# Finding Good Representations of Emotions for Text Classification

**Authors:** Ji Ho Park  
**arXiv:** [1808.07235](https://arxiv.org/abs/1808.07235)  
**Date:** 2018-08-22  
**Summary page:** https://www.summarizepaper.com/en/paper/1808.07235v1/

## Key points

- Study focuses on the importance of accurately interpreting human emotions for effective human-machine communication
- Emotion plays a crucial role in human-to-human interactions
- Representing emotions in text poses a challenge in natural language processing (NLP)
- Continuous vector representations like word2vec fail to consider emotions and may exhibit bias towards certain identities
- Proposed approach includes emotional word vectors (EVEC) learned from a convolutional neural network model using an emotion-labeled corpus constructed with hashtags
- Sentence-level representations are learned by leveraging a large corpus of texts through the pseudo task of recognizing emojis
- Improved representations enhance sentiment/emotion analysis tasks significantly
- Considering emotions within text helps build more robust machine learning models for affect-related text classification tasks like sentiment/emotion analysis and abusive language detection
- Study explores model bias in existing approaches, specifically addressing gender bias in various neural network models through experiments aimed at measuring and reducing biases in representations
- Research contributes to building more inclusive and robust classification models
- Emphasizes the need for accurate representation of emotions in NLP tasks to facilitate better human-machine communication
- Proposed methods offer promising avenues for improving affect-related text classification tasks while addressing biases present in current approaches.

## Summary

The study focuses on the importance of accurately interpreting human emotions for effective human-machine communication. Emotion plays a crucial role in human-to-human interactions, and representing emotions in text poses a challenge in natural language processing (NLP). While continuous vector representations like word2vec have become popular in NLP, they fail to consider emotions and may unintentionally exhibit bias towards certain identities. To address this issue, the thesis aims to improve existing representations at both the word and sentence levels by explicitly incorporating emotions and model bias into the training process. The proposed approach includes emotional word vectors (EVEC), which are learned from a convolutional neural network model using an emotion-labeled corpus constructed with hashtags. Additionally, sentence-level representations are learned by leveraging a large corpus of texts through the pseudo task of recognizing emojis. The results demonstrate that these improved representations, trained on millions of tweets with weakly supervised labels such as hashtags and emojis, enhance sentiment/emotion analysis tasks significantly. By considering emotions within text, more robust machine learning models can be built for affect-related text classification tasks like sentiment/emotion analysis and abusive language detection. Furthermore, the study explores model bias in existing approaches by examining automatic detection of abusive language. Specifically, it addresses gender bias in various neural network models through experiments aimed at measuring and reducing biases in representations. This research contributes to building more inclusive and robust classification models. Overall, this thesis emphasizes the need for accurate representation of emotions in NLP tasks to facilitate better human-machine communication. The proposed methods offer promising avenues for improving affect-related text classification tasks while addressing biases present in current approaches.