Identifying Topics in Captions

We classified captions into predefined topics (e.g., Love, Family, School, Work, Politics and Social Issues, Nature, Food, Emotions, Entertainment and Pop Culture, Caption Contest, Health, Law, Other) to explore the thematic dimensions of humor in the datasets.

We used a sentence embedding approach to classify captions into topics. For a detailed explanation of the classification process, see the Models & Methods: Topic Classification section.

1. Topic Distributions:

We also created an interactive pie chart to visualize the topic distributions in both datasets. You can explore it below:

This figure compares the normalized topic distributions of captions from NYCC and OHIC. While both datasets cover a similar range of themes, clear differences emerge in their relative emphasis across topics.

NYCC captions show a stronger concentration in Food, Work, Nature, and Health. Food is the most prominent topic, reflecting shared cultural experiences around eating, while Work and Nature highlight professional life and environmental humor.

In contrast, OHIC captions are dominated by Caption Contest, reflecting its self-referential and meme-driven nature. OHIC also emphasizes School, Emotions, Entertainment and Pop Culture, Law, and Love, showcasing a more expressive and pop-culture-oriented humor style.

Politics and Social Issues are limited in both datasets but slightly more represented in OHIC. These differences highlight contrasting humor strategies: NYCC favors situational and observational humor, while OHIC leans toward self-referential and emotionally expressive humor. A limitation of this analysis is that the content of the images, which could heavily influence caption themes, is not considered.

2. Topic vs. Funniness Scores:

This figure shows the distribution of funny scores across topics for both NYCC and OHIC captions. While the distributions reveal substantial overlap, humor appreciation is influenced more by caption quality, phrasing, and context than by topic alone.

For most topics, including Food, Entertainment and Pop Culture, Caption Contest, and Law, NYCC and OHIC captions achieve similar median funny scores, indicating comparable humorous effectiveness. However, Politics and Social Issues is the only topic where OHIC captions outperform NYCC, likely due to a more direct engagement with contemporary political content. In contrast, NYCC captions score higher for topics like Love, Family, School, Nature, and Health, suggesting more consistent humor in these themes.

The high variance in funny scores across all topics, as shown by wide interquartile ranges and long whiskers, highlights that humor effectiveness depends on execution rather than thematic category. The Other category, as expected, shows the widest and most uneven distribution due to its heterogeneous nature.

Overall, the topic classification highlights the diverse thematic dimensions of humor in the datasets. Relational topics like Love and Family resonate broadly with audiences, with NYCC captions achieving higher funniness scores in these themes. Institutional topics like Health and Law provide humor rooted in societal critique, with both datasets showing comparable effectiveness in Law. The prominence of topics like Nature in NYCC reflects its focus on observational humor, while the higher representation of Politics and Social Issues in OHIC highlights its engagement with contemporary political content. These findings underscore the role of context, stylistic choices, and audience preferences in shaping the thematic content and reception of humor.

Humor Types by Topic

This heatmap reveals how different topics influence and dictate the type of humor used in cartoon captions. The analysis highlights the interplay between humor types and thematic topics, showcasing how context shapes the humor style.

Key insights from the heatmap include:

• Irony and Satire: Dominates across all topics, especially in Politics and Social Issues and Work. Political humor is primarily critical or observational, while office-related humor often highlights the gap between corporate expectations and reality.
• Absurdist: Prominent in Food and Nature, often stemming from personifying animals or objects (e.g., a talking leaf). It also plays a significant role in Caption Contest captions, aligning with the visually bizarre nature of the cartoons.
• Affiliative: Most prominent in interpersonal topics like School, Family, and Love. This humor type fosters relatability, creating a "we've all been there" connection with the audience.
• Dark: While generally low across the dataset, it peaks in Health and Emotions, reflecting the grim reality and fatalistic attitude associated with health issues and psychological themes like therapy and depression.
• Offensive and Sexual: Extremely low across all topics, except for Love in the case of Sexual humor. This reflects NYCC’s editorial standards and its high-brow audience, which prioritizes indirect humor over visceral humor.

Overall, this analysis underscores how humor types adapt to the thematic context of captions. While some topics favor specific humor styles (e.g., Irony in Politics), others like Food and Nature lean toward Absurdist humor, reflecting the creative interplay between visuals and captions.

Word Clouds

Word clouds reveal the most frequent words associated with each humor type and topic, providing a quick overview of recurring themes in the datasets. These visualizations help identify the dominant vocabulary and recurring patterns in captions.

1. Word Clouds for Humor Type:

For humor types, the word clouds highlight the distinct vocabulary used in different styles of humor, reflecting the unique contexts and audiences of the NYCC and OHIC datasets:

• Affiliative: In NYCC, words like "home," "feel," and "right" emphasize a sense of belonging and emotional connection. In OHIC, words like "meme," "day," and "mom" reflect humor rooted in everyday life and internet culture.
• Irony/Satire: NYCC captions use words like "sure," "meant," and "guess," highlighting subtle and indirect commentary. OHIC captions, with words like "money," "right," and "upvote," focus on social critique and internet-specific humor.
• Absurdist: NYCC captions feature words like "sorry," "little," and "way," showcasing surreal or illogical scenarios. OHIC captions, with phrases like "oh man," "eat," and "right," lean into exaggerated and nonsensical humor.
• Dark: Both datasets share a focus on grim themes, with words like "death," "die," and "life" in NYCC and "kill," "dead," and "die" in OHIC, reflecting humor that makes light of taboo or morbid topics.
• Sexual: NYCC captions use words like "big," "size," and "honey," suggesting subtle innuendos. OHIC captions, with words like "girl," "ass," and "sex," are more explicit and direct in their approach.
• Offensive: NYCC captions include words like "Trump," "mother," and "white," reflecting political and societal critique. OHIC captions, with words like "gay," "stupid," and "karen," lean into internet-driven, provocative humor.

This analysis highlights the contrasting humor strategies of the two datasets. NYCC humor often relies on subtlety, cultural critique, and layered meanings, while OHIC humor is more direct, internet-centric, and emotionally charged. These differences underscore the influence of context and audience on the vocabulary and themes used in humor.

2. Word Clouds for Topics:

For topics, the word clouds illustrate the thematic focus of captions:

NYCC Caption Dataset:

• Love: Dominated by words like kiss, heart, marriage, and sex, reflecting interpersonal relationships and modern dating life.
• Family: Features terms like home, Mom, and kid, emphasizing domestic life and parental relationships.
• School: Includes school, class, and college, highlighting educational settings and recurring routines.
• Work: Words like job, boss, and office point to workplace stress and corporate culture.
• Politics and Social Issues: Focuses on politics, debate, and climate change, emphasizing systemic critique.
• Nature: Includes trees, planet, and global warming, blending environmental concerns with humor.
• Food: Features restaurant, pizza, and menu, reflecting shared cultural experiences around eating.
• Emotions: Marked by introspective terms like feel, cry, and therapy, emphasizing emotional vulnerability.
• Entertainment and Pop Culture: Includes movie, TV, and star, reflecting humor grounded in media and celebrity culture.
• Caption Contest: Self-referential terms like caption, joke, and funny dominate.
• Health and Law: Specialized vocabularies focus on doctor, hospital, lawyer, and court.
• Other: Generic terms like thing and people, serving as a residual category.

OHIC Caption Dataset:

• Love: Words like relationship, crush, and friend emphasize modern dating culture and youth-oriented humor.
• Family: Informal terms like mom, dad, and house focus on everyday domestic situations.
• School: Includes teacher, homework, and exam, reflecting shared student struggles.
• Work: Words like job, boss, and money highlight time pressure and compensation.
• Politics and Social Issues: Features government, election, and freedom, with explicit references to contemporary figures like Trump.
• Nature: Focuses on tree, sky, and rain, emphasizing situational contrasts over climate discourse.
• Food: Includes chicken, Taco Bell, and pizza, reflecting humor tied to pop culture and consumption habits.
• Emotions: Words like sad, happy, and anxiety emphasize explicit emotional expression.
• Entertainment and Pop Culture: Features anime, Fortnite, and Disney, relying on internet culture and fandoms.
• Caption Contest: Includes meme, post, and funny, highlighting meta-awareness of captioning as a social activity.
• Health: Pandemic-related terms like covid and vaccine dominate.
• Law: Includes lawyer, court, and crime, framed through dialogue and interpersonal tension.
• Other: Generic conversational tokens like day and thing.

Stylistic Analysis

Our stylistic analysis focused on three key aspects of the captions: word length distributions, formality scores, and part-of-speech distributions. These metrics provide insights into the linguistic characteristics of captions in the NYCC and OHIC datasets, highlighting differences in tone, structure, and word usage.

1. Word Length Distributions:

The word length distributions reveal that NYCC captions are more concentrated in length, particularly for the most highly ranked captions, while OHIC captions show a broader distribution. Interestingly, longer captions are more represented among the funniest captions in the OHIC dataset, suggesting a preference for more elaborate humor.

2. Part-of-Speech Distributions:

The part-of-speech distributions highlight differences in word usage between the two datasets. NYCC captions feature a higher density of nouns, adjectives, and verbs, reflecting a more descriptive and action-oriented style. OHIC captions, on the other hand, show a higher density of interjections and unknown words, indicative of a more casual and expressive tone. The higher proportion of interjections in OHIC aligns with its internet-based context, where captions often rely on exaggerated expressions and slang.

3. Formality Score Distributions:

The formality score distributions reveal distinct patterns between the two datasets. Captions in the NYCC dataset tend to have higher formality scores, with a peak in the mid-range, reflecting a balance between conversational and formal tones. In contrast, captions in the OHIC dataset exhibit a broader distribution, with a higher density of lower formality scores, indicating a more casual and internet-driven tone. This difference aligns with the institutional context of NYCC and the internet-based, meme-driven nature of OHIC.

Notably, the mean formality for NYCC captions is slightly higher than that of a US president speech, while the mean formality for OHIC captions is comparable to one, right above that of a script from "The Simpsons". Additionally, the formality scores for OHIC captions show higher variability compared to NYCC, reflecting a broader range of tones in the OHIC dataset. However, both datasets exhibit much lower formality than a typical New Yorker article.

These stylistic differences underscore the influence of context and audience on the linguistic characteristics of captions. While NYCC captions balance formality and descriptive language to appeal to a broader audience, OHIC captions embrace a more casual and expressive style, resonating with internet culture.

How Humor Evolves Over Time

The evolution of humor topics in the New Yorker captions reveals significant trends over time, reflecting societal events and cultural shifts. By analyzing the proportions of topics across years, we can observe how external factors influence the thematic focus of captions.

1. Humor Types Proportions Over Time:

This figure shows the yearly proportions of humor types in NYCC captions. The shift in humor styles reveals how the public adapted to the "shared challenges" of the last decade:

• 2016–2019: The "Ironic" Era: This period is dominated by Irony/Satire (peaking at 55.3% in 2017) and Absurdist humor. Interpretation: Users relied on sophisticated irony and surrealism to critique a turbulent political landscape. The high levels of absurdism suggest a preference for "nonsense" as a way to handle a world that felt increasingly illogical, while maintaining emotional distance.
• 2020–2021: The Reorientation: A massive structural shift occurred in 2021, with Affiliative humor jumping from 7.3% to 33.5%. Interpretation: This period marked a moment of reconnection. As the initial shock of the pandemic faded into a "new normal," humor shifted from "cleverness" to "shared humanity," prioritizing jokes that build bridges over jokes that mock or distort.
• 2022–2024: The "Relatability" and "Provocation" Era: Affiliative humor reached its zenith in 2023 (43.9%), representing a permanent shift toward "bottom-up" relatability. Interpretation: After prolonged isolation, the collective priority shifted to humor that validates shared social experiences. Additionally, 2023 saw a spike in Sexual (5.0%) and Offensive humor, reflecting a "social unmasking" as the public moved past restrictions and tested the limits of new norms.

2. Topic Proportions Over Time:

This figure presents a line plot of normalized topic distributions in NYCC captions over time. It highlight the dominance of certain themes, such as Food and Caption Contest, which consistently maintain high proportions across the years. These topics reflect universal and timeless themes that resonate broadly with audiences. Other topics, such as Love, Family, and Nature, remain steady over time, with occasional spikes reflecting specific cultural or societal moments.

3. Topic Variation Heatmap:

This figure presents a heatmap of the standardized variation (z-scores) of topic proportions in NYCC captions over time. Deviations from the mean highlight periods during which certain themes became unusually prominent or marginal, often aligning with major social, political, and global events.

The heatmap reveals several notable trends:

• Love and Family: These themes remain steady over the years, with sharp rises in 2019 for Love and 2022 for Family, possibly reflecting cultural shifts or societal well-being.
• School: Mentions of School increase substantially in 2021 and 2023, aligning with Covid’s long-term impact on education and the launch of ChatGPT, which affected university students.
• Work: Overrepresented in 2016, reflecting pre-pandemic concerns about corporate culture, with renewed interest in 2023 linked to post-pandemic work reorganization.
• Politics and Social Issues: Peaks in 2016 and 2019 align with the U.S. presidential election and Trump impeachment proceedings, followed by a decline after 2020 as political satire gave way to more personal themes.
• Nature: Peaked in 2024, reflecting increased public concern over climate change and environmental issues.
• Food: Consistently the most prominent topic, reflecting its universal appeal and cultural significance.
• Emotions: Declined in 2017 but rose steadily afterward, capturing post-pandemic emotional reflection and a broader openness to discussing mental health.
• Entertainment and Pop Culture: Steady over the years, with a slight increase from 2022, reflecting the rise of TikTok, Instagram reels, and global events like the Oscars and World Cup.
• Caption Contest: Dominates NYCC captions alongside Food, with a sharp rise in 2024, possibly due to increased audience familiarity with the contest format.
• Health: Peaked in 2020 and 2023 due to the Covid pandemic and its aftermath, with a decline in 2024 reflecting pandemic fatigue.
• Law: Peaked in 2016 during election-related legal disputes, with a decline in 2022–2024 as media focus on high-profile court cases waned.

Overall, these analyses demonstrate how external events and societal changes shape the thematic focus of humor over time. By examining these trends, we gain insights into the dynamic interplay between cultural context and humor.

Conclusion: NYCC and OHIC Comparison

The New Yorker Caption Contest dataset (NYCC) and Oxford Humor in Context dataset (OHIC) share similar humor distributions, but they differ in context, audience preferences, and humor styles. Below is a comparison of the two datasets:

	NYCC	OHIC
Context	Institutional humor, focused on cartoons and structured contests.	Internet culture, focused on memes and pop culture.
Audience	Niche, sophisticated audience with a preference for subtle, intellectual humor.	Broader, younger, internet-savvy audience with a preference for relatable and expressive humor.
Humor Style	Subtle, clever wordplay, and observational humor.	Direct, absurd, and emotionally expressive humor.
Structure	Highly structured, with detailed metadata (e.g., votes, captions, image descriptions).	Less structured, with fewer metadata fields but broader coverage of internet humor.
Engagement	Balanced engagement across contests, with most images receiving a moderate number of captions.	Highly uneven engagement, with a few images dominating in captions and votes.
Caption Length	Shorter captions (~8 words) perform better, emphasizing brevity and punchiness.	Shorter captions also perform better, though longer captions can succeed with elaborate humor.
Stylistic Features	Higher formality, descriptive language, and balanced tone.	Casual, expressive tone with a higher density of interjections and slang.
Topic Focus	Dominated by universal themes like Food, Work, and Nature.	Dominated by internet-driven themes like Caption Contest, Entertainment, and Emotions.
Temporal Trends	Peaks in Irony/Satire (2016–2019) and Affiliative humor (2020–2024).	Increased focus on Politics and Pop Culture during societal shifts.
Humor Distribution	Follows a power-law distribution, with a small subset of captions dominating funniness scores.	Also follows a power-law distribution, with a few captions dominating the "Funny" category.

This comparison highlights how context influences humor preferences. NYCC leans toward structured, institutional humor, appealing to a niche audience, while OHIC embraces the dynamic, fast-paced nature of internet culture, resonating with a broader audience.