SUGGESTED WORKFLOW
Analyse Chinese text
Choose a measurable task, establish a human reference set, and inspect errors before scaling up.
An editorial workflow to adapt, not a benchmark reproduced by this site.
Prepare
A permitted Chinese corpus, a construct definition, sampling rules, and human annotations for a pilot set.
Define the construct
Specify what the score or label means in your study. Write examples and counterexamples; sample across relevant subgroups and time periods.
Build a simple baseline
Start with a transparent dictionary or classifier. cntext provides Chinese text-analysis components; topic and embedding tools are alternatives for different tasks, not interchangeable measures.
Evaluate independently
Hold out human-labelled examples. Compare the baseline with a candidate model without tuning on the test set; inspect false positives and false negatives by subgroup.
Scale with a record
Pin package and model versions, save prompts where relevant, and document preprocessing. Recheck a sample when the domain or model changes.
Expected output
A versioned analysis script, a held-out evaluation, subgroup error notes, and a justified interpretation of the resulting measures.
Check your result
Evaluate against held-out human labels using metrics suited to the task. Inspect dialect, irony, domain terms, and group differences. Define acceptable errors before expanding the corpus.
Limits
A sentiment score or topic cluster is not itself a validated social-science construct. Do not infer individual attitudes from model output without validation. This is guidance, not a tested performance claim.