weibo sociolinguistics corpus

maikurokosmos_socioling_weibo_dataset is a Chinese-language corpus of real Weibo (微博) discourse, sampled from daily hot-search topics and annotated for sociolinguistic research. It is built to grow: most days a new hot-search topic is added, and topics are deliberately spread across different domains, so the corpus steadily accumulates posts across many subjects, registers, and communities, with a standing interest in fandom (饭圈) language and online-community slang.

view the dataset on hugging face ↗

1. at a glance

• Source: public posts from Weibo, sampled per daily hot-search topic (话题/热搜).
• Cadence: roughly one new topic per day, ~200 top posts collected each time. The corpus keeps growing.
• Size: 1,000+ rows so far, and counting.
• Language: Chinese (Mandarin; includes internet slang, fandom jargon, and code-mixing).
• Unit of analysis: one row = one Weibo post.
• Model: background extraction and labeling currently use Qwen3.7-Max (the latest at time of writing); model changes will be noted here.
• Annotations: two layers, an AI-generated first pass plus a human-reviewed subset (see the review layer).
• Format: CSV on Hugging Face (auto-converted to Parquet); single train split.
• License: CC BY-NC 4.0, for non-commercial academic research (see license).

2. how the corpus is built

Each day’s batch goes through the same pipeline:

1. Pick a topic. Select one Weibo hot-search topic that has meaningful volume / traction, aiming day to day for a topic from a different domain than recent ones.
2. Collect posts. Gather roughly 200 popular posts under that topic.
3. Ground the topic (AI, pass 1). The model (currently Qwen3.7-Max) reads the topic and produces a background brief: what the event is, plus a list of the key entities and terms involved. This context matters, because much of Weibo discourse is unintelligible without knowing who and what is being referred to.
4. Label the posts (AI, pass 2). A second pass by the same model, conditioned on that background brief, assigns the annotation labels to each post.
5. Human review. Each day a sample of up to 50 posts is drawn (partly at random, partly by rules that flag labels likely to be wrong) and checked by hand. The reviewer confirms or corrects each label, and the result is stored separately in the reviewed_* fields. Comparing the initial labels against this reviewed subset is how accuracy / F1 for the labeling is computed.
6. Publish. The batch is appended to the dataset.

3. what’s in each row

Every post carries some metadata, the raw text, an initial (AI) label set, and a parallel reviewed (human) label set.

4. annotation scheme (the codebook)

The full label set is below. Every dimension also has a reviewed_* counterpart (see the review layer). A couple of values carry non-standard spellings, e.g. outfit_or_appearence and idol_doration; these are the literal field values, kept as-is.

validity

Whether the post is usable, and if not, why.

value	meaning
general	a normal, on-topic post that carries interpretable content.
noise	off-topic or content-free noise (ads, spam, gibberish, unrelated chatter).
stanning_and_other_spamming	stan-style bulk posting, i.e. fan control-comment / vote-farming (饭圈控评、打投) and similar spam.

target_type

What the post is mainly about, or directed at.

value	meaning
person	a specific person.
outfit_or_appearence	someone’s outfit, looks, or appearance.
action	an action, behavior, or event.
opinion	a stated opinion or viewpoint.
group_or_organization	a group or organization.
works_or_creations	a work or creation (a song, show, artwork, product…).
scenery	scenery or a scene.
others	anything else.

stance

The post’s stance toward its target.

value	meaning
support	supportive / approving.
oppose	opposed / critical.
neutral	balanced or objective.
no_clear_stance	no discernible stance.
irrelevant	not actually about the target.

emotion_category & emotion_subtype

The dominant emotion (emotion_category) and a finer emotion within it (emotion_subtype).

emotion_category	emotion_subtype
anger	personal_anger, moral_outrage, fury
disgust	physical_disgust, moral_contemp, rejection
annoyance	annoyance
like	idol_doration, simple_liking, admiration, solidarity
sadness	disappointed, empathy, grieve, hopeless
happiness	simple_happy, pride, gloat, looking_forward_and_hope
fear	anxious, fear, concerned
surprise	astounded, happily_surprised, shock
none	general
uncertain	uncertain

mf_main & mf_direction (moral framing)

mf_main is the moral foundation a post invokes (from Moral Foundations Theory); mf_direction is which pole of that foundation is expressed, so its allowed values depend on mf_main.

mf_main	mf_direction
care_vs_harm	care · harm
fairness_vs_cheating	fairness · cheating
loyalty_vs_betrayal	loyalty · betrayal
authority_vs_subversion	authority · subversion
sanctity_vs_degradation	sanctity · degradation
liberty_vs_oppression	liberty · oppression
none	none

mf_direction also allows uncertain as a standalone value.

satire

value	meaning
yes	uses irony, sarcasm, or satire, so surface wording may invert the intended meaning.
no	literal / sincere.

the review layer

Every field above has a reviewed_* twin, plus a review_selected flag.

• review_selected: whether the row was pulled for human review. Selection is partly random, partly rule-based, so rows whose labels look likely to be wrong are flagged for a closer look.
• reviewed_*: the human verdict. If it matches the initial label, the original was judged correct; if it differs, it has been corrected.
• Comparing the initial labels against reviewed_* over the reviewed subset is exactly how the labeling’s accuracy / F1 is measured.

5. what you can study with it

Because topics are deliberately spread across different domains, and because some labels specifically flag fandom (饭圈) stanning and similar spam, this corpus is aimed at:

• Community & subculture language: how particular communities and subcultures actually talk; the stanning_and_other_spamming validity value and fandom-flavored emotions like idol_doration are built in on purpose.
• Comparing communities: whether different communities show different linguistic habits, framing, and emotional registers around comparable events.
• Tracking internet slang: following buzzwords and memes as they surface, spread, and fade across topics and days.
• Language change, sociolinguistically: how usage shifts over time, tracked across a steadily growing, dated corpus.
• Stance, emotion & moral framing in the wild: Chinese-language stance / emotion / moral-foundation classification on messy real text.
• AI-vs-human labeling: the initial vs. reviewed_* split is a ready-made benchmark, and the basis for the reported accuracy / F1.

6. limitations & responsible use

7. license & attribution

8. help audit, or send feedback