in the yard. Figure 61. Positive Dog Bark. Figure 61 shows a dog chasing a tennis ball through a yard. The midday sun brightly lights the yard. There are some flowers on the sides of the yard. Figure 62. Negative Dog Bark. Figure 62 shows a silhouette of a dog in the middle of a yard. The only light comes from the full moon in the background. There are trees in the yard without any leaves on them. Sentence 29 The door slammed shut behind him. Figure 63. Positive Door Slam. Figure 63 shows a man walking through a door. There are brightly colored plants all around. The midday sun lights up the surrounding area. Figure 64. Negative Door Slam. Figure 64 shows a man walking through a door. The room he is entering is much brighter than the one he was in. The man's shoulders are slumped down indicating he is in a negative mood. Sentence 30 The flowers swayed in the breeze. Figure 65. Positive Flowers. Figure 65 shows a park bench in the middle of a flower garden. The flowers are all full of bright colors. It is a nice and sunny day. Figure 66. Negative Flowers. Figure 66 shows an old and broken up house in the middle of a flower garden. Lots of fog and leafless trees are seen in the surrounding area. The flowers in the garden are spread out and have dull colors. Sentence 31 The athlete triumphantly crossed the finish line. Figure 67. Positive Athlete. Figure 67 shows a woman crossing the finish line after a race. She raises her arms triumphantly as she crosses the line. There are a few other racers behind her and a full crowd in the stands cheering her on. Figure 68. Negative Athlete. Figure 68 shows a man crossing the finish line. He seems to be in a bad mood. It is a dark and overcast day. Sentence 32 The couple held hands as they walked along the beach. Figure 69. Positive Beach Walk. Figure 69 shows a couple holding hands walking across a beach. Waves are slowly crashing into the beach during the sunset. Birds are flying around overhead. Figure 70. Negative Beach Walk. Figure 70 shows a couple holding hands walking across a beach. It is a dark and overcast day. The waves are crashing into the shore. Sentence 33 The fox jumped over the log, dashed between the trees, and disappeared into the forest. Figure 71. Positive Fox. Figure 71 shows a happy looking fox jumping over a log. The fox and log are in the middle of a brightly lit forest. Flowers and butterflies can be seen scattered around. Figure 72. Negative Fox. Figure 72 shows a fox standing on a log in the middle of a foggy forest. The trees do not have any leaves on them. There are dead leaves scattered across the ground. Sentence 34 The sound of footsteps echoed in the hallway. Figure 73. Positive Hallway. Figure 73 shows a brightly lit hallway that has picture frames neatly arranged along the wall. There are brightly colored walls and doors in the hallway. Figure 74. Negative Hallway. Figure 74 shows a long and narrow hallway that has lights spread out along the ceiling. Dark colored doors line the walls. Sentence 35 By the tranquil lake, beneath the setting sun. Figure 75. Positive Tranquil Lake. Figure 75 shows a small group of people having a picnic beneath a tree. Next to them is a lake reflecting the setting sun on it. Trees full of green leaves are seen around the lake. Figure 76. Negative Tranquil Lake. Figure 76 shows a person standing on the edge of a lake. Through the dark and overcast sky, the setting sunlight peeks through the clouds. The lake water is calm. Appendix 2 Examples of the various pages shown during a survey. Informed Consent Form Figure 77. Informed Consent Form. Figure 77 shows what the informed consent form web page looks like to a survey participant. Demographics Form Figure 78. Demographics Form Figure 78 shows what the demographics form looks like to every user who participated in the survey. Sample Page from a Survey Figure 79. Sample Page from a Survey. Figure 79 shows a sample of what a survey question looks like to a user. This specific example uses Figure 1 to show what a survey that displays images with a positive context would look like. Surveys with negative or no context look the same, just different or no images. References [1] R. Abinaya, P. Aishwaryaa, S. Baavana, and N. D. T. Selvi, “Automatic sentiment analysis of user reviews,” 2016 IEEE Technological Innovations in ICT for Agriculture and Rural Development (TIAR), Jul. 2016. doi:10.1109/tiar.2016.7801231 [2]S. De, S. Dey, S. Bhattacharyya, and S. Bhatia, Eds., Advanced Data Mining Tools and Methods for Social Computing, 1st ed. 2022. [3] N. S. Bruce and W. J. Davies, “The effects of expectation on the perception of soundscapes,” Applied Acoustics, vol. 85, pp. 1–11, Nov. 2014. doi:10.1016/j.apacoust.2014.03.016 [4] P. Ekman, “Universal emotions,” Paul Ekman Group, https://www.paulekman.com/universalemotions/ (accessed Feb. 21, 2024). [5] Ball, Robert, Joshua Jensen, and Samuel Romine. "Crowdsourcing a More Realistic Emotional Lexicon Process." International Conference on Human-Computer Interaction. Cham: Springer Nature Switzerland, 2023 [6] V. Tripathi, A. Joshi, and P. Bhattacharyya, “Emotion Analysis from Text: A Survey.” Accessed: Apr. 19, 2021. [Online]. Available: https://www.cfilt.iitb.ac.in/resources/surveys/emotion-analysis-survey-2016-vaibhav.pdf [7] S. Romine, J. Jensen, and R. Ball, “Comparing Sentiment Analysis and Emotion Analysis of Algorithms vs People,” Weber State University. [8] “The Feelings Wheel: unlock the power of your emotions,” Calm Blog, Aug. 29, 2023. https://www.calm.com/blog/the-feelings-wheel [9] A. Ortis, G. M. Farinella, and S. Battiato, “Survey on visual sentiment analysis,” IET Image Processing, vol. 14, no. 8, pp. 1440–1456, May 2020. doi:10.1049/iet-ipr.2019.1270 [10] S.-G. Ko, T.-H. Lee, H.-Y. Yoon, J.-H. Kwon, and M. Mather, “How does context affect assessments of facial emotion? The role of culture and age.,” Psychology and Aging, vol. 26(1), 2011. [11] S. Zhao, H. Yao, Y. Gao, G. Ding, and T.-S. Chua, “Predicting personalized image emotion perceptions in social networks,” IEEE Transactions on Affective Computing, vol. 9, no. 4, pp. 526–540, Oct. 2018. doi:10.1109/taffc.2016.2628787 [12] H. H. Do, P. Prasad, A. Maag, and A. Alsadoon, “Deep learning for aspect-based sentiment analysis: A comparative review,” Expert Systems with Applications, vol. 118, pp. 272–299, Mar. 2019. doi:10.1016/j.eswa.2018.10.003 [13] I. Carvalho, H. G. Oliveira, and C. Silva, “The importance of context for sentiment analysis in dialogues,” IEEE Access, vol. 11, pp. 86088–86103, 2023. doi:10.1109/access.2023.3304633 [14] D. Carnine, E. J. Kameenui, and G. Coyle, “Utilization of contextual information in determining the meaning of unfamiliar words,” Reading Research Quarterly, vol. 19, no. 2, p. 188, 1984. doi:10.2307/747362 [15] N. M. Hakak, M. Mohd, M. Kirmani, and M. Mohd, “Emotion analysis: A survey,” 2017 International Conference on Computer, Communications and Electronics (Comptelix), pp. 397– 402, Jul. 2017. doi:10.1109/comptelix.2017.8004002 [16] P. Nandwani and R. Verma, “A review on sentiment analysis and emotion detection from text,” Social Network Analysis and Mining, vol. 11, no. 1, Aug. 2021. doi:10.1007/s13278-02100776-6 [17] M. Wankhade, A. C. Rao, and C. Kulkarni, “A survey on sentiment analysis methods, applications, and challenges,” Artificial Intelligence Review, vol. 55, no. 7, pp. 5731–5780, Feb. 2022. doi:10.1007/s10462-022-10144-1 [18] D. M. Hussein, “A survey on sentiment analysis challenges,” Journal of King Saud University - Engineering Sciences, vol. 30, no. 4, pp. 330–338, Oct. 2018. doi:10.1016/j.jksues.2016.04.002 [19] S. Zhao et al., “Exploring principles-of-art features for Image emotion recognition,” Proceedings of the 22nd ACM international conference on Multimedia, Nov. 2014. doi:10.1145/2647868.2654930 [20] S. Zhao et al., “Computational emotion analysis from images: Recent advances and future directions,” Human Perception of Visual Information, pp. 85–113, Jul. 2021. doi:10.1007/978-3030-81465-6_4 [21] U. Doshi, V. Barot, and S. Gavhane, “Emotion detection and sentiment analysis of static images,” 2020 International Conference on Convergence to Digital World - Quo Vadis (ICCDW), Feb. 2020. doi:10.1109/iccdw45521.2020.9318713 [22] N. Mittal, D. Sharma, and M. L. Joshi, “Image sentiment analysis using Deep Learning,” 2018 IEEE/WIC/ACM International Conference on Web Intelligence (WI), pp. 684–687, Dec. 2018. doi:10.1109/wi.2018.00-11 [23] T. Zhu et al., “Multimodal sentiment analysis with image-text Interaction Network,” IEEE Transactions on Multimedia, vol. 25, pp. 3375–3385, 2023. doi:10.1109/tmm.2022.3160060 [24] S. Poria et al., “Context-dependent sentiment analysis in user-generated videos,” Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2017. doi:10.18653/v1/p17-1081 [25] B. Agarwal, N. Mittal, P. Bansal, and S. Garg, “Sentiment analysis using common-sense and context information,” Computational Intelligence and Neuroscience, vol. 2015, pp. 1–9, 2015. doi:10.1155/2015/715730 [26] L. F. Barrett, B. Mesquita, and M. Gendron, “Context in emotion perception,” Current Directions in Psychological Science, vol. 20, no. 5, pp. 286–290, Oct. 2011. doi:10.1177/0963721411422522 [27] J. Lee, S. Kim, S. Kim, J. Park, and K. Sohn, in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 10143–10152 [28] R. Kosti, J. M. Alvarez, A. Recasens, and A. Lapedriza, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1667–1675 [29] N. Ngo and D. M. Isaacowitz, “Use of context in emotion perception: The role of top-down control, cue type, and perceiver’s age.,” Emotion, vol. 15, no. 3, pp. 292–302, Jun. 2015. doi:10.1037/emo0000062 [30] L. F. Barrett and E. A. Kensinger, “Context is routinely encoded during emotion perception,” Psychological Science, vol. 21, no. 4, pp. 595–599, Feb. 2010. doi:10.1177/0956797610363547 [31] K. H. Greenaway, E. K. Kalokerinos, and L. A. Williams, “Context is everything (in Emotion Research),” Social and Personality Psychology Compass, vol. 12, no. 6, May 2018. doi:10.1111/spc3.12393 [32] W. Frankenstein, K. Joseph, and K. M. Carley, “Contextual sentiment analysis,” Lecture Notes in Computer Science, pp. 291–300, 2016. doi:10.1007/978-3-319-39931-7_28 [33] J. A. Deonna, “Emotion, perception and perspective,” dialectica, vol. 60, no. 1, pp. 29–46, Mar. 2006. doi:10.1111/j.1746-8361.2005.01031.x [34] M. Calbi et al., “How context influences our perception of emotional faces: A behavioral study on the Kuleshov effect,” Frontiers in Psychology, vol. 8, Oct. 2017. doi:10.3389/fpsyg.2017.01684 [35] V. Campos, B. Jou, and X. Giró-i-Nieto, “From pixels to sentiment: Fine-tuning CNNS for visual sentiment prediction,” Image and Vision Computing, vol. 65, pp. 15–22, Sep. 2017. doi:10.1016/j.imavis.2017.01.011 [36] S. Fan et al., “The role of visual attention in sentiment prediction,” Proceedings of the 25th ACM international conference on Multimedia, pp. 217–225, Oct. 2017. doi:10.1145/3123266.3123445 [37] H. Ou, C. Qing, X. Xu, and J. Jin, “Multi-level context Pyramid Network for visual sentiment analysis,” Sensors, vol. 21, no. 6, p. 2136, Mar. 2021. doi:10.3390/s21062136 [38] A. Ortis, G. Farinella, and S. Battiato, “An overview on image sentiment analysis: Methods, datasets and current challenges,” Proceedings of the 16th International Joint Conference on eBusiness and Telecommunications, 2019. doi:10.5220/0007909602900300 [39] G. Ahuja, A. Alaei, and U. Pal, “A new multimodal sentiment analysis for images containing textual information,” Multimedia Tools and Applications, Aug. 2024. doi:10.1007/s11042-024-19999-8 [40] Q. Chen, E. Ragusa, I. Chaturvedi, E. Cambria, and R. Zunino, “Text-image sentiment analysis,” Lecture Notes in Computer Science, pp. 169–180, 2023. doi:10.1007/978-3-03123804-8_14