Abstract
Humor is an essential component in personal communication. How to create computational models to discover the structures behind humor, recognize humor and even extract humor anchors remains a challenge. In this work, we first identify several semantic structures behind humor and design sets of features for each structure, and next employ a computational approach to recognize humor. Furthermore, we develop a simple and effective method to extract anchors that enable humor in a sentence. Experiments conducted on two datasets demonstrate that our humor recognizer is effective in automatically distinguishing between humorous and non-humorous texts and our extracted humor anchors correlate quite well with human annotations.
Introduction
Humor represents a significant challenge for Natural Language Processing (NLP) due to its subjective nature and reliance on non-literal linguistic structures. While previous work has largely focused on binary classification—determining whether a text is humorous or not—this research from Carnegie Mellon University extends the scope to Humor Anchor Extraction. This involves identifying the specific linguistic "anchors" that trigger a humorous effect, such as the wordplay in a pun.
Humor Recognition via Latent Structures
The authors identify four primary semantic structures that characterize humor and design specific features for each:
- Incongruity: The semantic distance between concepts, modeled using Word2Vec to find clashing ideas.
- Ambiguity: Multiple word senses or semantic distance between senses, identified via WordNet.
- Interpersonal Effect: Sentiment and subjectivity, reflecting the "hostility" or strong affect often found in humor.
- Phonetic Style: Alliteration and rhyme patterns, common in one-liners, identified using the CMU Pronouncing Dictionary.
A Random Forest classifier is employed to integrate these features, distinguishing humorous text from carefully selected non-humorous controls.
Humor Anchor Extraction: The Maximal Decrement Algorithm
A key contribution of this work is the Maximal Decrement Algorithm (MDE) for identifying humor anchors. The process begins by identifying candidate words through syntactic parsing. The algorithm then measures the "humor score" of the full sentence and systematically removes subsets of candidates. The subset whose removal causes the largest drop in the predicted humor score is identified as the anchor.
This approach highlights the "fragility" of humor; removing just one or two key words often causes the statistical signature of the joke to collapse, proving that humor is often concentrated in specific linguistic triggers.
Experimental Setup and Datasets
The models were evaluated on two datasets: Pun of the Day (2,423 puns) and 16,000 One-Liners. To ensure the models were recognizing humor rather than just topical differences, negative samples were drawn from news sources and proverbs, then filtered to match the positive samples in length and vocabulary. This prevents the classifier from "cheating" by simply recognizing domain-specific vocabulary.
Results
The combined model, integrating Word2Vec with human-centric features (HCF), achieved state-of-the-art results in humor recognition.
As shown in the table above, the Word2Vec+HCF model reached an accuracy of 85.4% on puns and 79.7% on one-liners. For anchor extraction, the MDE method achieved an F1 score of 0.756 on puns when compared to human annotations, demonstrating that the model's identified anchors correlate strongly with human perception.
Conclusion
This research demonstrates that humor can be modeled through latent linguistic structures and that its "anchors" can be effectively extracted by measuring score decrements. While the model performs well on short texts, future work aims to address multi-word semantic units and incorporate broader cultural context to handle more complex forms of humor like irony and sarcasm.
Summary updated