Skip to navigation

How can I prepare Weibo comments and Xiaohongshu notes for research without writing a collector?

You do not need a new collector: write down the research question, the selection rules and the unit of analysis first, then let a tool-connected assistant retrieve a bounded sample or organize your saved originals. You review the sample choices and the coding evidence yourself.

For example, to study help-seeking in tutorial comments, separate step questions, substitutions, experiences and corrections. Every row of the dataset must lead back to an original comment, not only to an AI theme summary.

Define four parts of the design

PartSpecify
QuestionThe expression or interaction being studied
Objects and timePlatforms, terms, formats and target period
Inclusion/exclusionRelevance, promotion, retelling, duplicates and missing context
UnitPost, comment or conversation, including multi-label rules

Start with a bounded batch to settle the rules before expanding. Selecting high-interaction content can be a sampling rule; record the rule, and do not describe the result as a platform-wide random sample.

Collect and preserve material

With verified links or content IDs, read the targets directly. With only a topic, search first and select manually. Use Weibo search, Weibo comments, Xiaohongshu selection and Xiaohongshu comments.

  1. Set the scope. Give the assistant the initial page count and the selection rules.
  2. Save every request. Keep the actual inputs, the responses, the request IDs and the collection intervals.
  3. Check dates separately. Compare each item’s publication time with the target period yourself. Asking for “last month” in natural language does not add a date filter that Weibo search does not declare.

Xiaohongshu has a field-extraction and excerpt-review workflow in the comment guide linked above. For actions without a recorded fixed output, such as Weibo’s, verify the actual structure and build the working table by hand first. Asking an AI to “export” the data does not confirm the fields and does not produce one shared format across platforms.

Which columns support coding?

Define your own research columns, and store identifiers as text:

ColumnPurpose
Research record IDCitation and internal tracing
Platform, original object/comment identity and sourceEstablish the object without using display names as IDs
Exact wording and necessary parent repliesPreserve expression and interaction
Publication and collection timesSeparate content age from observation
Selection state and reasonIncluded, excluded or needs material
Code and interpretationResearcher judgment separate from wording
Original location and request IDRevisit saved evidence

Leave missing values empty and keep a reported zero as zero. Preserve quotation marks and line breaks. After importing the CSV into your analysis tool, check that text identifiers, missing values and characters were not changed.

How do you ground coding?

  1. Define the labels. For example, define step question, substitution, experience, correction and other, and state when one record may carry several labels and when a record needs context.
  2. Code with reasons. Record the reason for each record’s code.
  3. Check context. “Would another pan work?” asks about conditions. “It worked this time” is an experience; whether the experience contains usable conditions depends on the context. “Same question” needs the parent comment to show what was asked.
  4. Review AI suggestions. An AI model can propose labels; the final table keeps the human review and the rule version.
  5. Deduplicate and count. Deduplicate by actual identifiers and keep the conversation relationships. Leave unresolved material out of the reviewed denominator, and do not estimate the valid sample size from requested counts.

For example, you coded 12 excerpts as “changed conditions,” then tightened the rule to require a named change. Review those 12 excerpts again, flag the unclear ones for follow-up and record the rule revision. When you report the share of reviewed excerpts in a category, use one rule version and state the actual number reviewed.

How do you document methods and quotations?

Fill the statement below with your actual records:

We studied [question] using [platforms and terms/object list], targeting [period] and collecting during [actual interval]. Under [selection rules], we reviewed [actual number and unit], preserving wording, sources and request records and applying [coding-rule version]. Unread or unavailable scope was [scope].

Dynamic platforms can return different results on a rerun, so keep the originals and the collection conditions next to the source links. In the paper, show research IDs and the context needed; keep the source relationships in your research records. Names and avatars that play no part in the analysis do not belong in the presentation table.

Prompt for the assistant:

Retrieve an initial batch under this question and selection plan. Preserve wording, parent relationships, times and sources, separating inclusion, exclusion and missing-material reasons. Propose codes for review rather than replacing originals with themes, and report actual sample scope.

Budget terms, platforms and actual pages with current pricing. Reuse saved originals, and expand the sample only after the rules are settled.

Use public content for industry observations

For an appliance category, record the questions in works and comments, verify the use conditions with product reviews, and review existing index material for topic changes. Keep the dates, sources, candidate scope and exclusion reasons, and record your hypotheses separately from the quotations. Repeated noise complaints can guide the next round of product research; market size, revenue and share need their own data.