Abstract:The growing ubiquity of Social Media data offers an attractive perspective for improving the quality of machine learning-based models in several fields, ranging from Computer Vision to Natural Language Processing. In this paper we focus on Facebook posts paired with reactions of multiple users, and we investigate their relationships with classes of emotions that are typically considered in the task of emotion detection. We are inspired by the idea of introducing a connection between reactions and emotions by means of First-Order Logic formulas, and we propose an end-to-end neural model that is able to jointly learn to detect emotions and predict Facebook reactions in a multi-task environment, where the logic formulas are converted into polynomial constraints. Our model is trained using a large collection of unsupervised texts together with data labeled with emotion classes and Facebook posts that include reactions. An extended experimental analysis that leverages a large collection of Facebook posts shows that the tasks of emotion classification and reaction prediction can both benefit from their interaction.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to improve the accuracy of sentiment detection and prediction of Facebook users' reactions through joint learning models. Specifically, the author focuses on how to utilize the relationship between social network data (especially Facebook posts and their users' "reactions") and sentiment categories to improve the performance of machine - learning models in these two tasks. ### Decomposition of the Main Problem 1. **Sentiment Detection**: - Identify the user's emotional state from the text. - Use the six basic emotions defined by Ekman: anger, disgust, fear, happiness, sadness, and surprise. 2. **Prediction of Facebook Reactions**: - Predict users' reactions to Facebook posts, including LOVE, HAHA, WOW, SAD, ANGRY, and LIKE. - Note that LIKE is a general and ambiguous positive feedback, while other reactions are more specifically associated with certain emotional categories. 3. **Joint Learning**: - Propose a neural network model that can simultaneously learn sentiment detection and prediction of Facebook reactions. - Link these two tasks by introducing First - Order Logic (FOL) formulas as constraints and train in a multi - task environment. ### Key Points of the Solution - **Model Architecture**: - Use a Bidirectional Long - Short - Term Memory (Bidirectional LSTM) network to encode the input text. - Two predictors are used for reaction classification and sentiment classification respectively, and they share the same latent representation. - **Logical Constraints**: - Convert FOL formulas into polynomial constraints and incorporate them into the learning process in the form of soft constraints. - These logical rules describe the relationships between emotions and reactions, for example: \[ \forall x \, HAHA(x) \Rightarrow happiness(x) \] \[ \forall x \, SAD(x) \Rightarrow sadness(x) \] \[ \forall x \, ANGRY(x) \Rightarrow anger(x) \lor disgust(x) \] - **Dataset**: - Use partially labeled and unlabeled data for training, including: - Facebook posts and their reaction labels. - Sentences labeled with sentiment categories. - A large number of unlabeled Facebook posts. Through this method, the author aims to explore the interaction between sentiment detection and reaction prediction, thereby improving the performance of both tasks. The experimental results show that the introduction of logical constraints can significantly improve the performance of the model, especially in the sentiment classification task.

Jointly Learning to Detect Emotions and Predict Facebook Reactions

Emotion Detection in Online Social Networks: A Multilabel Learning Approach

Social Emotion Mining Techniques for Facebook Posts Reaction Prediction

Emotional Reaction Analysis Based on Multi-Label Graph Convolutional Networks and Dynamic Facial Expression Recognition Transformer

Emotion Detection In Online Social Network Based On Multi-Label Learning

Recent Advancement of Emotion Cognition in Large Language Models

Language Models as Emotional Classifiers for Textual Conversation

A First Step in Combining Cognitive Event Features and Natural Language Representations to Predict Emotions

Modeling Label Semantics for Predicting Emotional Reactions

Face Behavior a la carte: Expressions, Affect and Action Units in a Single Network

Emotion Detection and Analysis on Social Media

Facial Emotion Recognition with Noisy Multi-task Annotations

Classifying Emotions and Engagement in Online Learning Based on a Single Facial Expression Recognition Neural Network

MODELLING A NOVEL APPROACH FOR EMOTION RECOGNITION USING LEARNING AND NATURAL LANGUAGE PROCESSING

Association Extraction and Recognition of Multiple Emotion Expressed in Social Texts

Emotion Recognition under Consideration of the Emotion Component Process Model

Joint Contrastive Learning with Feature Alignment for Cross-Corpus EEG-based Emotion Recognition

Joint Emoji Classification And Embedding Learning

Affective Behaviour Analysis via Progressive Learning

Robust Facial Reactions Generation: An Emotion-Aware Framework with Modality Compensation

Multi-task, multi-label and multi-domain learning with residual convolutional networks for emotion recognition