[Paper Review] Annotation and Detection of Emotion in Text-based Dialogue Systems with CNN

Jialiang Zhao, Qi Gao|arXiv (Cornell University)|Oct 3, 2017

Sentiment Analysis and Opinion Mining8 references3 citations

TL;DR

This paper proposes EmoNet, a deep convolutional neural network (CNN)-based model for emotion detection in Chinese text-based dialogues, bypassing traditional preprocessing like segmentation and keyword extraction to preserve linguistic order and context. By leveraging a deep CNN architecture with residual connections and end-to-end learning, EmoNet achieves a top-1 accuracy of 72.8% on a 12,000-dialogue dataset, outperforming state-of-the-art methods in multi-class emotion detection.

ABSTRACT

Knowledge of users' emotion states helps improve human-computer interaction. In this work, we presented EmoNet, an emotion detector of Chinese daily dialogues based on deep convolutional neural networks. In order to maintain the original linguistic features, such as the order, commonly used methods like segmentation and keywords extraction were not adopted, instead we increased the depth of CNN and tried to let CNN learn inner linguistic relationships. Our main contribution is that we presented a new model and a new pipeline which can be used in multi-language environment to solve sentimental problems. Experimental results shows EmoNet has a great capacity in learning the emotion of dialogues and achieves a better result than other state of art detectors do.

Motivation & Objective

• To develop a robust emotion detection system for Chinese dialogues that maintains linguistic structure and context.
• To overcome limitations of traditional preprocessing (e.g., segmentation, keyword extraction) that distort word order and lose emotional nuance.
• To evaluate the effectiveness of deep CNNs in end-to-end emotion classification without manual feature engineering.
• To compare EmoNet’s performance against existing state-of-the-art models in Chinese text emotion detection.

Proposed method

• EmoNet employs a deep 9-layer CNN with 32, 64, 128, and 256-dimensional convolutional filters, followed by max-pooling and fully connected layers.
• The model processes raw UTF-8 encoded Chinese characters directly, avoiding segmentation and stemming to retain word order and syntactic context.
• Stop words are removed from input sequences to reduce noise and improve training efficiency, while preserving sequence order.
• A re-sampling technique is used to standardize input length, replacing traditional vectorization via LUT-based word frequency encoding.
• Hyperparameters such as learning rate (γ = 5×10⁻⁶) and L2 regularization (L = 1.5×10⁻⁴) are optimized via grid search for best performance.
• The network uses ReLU activations, dropout for regularization, and softmax for multi-class output (positive, negative, wondering, neutral).

Experimental results

Research questions

• RQ1 Can a deep CNN model detect emotions in Chinese dialogues without segmentation or keyword extraction, preserving linguistic order?
• RQ2 How does EmoNet’s performance compare to existing state-of-the-art emotion detectors in Chinese text-based dialogue systems?
• RQ3 What impact does the absence of preprocessing steps have on emotion detection accuracy and model generalization?
• RQ4 Why is the accuracy for positive emotions lower than for other classes, and how does this relate to irony and ambiguous expressions in training data?

Key findings

• EmoNet achieved a top-1 accuracy of 72.8% on a 12,000-dialogue test set after 100 epochs of training, outperforming Multi-Modal Net (65.48%) and ESiN in overall accuracy.
• The model achieved the highest accuracy for 'wondering' emotions (85%), attributed to the presence of explicit markers like question marks and interrogative words.
• Neutral emotion detection had the highest accuracy (73%), though many false positives were misclassified as neutral due to the absence of distinct emotional cues.
• Positive emotion detection had the lowest accuracy (57%), likely due to irony and ambiguous expressions in the training data that are hard to distinguish even for humans.
• The model’s performance on neutral states (73%) was better than Multi-Modal Net (71.37%) and comparable to ESiN’s peak performance (over 80%).
• The optimized hyperparameters (γ = 5×10⁻⁶, L = 1.5×10⁻⁴) yielded the best results, with loss decreasing steadily over the first epoch as shown in the training curve.