[Paper Review] Convolutional Neural Network for Humor Recognition.

Lei Chen, Chongmin Lee|arXiv (Cornell University)|Feb 8, 2017

Humor Studies and Applications19 citations

TL;DR

This paper proposes a Convolutional Neural Network (CNN)-based approach for automatic humor recognition in spoken utterances, using a newly created, open, and diverse TED talk corpus. The CNN outperforms traditional linguistic feature-based methods by achieving higher accuracy and automatically learning relevant humor cues from raw text.

ABSTRACT

For the purpose of automatically evaluating speakers' humor usage, we build a presentation corpus containing humorous utterances based on TED talks. Compared to previous data resources supporting humor recognition research, ours has several advantages, including (a) both positive and negative instances coming from a homogeneous data set, (b) containing a large number of speakers, and (c) being open. Focusing on using lexical cues for humor recognition, we systematically compare a newly emerging text classification method based on Convolutional Neural Networks (CNNs) with a well-established conventional method using linguistic knowledge. The CNN method shows its advantages on both higher recognition accuracies and being able to learn essential features automatically.

Motivation & Objective

• To develop an automatic method for evaluating humor usage in spoken presentations.
• To address the lack of balanced, diverse, and open-source humor datasets in existing research.
• To compare deep learning (CNN) with conventional linguistic feature-based approaches for humor recognition.
• To investigate whether CNNs can learn effective humor features directly from text without manual feature engineering.

Proposed method

• The authors constructed a new presentation corpus from TED talks, including both humorous and non-humorous utterances from a large number of speakers.
• The corpus is designed to be homogeneous, balanced, and publicly available to support reproducible research.
• A CNN model is trained on raw text sequences to classify utterances as humorous or not, using word embeddings as input.
• The CNN employs multiple filter sizes to capture local n-gram patterns across the input sentence.
• The model uses max-pooling over the feature maps to extract the most salient representations.
• The performance of the CNN is compared against a conventional method relying on manually engineered linguistic features (e.g., lexical and syntactic cues).

Experimental results

Research questions

• RQ1 Can a CNN-based model achieve higher humor recognition accuracy than traditional linguistic feature-based methods?
• RQ2 Does the CNN model effectively learn humor-relevant features directly from raw text without explicit feature engineering?
• RQ3 How does the performance of the CNN model compare when trained on a diverse, speaker-balanced, and open corpus?
• RQ4 To what extent can the CNN generalize across different speakers and speaking styles in natural spoken language?

Key findings

• The CNN-based method achieved higher humor recognition accuracy compared to the conventional linguistic feature-based approach.
• The CNN model demonstrated the ability to automatically learn essential humor-related features from raw text, reducing reliance on manual feature engineering.
• The proposed corpus provides a balanced, diverse, and open resource suitable for training and evaluating humor recognition systems.
• The use of a homogeneous data source with both positive and negative instances improved model generalization and reliability.