Feeds Don't Tell the Whole Story: Measuring Online-Offline Emotion Alignment

Abstract

In contemporary society, social media is deeply integrated into daily life, yet emotional expression often differs between real and online contexts. We studied the Persian community on X to explore this gap, designing a human-centered pipeline to measure alignment between real-world and social media emotions. Recent tweets and images of participants were collected and analyzed using Transformers-based text and image sentiment modules. Friends of participants provided insights into their real-world emotions, which were compared with online expressions using a distance criterion. The study involved N=105 participants, 393 friends, over 8,300 tweets, and 2,000 media images. Results showed only 28% similarity between images and real-world emotions, while tweets aligned about 76% with participants' real-life feelings. Statistical analyses confirmed significant disparities in sentiment proportions across images, tweets, and friends' perceptions, highlighting differences in emotional expression between online and offline environments and demonstrating practical utility of the proposed pipeline for understanding digital self-presentation.

Figures (4)

• Figure 1: Overview of the proposed experimental pipeline. The workflow includes participant recruitment, collection of tweets and images, AI-based sentiment and facial expression analysis, aggregation of friends’ offline assessments, computation of sentiment distribution distances, and visualization of alignment results.
• Figure 2: Overview of the experimental website used in the post-survey study: participants first log in with their X platform ID and a secret key to access their social media data (Step 1), and then view the analysis results (Step 2), where donut charts show the distribution of emotional tones (e.g., happy, sad, neutral, angry) and bar charts present similarity scores (0–100) comparing real-world personality with expressions across tweets, uploaded images, and friends’ content.
• Figure 3: A screenshot of 10 tweets from the gathered dataset is shown here.
• Figure 4: Architecture of the final hybrid sentiment classification model. The model takes a Persian input sentence, generates two sets of contextual embeddings using ParsBERT and LaBSE models, and concatenates the generated embeddings into a single vector. The final embedding vector is passed through a series of fully connected layers and non-linear activation functions (e.g., ReLU and Leaky ReLU) to produce a probability distribution over five sentiment classes. In parallel, a rule-based model based on keyword detection attempts to predict sentiment. The final output is selected based on a decision rule: if the neural model's confidence is $\geq 80$ or the rule-based system cannot make a prediction, the model's output is used; otherwise, the rule-based result is chosen.