Source-linked AI summary
Supporting Perspective Acquisition and Opinion Formation on Societal Issues Through AI-Generated Japanese Rap Battle Debates
Ryota Mibayashi, Toru Urakawa, Dai Takanashi, Tomoya Morohoshi, Kanata Yamagishi, Ryuho Sekikawa, Yasuhiko Nishimura, Yuta Takeuchi, Hideaki Tamori, Takehiro Yamamoto, Hidenari Kiyomitsu, Hiroaki Ohshima
TL;DR
Personalized recommendations can reduce exposure to diverse opinions, making balanced understanding harder. This paper generates and presents AI-written rap battle debates with synthesized audio, then evaluates them with a user study. After viewing, participants identified nearly as many additional opinions and discussion points as they had before, and some reconsidered or changed their stances.
Problem
Personalized recommendation can reduce exposure to diverse opinions, making balanced understanding harder.
Source: §I · p. 1Method
The study generated rap battle-style debates with GPT-5.4-mini and presented them with synthesized audio through a web-based interface for evaluation by six participants viewing 40 societal-issue topics.
Source: §I · p. 2Results
247 additional opinions and discussion points were identified after viewing, compared with 254 identified beforehand, and participants reconsidered or changed stances in 52.5% of evaluated cases.
Source: §I · p. 2, B. Acquisition of New Opinions and… · p. 4Takeaways & Limitations
Abstract
from arXiv · showhide
Acquiring diverse perspectives and forming informed opinions on societal issues are essential for critical thinking and informed decision-making. Although observing debates between opposing viewpoints can promote perspective acquisition, conventional debates require substantial time and human resources, limiting their accessibility. This study investigates whether brief AI-generated rap battle-style debates can support perspective acquisition and opinion formation on societal issues. Rap battles are debate-style performances in which opposing viewpoints are expressed through rhymed verses, making them engaging while preserving the structure of argumentative dialogue. To conduct this investigation, we automatically generated rap battle-style debates using GPT-5.4-mini and presented them to participants with synthesized audio through a web-based viewing interface. We evaluated the educational effectiveness of this approach through a user study involving six participants who viewed 40 debates generated from topics on societal issues used in the All Japan Junior and Senior High School Debate Championships. The results showed that watching approximately 1.5 minute rap battle-style debates nearly doubled the number of opinions and discussion points identified by participants. Furthermore, participants reconsidered or changed their stances in 52.5% of the evaluated cases. These findings demonstrate that even brief AI-generated rap battle-style debates can effectively support perspective acquisition and opinion formation on societal issues.
I. INTRODUCTION
The introduction frames limited exposure to diverse opinions and the time demands of conventional debates as barriers to perspective acquisition. It proposes brief, AI-generated rap battle debates and evaluates them as a way to support opinion formation on societal issues.
I. INTRODUCTION
- Conventional debates can take several tens of minutes per topic, while observing debates can help audiences understand diverse viewpoints and opposing arguments.The introduction notes that listening to debates may foster broader perspective-taking and deliberative capacities.Source: §I · p. 1
- The study investigates whether AI-generated rap battle debates can support perspective acquisition and opinion formation using synthesized rap audio and a web-based viewing interface.The debates were generated with GPT-5.4-mini while preserving the consistency of opposing arguments.Source: §I · p. 2
- The user study used 40 championship-topic debates viewed by six participants, and the authors report that the format maintained engagement and supported comprehension.Rap battles use concise, rhymed exchanges between opposing viewpoints and are presented as an engaging way to expose audiences to diverse perspectives.Source: §I · pp. 1–2
- The paper presents an AI framework for automatically generating and presenting rap battle-style debates in a short format.The introduction positions the framework as a way to deliver debate content in an engaging format.Source: §I · p. 2
- In a user study spanning 40 debate topics, brief AI-generated rap battle-style debates nearly doubled identified opinions and discussion points and prompted participants to reconsider or change their stances in 52.5% of evaluated cases.Source: §I · p. 2
II. RELATED WORK
Related work covers rhyme-aware lyric generation, rap battle-specific generation, and large language model approaches to generating opinions on topics. The paper proposes using GPT-5.4-mini with in-context learning to combine rhyme and coherent debate content.
II. RELATED WORK
- Prior rap-lyric generation explored rhyme-aware word replacement and learning rhyme patterns from lyric corpora –.These approaches address rhyme in lyrics rather than the paper’s combined goal of rhyme and coherent debate content.Source: §II · p. 2
- Rap battle-specific generation has used LSTM- or Transformer-based models, including approaches in,,.Source: §II · p. 2
- The paper proposes GPT-5.4-mini with in-context learning to generate rap battles combining rhyme with coherent debate content.The passage also notes prior large language model methods for generating diverse opinions on given topics.Source: §II · p. 2
III. DEBATE VIEWING SYSTEM
The debate-generation method uses GPT-5.4-mini prompts to produce two opposing, consistent sides in a rap battle format. The requested output combines rhymed Japanese verses, English translations, and a structured exchange.
A. Rap Battle Format Debate Generation
- Each generated debate has two exchanges, four verses total, and eight lines per verse, with the second exchange continuing the topic without repeating earlier arguments.The two rappers represent opposing sides and retain their stance across their respective verses.Source: A. Rap Battle Format Debate Generation · p. 2
- The prompt requests Japanese verses with corresponding English translations, named MCs, and lines of approximately 15 characters.The output template labels each verse and line and distinguishes the affirmative and opposing perspectives.Source: A. Rap Battle Format Debate Generation · p. 2
- The generation prompt asks GPT-5.4-mini to include rhymed expressions while maintaining opposing arguments in a rap battle debate.The system overview shows topic input followed by automatic debate and speech generation and visualized playback.Source: A. Rap Battle Format Debate Generation · p. 2, Figure 2 · p. 3
B. Rap Battle Format Debate Viewing System
The web-based system generates a rap battle debate from a user-entered topic, synthesizes its audio, and displays the text and playback together. Generation and viewing take about two minutes in total.
B. Rap Battle Format Debate Viewing System
- The web system accepts a debate topic, generates a rap battle debate, synthesizes rap audio, and presents the result through an integrated interface.The system builds on an earlier prototype.Source: B. Rap Battle Format Debate Viewing… · p. 3
- The system uses GPT-5.4-mini for debate generation and Style-Bert-VITS2 for speech synthesis, then displays the audio and text with rhyming phrases highlighted.Style-Bert-VITS2 was fine-tuned using rap audio from two rappers.Source: B. Rap Battle Format Debate Viewing… · p. 3
- A debate video takes about 1.5 minutes, and generation plus viewing can be completed in about two minutes, at lower human-effort and time cost than an actual parliamentary debate.The topic-to-generation-and-visualization process itself takes approximately 20 seconds.Source: B. Rap Battle Format Debate Viewing… · p. 3
A. Experimental Setup for the Evaluation
The evaluation used 40 championship debate topics and pre/post viewing responses to assess participants’ understanding, opinions, and stances. Six Japanese-speaking students evaluated two different sets of 20 topics.
A. Experimental Setup for the Evaluation
- The study used 40 societal-issue topics from past All Japan Junior and Senior High School Debate Championships debates.Participants answered evaluation questions before and after watching a generated debate for each topic.Source: A. Experimental Setup for the Evaluation · p. 3
- The 16 evaluation items covered prior understanding, debate educational effects, debate quality, and overall system evaluation.The measures included 5-point ratings, a 6-category stance scale, and free-text responses; stance and opinion responses were collected before and after viewing, while system items were answered once at the end.Source: A. Experimental Setup for the Evaluation · pp. 3–4
- Six participants—four graduate and two undergraduate students—were divided into groups, each evaluating a different set of 20 topics.Each set included 10 junior-high and 10 high-school debate topics.Source: A. Experimental Setup for the Evaluation · p. 4
B. Acquisition of New Opinions and Discussion Points
Viewing the debates nearly doubled the opinions and discussion points participants identified, with stance changes and new perspectives observed across prior-understanding levels. Stance changes were more often refinements than complete reversals.
B. Acquisition of New Opinions and Discussion Points
- Participants reported 247 new opinions and discussion points after viewing, nearly doubling the 254 identified beforehand.The subjective rating for perceived acquisition of new opinions and discussion points was 3.82.Source: B. Acquisition of New Opinions and… · p. 4
- The “Very low” prior-understanding group reached 3.32 times its original number of opinions and discussion points.Participants in the “Very high” group also reported additional opinions and discussion points.Source: B. Acquisition of New Opinions and… · p. 4
C. Effect on Stance Formation
- 63 of 120 responses (52.5%) showed a stance change after viewing, while the subjective reconsideration rating was 3.09.The stance measure compared participants’ responses before and after watching.Source: C. Effect on Stance Formation · p. 4, Figure 4 · p. 4
- Stance changes more often refined or moderated views than reversed them, with SD→SD (22) and D→D (16) common among initially disagreeing participants.Frequent transitions included N→SD (11), SD→SA (7), N→SA (7), and U→D (9).Source: C. Effect on Stance Formation · pp. 4–5
- All 21 initially “Undecided” responses shifted to another stance; 19 (90.5%) moved to agreement or disagreement and two (9.5%) to neutral.The reported pattern indicates stance formation among participants who began undecided.Source: C. Effect on Stance Formation · p. 5
D. Evaluation of Rap Quality
Human ratings were relatively high for interactive exchanges, preservation of meaning, and naturalness as rap, while rhyme inclusion was the lowest-rated aspect.
D. Evaluation of Rap Quality
- Inclusion of responses scored highest at 4.66, while inclusion of rhymes scored lowest at 3.93.Preservation of meaning scored 4.54 and naturalness as rap scored 4.40; rhyme generation was identified as the primary area for improvement.Source: D. Evaluation of Rap Quality · p. 5
- Preservation of meaning scored 4.54, and naturalness as rap scored 4.40.Both aspects received high average scores in the evaluation.Source: D. Evaluation of Rap Quality · p. 5
E. Analysis of Generated Debates
The ambulance-fee example presents opposing arguments and responsive turns, while system ratings indicate that the format supported comprehension and engagement. Participants also reported increased perspectives and stance changes after brief viewing.
E. Analysis of Generated Debates
- In the ambulance-fee example, the two sides emphasize emergency-patient prioritization and equal access, and later turns respond to opponents’ arguments.The example’s English translation contained end-of-line rhymes, while Japanese rhyme generation remained limited.Source: E. Analysis of Generated Debates · pp. 5–6
F. Evaluation of the Debate Viewing System
- All system evaluation questions averaged at least 4.00; interface neutrality scored 4.67, comprehension 4.00, and engagement 4.67.Participants’ willingness to use the system for other topics also averaged 4.67.Source: F. Evaluation of the Debate Viewing… · p. 6
- After approximately 1.5-minute debates, participants nearly doubled identified opinions and discussion points and changed stances in 52.5% of evaluated cases.The study characterizes these results as support for perspective acquisition and stance formation within a short viewing time.Source: F. Evaluation of the Debate Viewing… · p. 6
V. CONCLUSIONS AND FUTURE WORK
The study concludes that brief AI-generated rap battle-style debates can support perspective acquisition and opinion formation on societal issues. Future work will test their educational impact at larger scale and improve rhyme quality while preserving debate consistency.
V. CONCLUSIONS AND FUTURE WORK
- Brief AI-generated rap battle-style debates can support perspective acquisition and opinion formation on societal issues.The six-participant user study also found that they maintained learner engagement and supported comprehension.Source: §V · p. 6
- Larger-scale user studies are planned to further investigate the approach’s educational impact.Source: §V · p. 6
- Future work will improve rhyme quality while preserving debate consistency.Source: §V · p. 6