Cards Against AI Predicting Humor in a Fill-in-the-blank Party Game Dan Ofer The Hebrew University of Jerusalem

2025-04-30 0 0 327.2KB 7 页 10玖币

侵权投诉

Cards Against AI: Predicting Humor in a Fill-in-the-blank Party Game

Dan Ofer

The Hebrew University of Jerusalem

dan.ofer@mail.huji.ac.il

Dafna Shahaf

The Hebrew University of Jerusalem

dshahaf@cs.huji.ac.il

Abstract

Humor is an inherently social phenomenon,

with humorous utterances shaped by what is

socially and culturally accepted. Understand-

ing humor is an important NLP challenge,

with many applications to human-computer in-

teractions. In this work we explore humor

in the context of Cards Against Humanity –

a party game where players complete ﬁll-in-

the-blank statements using cards that can be

offensive or politically incorrect. We intro-

duce a novel dataset of 300,000 online games

of Cards Against Humanity, including 785K

unique jokes, analyze it and provide insights.

We trained machine learning models to predict

the winning joke per game, achieving perfor-

mance twice as good (20%) as random, even

without any user information. On the more dif-

ﬁcult task of judging novel cards, we see the

models’ ability to generalize is moderate. In-

terestingly, we ﬁnd that our models are primar-

ily focused on punchline card, with the context

having little impact. Analyzing feature impor-

tance, we observe that short, crude, juvenile

punchlines tend to win.

1 Introduction

Humor is a universal phenomenon, fulﬁlling im-

portant social roles: approaching social taboos, ex-

pressing criticism against individuals and institu-

tions, and consolidating a sense of belonging to a

group (Ziv,2010). Humorous utterances are shaped

by what is socially and culturally accepted.

Humor underpins many social interactions (Beach

and Prickett,2017;Urbatsch,2022). It increases

likeability and trust (Meyer,2015). Thus, humor is

also a crucial component in developing personable

human-computer interactions.

Speciﬁcally, we focus on the task of humor recog-

nition – determining whether a sentence in a given

context is funny. This task is difﬁcult, as humor is

a diverse, amorphous and complex phenomenon.

It requires creativity and common sense, and is

very challenging to model (Winters,2021;Attardo,

2010), considered by some researchers to be AI-

complete (Stock and Strapparava,2003). Thus,

designing a general humor recognition algorithm

currently seems beyond our reach, and works on

computational humor tend to focus on narrow, spe-

ciﬁc types of humor, such as knock-knock jokes,

one-liners, or even that’s-what-she-said jokes (Mi-

halcea and Strapparava,2006;Taylor and Mazlack,

2004;Kiddon and Brun,2011)

In this work we explore humor in the context of

the immensely popular card game

“Cards Against

Humanity” (CAH)

. The game mechanics are sim-

ple: Players are dealt ten cards (“punchlines”).

The judge of the round draws a “prompt” card

posing a question or a “ﬁll-in-the-blank” state-

ment. Each player submits an answer from their

hand, and the judge picks the winner. An exam-

ple prompt is “TSA guidelines now prohibit ___

on airplanes”. Candidate punchlines are “Gob-

lins”, “BATMAN!!!”, “Poor people”, and “The

right amount of cocaine”. Importantly, many cards

are offensive or politically incorrect.

We introduce a novel dataset of 300K online CAH

games. While most current humor datasets are

lacking in size (Weller and Seppi,2020), or have

weak labels (e.g., upvotes without total views), our

dataset is large and strongly labeled. We train ma-

chine learning models

to predict the winning joke

per round and show models can somewhat general-

ize to novel (unseen) punchline cards. Surprisingly,

we ﬁnd that our models primarily focus on the

punchline card alone, and the impact of the prompt

is limited. We also identify potential behavioral

biases in the data.

Our main goal here is to explore humor through a

data-driven lens, and we believe CAH provides a

1Code available at https://github.com/ddofer/CAH

arXiv:2210.13016v1 [cs.LG] 24 Oct 2022

unique perspective to this end. Most existing stud-

ies on humor recognition formulate the problem

as a binary classiﬁcation task and try to recognize

jokes via a set of linguistic features (Yang et al.,

2015;Purandare and Litman,2006;Zhang and Liu,

2014). One of the common problems those works

face is the construction of negative instances, which

are often sampled from a different domain (e.g.,

news). In contrast, the CAH task does not suffer

from this problem.

Perhaps the closest setting to ours is humorous ﬁll-

in-the-blank (Hossain et al.,2017;Garimella et al.,

2020), where users complete a joke however they

see ﬁt. However, our setting is a lot more restricted:

players choose (rank) an answer from a small set of

options, enabling comparisons that would be hard

to test on other corpora.

From a humor-theory point of view, we believe

CAH serves as an interesting example of frame

blends and frame shifts (Hofstadter and Gabora,

1989;Coulson,2001), where a speaker’s mental

model suddenly shifts to new situations, or two

distinct situations create a hybrid. CAH provides a

relatively clean setting to explore this phenomenon,

as the jokes are short, with simple syntax and nar-

rative structure.

To the best of our knowledge, CAH has only been

explored in the literature through pedagogical, eth-

ical or sociological lenses (e.g., (Strmic-Pawl and

Wilson,2016)), not computational or linguistic

ones. We note the data contains offensive humor,

and should be very carefully used as training data.

However, we believe it is important to study offen-

sive humor too and understand its role in generating

and reinforcing social boundaries and inequalities.

2 Data

The dataset consists of games played on

the online CAH labs website,

https://lab.

cardsagainsthumanity.com

. The players played

the game voluntarily, for fun; they are not our anno-

tators or workers. In each round a user is presented

with a random prompt card, 10 potential punch-

lines cards, and chooses the funniest punchline.

The raw data had 298,955 past games (i.e., we did

not perform any additional experimentation our-

selves).There are 581 unique black prompt cards

and 2,128 white punchline cards, including cards

from the ofﬁcial CAH game and expansions, re-

sulting in 1,236,368 possible unique jokes (where

Figure 1: Card counts. Log scale histogram of prompt

and punchline cards occurrence frequency (i.e., how

many times cards appeared). Prompts have a more

relatively uniform distribution, but both prompts and

punchline cards have a “tail” of rare cards. The spikes

of frequent cards are presumably due to cards from the

standard game, as opposed to experimental cards or ex-

pansions.

a joke is the result of ﬁlling in the blank of the

prompt card with a punchline). Each round is ef-

fectively unique due to the large number of com-

binations. The data we received from CAH did

not include any demographic or geographic char-

acteristics, user identiﬁers or personally identiﬁ-

able information. 5% of games were skipped by

users and were excluded, as were a minority of

prompts that required picking more than one punch-

line. Data is available upon request to CAH at

mail@cardsagainsthumanity.com.

2.1 Data analysis and observations

The frequency of different prompts or punchlines

presented to users is not a uniform distribution (Fig

1). The odds of a punchline card being picked

and winning is also unevenly distributed – perhaps

unsurprisingly, some punchlines are funnier than

others (Fig 2). The data is sparse: the number of po-

tential games is immense (

7.06 ×1054)

. Viewed at

the level of unique jokes (prompt+punchline com-

bined), only 784,974 appear at least once across

the games, out of the 1.23M possible (60%), with

few repeats. If we consider only cases where we

have feedback (a “winning pick”), then we have

only 248,896 jokes with feedback, and of these

77% were picked only once, out of 300,000 games.

A further 17% were picked only twice.

2.1.1 Popular punchlines

Across all games, all punchlines appeared at least

14 times, with

µ= 1149

σ= 334

. We considered

a punchline successful if its win rate is over 20%

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

CardsAgainstAI:PredictingHumorinaFill-in-the-blankPartyGameDanOferTheHebrewUniversityofJerusalemdan.ofer@mail.huji.ac.ilDafnaShahafTheHebrewUniversityofJerusalemdshahaf@cs.huji.ac.ilAbstractHumorisaninherentlysocialphenomenon,withhumorousutterancesshapedbywhatissociallyandculturallyaccepted.Understa...

展开>> 收起<<

Cards Against AI Predicting Humor in a Fill-in-the-blank Party Game Dan Ofer The Hebrew University of Jerusalem.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Cards Against AI Predicting Humor in a Fill-in-the-blank Party Game Dan Ofer The Hebrew University of Jerusalem

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: