DISTANCED LAUGHTER - PUBLIC DATA DOCUMENTATION
Author: Yue Xing, University of Illinois Urbana-Champaign
Contact: yxing16@illinois.edu
Release: 2026-09-21

RELATED STUDY
Mapping Audience Emotion in YouTube Short-Form Drama Comments:
A Multi-Method Computational Reception Study.

1. DATA ORIGIN AND UNIT OF OBSERVATION
Data were collected in May 2026 using the YouTube Data API v3.
The cleaned corpus contains 16,683 English-language top-level comments.
Short-form: 15 female-oriented videos from DramaTime and 15 male-oriented
videos from Dragon King Drama. Long-form: 3 female-oriented and 3 male-oriented
series from Tencent Video, with comments collected at the episode level.
The short-form videos are compilations, not short individual uploads.
Female-oriented and male-oriented describe content categories, not viewer gender.
Collection used relevance order, up to 500 comments per video, excluding replies.
cld3 language detection retained comments classified as English.
The original cleaned file has unique comment IDs and no missing text.

2. PUBLIC RELEASE AND CLEANING
The main CSV preserves one row per original cleaned comment, in source order.
It contains numeric/text-derived features and saved model labels; it is not
a release of verbatim comment text. Comment text, platform video/comment IDs,
user IDs, exact timestamps, and model rationales have been omitted.
New local observation and content IDs support grouping within this release.
They are not hashes of the original platform identifiers. Removing direct
identifiers reduces traceability; it does not guarantee complete anonymity.
Raw source files were preserved. No new API data or LLM labels were collected
for this website. Emoji and syuzhet features were recomputed from source text.

Source metadata contains three invalid video IDs (#NAME?). That metadata table
is not redistributed here and was not used to join the public comment dataset.
All content IDs are derived from valid video IDs in the comment table itself.
There are 235 videos with retained English comments: 80 long/female-oriented,
125 long/male-oriented, and 15 in each short-form group.

3. FILES
distanced-laughter-analysis.csv: 16,683 rows, one observation per comment.
llm-labels.csv: 1,000 rows, one observation per saved model-labeled comment.
group-summary.csv: 4 rows, descriptive corpus totals and sentiment means.
emotion-distribution.csv: 28 rows, 7 emotion categories across 4 groups.
emoji-distribution.csv: occurrence counts by symbol and group.
emotion-distribution.png/.svg: static full-group heatmap.
emoji-comparison.png/.svg: selected emoji comparison between formats.

4. MAIN DATA DICTIONARY
observation_id: release-local unique comment identifier (obs_00001, etc.).
content_id: release-local video identifier; comments on the same source video
share this identifier. Different episodes are different content units here.
drama_type: short or long (narrative format, not uploaded video duration).
genre_label: female-oriented or male-oriented industrial genre category.
word_count: count of whitespace-separated nonempty tokens; includes emoji-only
tokens when separated by whitespace, so it is an approximate text-length measure.
emoji_count: number of emoji occurrences detected with R emoji::emoji_count.
has_emoji: TRUE if emoji_count > 0; FALSE otherwise.
syuzhet_score: unnormalized lexicon sentiment score from
syuzhet::get_sentiment(text, method="syuzhet"). Not a probability or fixed scale.
comment_like_count: likes reported at collection time.
reply_count: replies reported at collection time; reply text was not analyzed.
llm_sampled: TRUE for the 1,000 comments in the saved model-labeled sample.
primary_emotion: saved primary model label; blank for unselected comments.
secondary_emotion: optional saved secondary label; blank means absent or
unavailable. Consult llm_sampled to distinguish unselected comments.

5. LLM SAMPLING AND LABELS
250 comments were sampled per group (random seed 42) from comments with at
least 3 whitespace-separated tokens. A 200-comment exploratory coding stage
informed seven categories consolidated by the researcher. The saved labels use
gpt-4o-2024-08-06 with temperature 0. Every sampled comment has one primary label.
The website uses saved labels and does not invoke a model interactively.
Category definitions:
affection_love: attachment to characters, actors, or romantic pairs.
critical_negative: negative evaluation of plot, acting, or production quality.
humor_amusement: finding the content funny or enjoying absurd/exaggerated elements.
anticipation_curiosity: eagerness for narrative continuation or developments.
empathy_sadness: sadness, sympathy, or being moved by a character's situation.
surprise: shock or astonishment at unexpected developments.
none_unclear: no clearly identifiable emotion, questions, or meta-comments.

6. DENOMINATORS
Corpus counts: long/female-oriented 8,830; long/male-oriented 2,034;
short/female-oriented 4,687; short/male-oriented 1,132.
Emotion-bearing denominators: 242, 231, 244, 238 in the same order.
All-sample denominator: 250 in every group.
emotion-distribution.csv: count is the label count; total_sample is 250;
emotion_bearing excludes none_unclear; pct_all = 100*count/total_sample;
pct_emotion = 100*count/emotion_bearing, blank for none_unclear.
The website defaults to pct_emotion and lets readers select pct_all.
The 45.0% critical-negative result is 107/238, versus 42.8% (107/250)
if unclear cases remain in the denominator. Displayed values round to 1 decimal.

Emoji: extraction uses emoji::emoji_extract_all; emoji_modifier_remove normalizes
modifiers. Repeated emoji count separately. Emoji composition uses all normalized
emoji occurrences in the selected group, not only the displayed top symbols.
Pooled format results sum counts across orientations; they are not equal-weighted.
Emoji-comment rate = comments containing at least one emoji / all comments.
Emoji distribution fields: emoji (symbol), name (conventional name), count,
total_tokens (group denominator), pct_tokens (100*count/total_tokens).
Group-summary fields: comments, videos (distinct videos with retained comments),
comments_with_emoji, emoji_tokens, emoji_comment_pct, mean_syuzhet.
The earlier top-five emoji-subgroup chart is not used: its denominator did not
include all Smileys & Emotion subgroups. This release uses individual symbols.

7. LIMITATIONS AND REPRODUCIBILITY
This is an observational, self-selected, English-language, relevance-ranked
sample from selected channels. Groups are unequal and comments are clustered
within videos and series. Genre and channel are confounded in the short-form
sample. Compilation versus episode presentation may also affect responses.
Descriptive differences do not establish causation, viewer demographics, or
representativeness of all audiences. LLM categories and emoji names are
interpretations, not ground truth about inner feelings.
The public release reproduces the displayed distributions and enables analysis
of derived features. It cannot independently reproduce text-based scores or
validate model labels because original comment text is not redistributed.
Emoji-removal robustness and human-validation results are not presented here
because their row-level records were not available in the supplied package.

8. CITATION
Xing, Yue. 2026. Distanced Laughter: Audience Emotion in YouTube Drama Comments.
Public feature dataset and visualizations, September 21 release.
Please cite the researcher and retain this documentation when reusing the data.
