A Complete and Modestly Funny System for Generating and Performing Japanese Stand-Up Comedy

Jonas Sjöbergh, Kenji ArakiCOLING 2008

TL;DR

This paper presents a complete system for generating and performing Japanese stand-up comedy, combining multiple joke generation modules and presenting them via robots. The system achieved promising results in evaluations, with nearly half of the participants rating the overall impression as funny.

What Problem Does This Paper Address?

The paper tackles the challenge of generating not just individual jokes, but complete, coherent comic performances. While humor recognition and generation are active research areas, most systems focus on single joke types. This work aims to bridge the gap by creating a multi-module system that produces a structured stand-up routine, mimicking the Japanese "manzai" style with a "boke" (funny man) and "tsukkomi" (straight man) dynamic.

Methods

The system integrates several joke generation components:

  • Introduction/Closing: A simple selection from a predefined list.
  • Proverb Twisting: Takes common Japanese proverbs and modifies them by substituting words with similar-sounding, often vulgar, alternatives. This involves a morphological analyzer (ChaSen) to get pronunciations and a similarity table for sound matching.
  • Wordplay Jokes (Database): Utilizes a large database of Japanese wordplay jokes, selecting those that incorporate nouns already present in the script.
  • Punning Riddle Generator: Creates riddles by finding nouns in the script that sound similar to dirty words, then constructing a riddle using a pattern and a hint that describes the dirty word.
  • Put-Downs and Come-Backs: A list of generic put-down phrases is used by the "tsukkomi" robot. The "boke" robot can also make come-backs.
  • Performance Assembly: Jokes are chained together, with "boke" telling a joke and "tsukkomi" delivering a put-down.
  • Speech Synthesis and Robot Performance: The generated script is converted to speech using the AquesTalk TTS system with different voices, and then performed by two Robovie-i robots.

Datasets

  • Proverb List: A collection of nearly 1,000 Japanese proverbs and sayings, filtered for commonality by checking web frequency.
  • Dirty Words Collection: A list of a few hundred Japanese dirty words, categorized by type (sex, feces, insults).
  • Wordplay Joke Database: A collection of approximately 2,200 Japanese wordplay jokes, automatically collected from online lists.
  • Put-Down Phrases: A short list of generic put-down phrases provided by an amateur comedian.

Key Insights and Ideas

  • Coherent Performances are Key: The paper highlights that simply generating individual jokes isn't enough; a structured performance with interconnections (like the manzai dynamic) is crucial for a more engaging experience.
  • Mistakes Can Be Funny: The system's inherent errors in joke generation can sometimes be perceived as funny due to the "boke" character's role of saying silly things.
  • Robot Speakers are a Bottleneck: The quality of the robot's speakers significantly impacts the intelligibility and perceived funniness of the performance, suggesting that hardware limitations can be a major hurdle.
  • Humor is Subjective: Despite efforts to generate funny content, individual tastes in humor vary greatly, making objective evaluation challenging.

Key Results

  • Overall Impression: The system's performances received an average overall impression score of 3.3 out of 5 (where 5 is very funny).
  • High Scores: Nearly half (45%) of the total impression scores were 4 or 5, indicating a significant portion of evaluators found the performances funny.
  • Individual Joke Performance:
    • Database jokes (specifically "Database 1a") scored highest on average (3.6/5).
  • Proverb jokes averaged 2.8/5.
  • Riddle jokes averaged 2.4/5.
  • Come-back jokes scored 2.6/5.
  • Qualitative Feedback: Evaluators noted that while the system makes mistakes, joke generation is robust. However, some found dirty jokes too direct and suggested subtlety is funnier. The robot speakers were identified as a major issue.

Limitations and Future Work

  • Robot Speaker Quality: The paper explicitly identifies the weak speakers of the Robovie-i robots as a significant limitation, impacting sound quality and intelligibility. This is noted as an easy fix.
  • Joke Funniness: The generated jokes, particularly riddles and come-backs, were not always perceived as very funny, suggesting room for improvement in joke generation algorithms.
  • Subjectivity of Humor: The authors acknowledge that humor is highly subjective, making it difficult to satisfy all evaluators.
  • Future Work: The authors suggest improving the robot's speaker quality as a primary next step. Further work could focus on enhancing the funniness of individual joke types and potentially exploring more sophisticated ways to connect jokes within a performance.

Summary updated

BibTeX
@inproceedings{sjobergh2008complete,
  author = {Jonas Sjöbergh and Kenji Araki},
  title = {{A Complete and Modestly Funny System for Generating and Performing Japanese Stand-Up Comedy}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2008},
}
Figure