Datasets·Customer reviews·eval11_douban_100k_4lens_20260420

Douban Chinese Movie Reviews

Douban is a 100,000-review Chinese movie-review corpus. Each of the four production lenses (Behavioral Medicine v3, All Psychology, Sociology, Information Systems) is applied directly to the original Chinese text via that lens's Chinese-language dictionary. No external translation step is involved — the per-lens translation pipeline is internal to each lens's multilingual loading matrix. The texts predominantly discuss a range of sentiments regarding film quality , character performances , and emotional impact . Many reviews express feelings of disappointment or surprise related to the films' execution, often contrasting high expectations with the actual viewing experience. Common themes include humor and emotional resonance , with some passages highlighting specific scenes or dialogue that elicited strong reactions, whether positive or negative. Additionally, there are mentions of technical aspects such as visual effects and cinematography , which are sometimes praised while the narrative coherence is critiqued. Overall, the reviews reflect a diverse array of personal experiences and evaluations of cinematic works. [Summary on 50 random texts by ChatGPT 4o Mini].

Distribution of Douban movie rating (1-5 stars, ordinal)
1
2
3
4
5
8,987 at floor29,920 at ceiling
100,000
items
20,000
holdout n
Douban movie rating (1-5 stars, ordinal)
target
Ordinal
kind
14
systems compared
Criterion validity

Reported holdout systems from the verified card

Ordinal prediction uses Quad. κ as the task-primary metric. Secondary columns keep the companion metrics visible so binary, ordinal, regression, and multiclass cards are not compared through one flattened score.

Model-family mix
PsyProxy · 4Baseline · 10
SystemFamilyVariantQuad. κWithin-oneMAEPrimary scale
psyproxyPsyProxy — Behavioral Sciences Lens v0.5 · 1000d
PsyProxystrict0.4960.8190.82
psyproxyPsyProxy — Sociology Lens v0.5 · 1000d
PsyProxystrict0.4910.8160.83
psyproxyPsyProxy — Health Lens v0.9 · 1100d
PsyProxystrict0.4800.8190.83
psyproxyPsyProxy — Technology Lens v0.5 · 800d
PsyProxystrict0.4760.8120.84
baselineIMDB movie reviews (ACL) binary · via best lens
Baselinestrict
baselineAmazon Video-Games reviews ordinal · via best lens
Baselinestrict
baselineAmazon Video-Games reviews continuous · via best lens
Baselinestrict
baselineDisneyland TripAdvisor reviews continuous · via best lens
Baselinestrict
baselineSentiment140 tweets binary · via best lens
Baselinestrict
baselineDruglib drug reviews ordinal · via best lens
Baselinestrict
baselineDruglib drug reviews regression · via best lens
Baselinestrict
baselineAmazon Video-Games reviews binary · via best lens
Baselinestrict
baselineLIAR fact-check statements ordinal · via best lens
Baselinestrict
baselineDisneyland TripAdvisor reviews binary · via best lens
Baselinestrict