Xiaohongshu Sentiment Analysis: Find the Objections Behind the Comments

Xiaohongshu sentiment analysis is most useful to a brand when it identifies what people are evaluating, why they hesitate and what evidence would resolve the uncertainty. A positive or negative label alone cannot tell a product team whether to change a feature, explain a use condition or investigate an unverified complaint.

Start with the objection behind the wording. Then separate the topic, the author's relationship to the product and the confidence of your interpretation. This guide provides a practical coding worksheet for international brand managers and research teams. Every Chinese comment below is invented for illustration; none was collected from Xiaohongshu, and none represents a measured customer preference.

Define the decision before scoring the comments

Write one question that the analysis could help answer. For a hypothetical travel-accessory brand, it might be: “What should we investigate before describing a compact umbrella as convenient for commuters?” That is more actionable than asking whether people like umbrellas.

Choose aspects that relate to the decision: carrying weight, packed length, drying, opening mechanism, durability or price. Allow new aspects to emerge during review rather than forcing every comment into the initial list. Separate sentiment toward an aspect from sentiment toward the product overall.

This distinction has an established research basis. The ASAP paper, published at NAACL in 2021, treats aspect-category sentiment and overall review rating as related but distinct tasks in Chinese restaurant reviews. That supports the analytical distinction, not a claim that a model trained on restaurant reviews will perform reliably on Xiaohongshu umbrella comments.

For selecting material in a genuine study, use the broader Xiaohongshu market research workflow. This article focuses on interpretation after an appropriately scoped evidence set exists.

Keep source type separate from sentiment

A question about breakage is not a report of breakage. A creator's recommendation is not independently verified ownership. A statement of personal use remains a claim unless supporting evidence establishes more.

Use separate fields for communicative function and sentiment. Suggested functions are question, claimed experience, recommendation or promotion, comparison, irony candidate and unclear. These can overlap: a promotional message may also contain a negative comparison. Record a disclosed commercial relationship separately; do not diagnose an undisclosed advertisement solely from enthusiastic language.

For sentiment, allow positive, negative, mixed, no evaluative position and unresolved. “No evaluative position” means the text does not express a judgment on the selected aspect. “Unresolved” means the available context does not support a confident interpretation. Combining those categories as “neutral” hides a useful distinction.

An annotated Chinese comment worksheet

The following comments and translations are entirely illustrative. They were written for this guide, are not quotations or customer findings, and cannot be counted as research observations. The proposed interpretations are examples of a coding process, not authoritative readings of every similar expression.

The first and fourth rows concern a similar use situation but carry different evidential weight. One asks whether something could happen; the other expresses a complaint in the invented wording. Neither proves a defect or its frequency. The second row also shows why assigning one label to a whole comment can erase the trade-off a buyer is describing.

Read negation, irony and context before translating the verdict

Keep the Chinese text beside the English translation and the surrounding material needed to understand it. Record what a pronoun refers to, whether the comment replies to another person, and whether the criticism concerns delivery, a competing product or the featured item.

Xu and Huang's 2016 study of Chinese polarity shifting explains how linguistic patterns can reverse, weaken or cancel an apparent sentiment. The practical lesson is to review the whole expression rather than count positive and negative words. The qualified weight example above illustrates this editorial recommendation; it is not copied from that study.

Likewise, the 2022 NAACL Findings paper on sarcasm recognition examines conflict between literal and implied sentiment. Its findings motivate caution about literal readings, not a platform-specific accuracy promise. When context remains missing, retain uncertainty instead of forcing a confident English verdict.

Ask a fluent Chinese reviewer to check decision-relevant ambiguous items. Save both the proposed reading and the reason for changing it. Translation software can assist that review, but its fluency does not validate the interpretation.

Build a codebook another researcher can use

Define the unit of analysis: a comment, a sentence or an aspect within a comment. If one comment has two aspect labels, do not later describe those labels as two people. Keep post and reply identifiers so repeated text, cross-posts and multiple replies do not silently become independent testimony.

For each retained unit, record the source reference, relevant context, original wording, translation, function, aspect, sentiment, commercial context, uncertainty and reviewer rationale. Keep missing purchase verification explicitly unknown. Avoid inferring age, income, location or other personal characteristics from a profile.

Trial the codebook on varied material, including questions, mixed judgments and ambiguous language. Have two reviewers independently code a shared subset, compare disagreements, revise definitions and recheck affected items. Disclose how much received this review. Do not invent an agreement statistic or claim that this procedure guarantees accuracy.

AAPOR's transparency guidance provides a useful benchmark for disclosing content selection, units, coding and AI assistance. Applying those principles makes the analysis inspectable; it does not certify the study or make search-selected comments representative of Chinese consumers.

Use automation with an explicit validation boundary

If a tool proposes labels, preserve its model or version, instructions, input fields and review process. Evaluate it against human-reviewed examples from the actual category and intended task. An unrelated benchmark cannot establish its accuracy for your study.

Inspect errors by type, especially when an incorrect label could alter the business recommendation. A tool that turns questions into complaints may inflate an apparent product problem. A tool that collapses mixed comments into positive ones may conceal a trade-off. Keep unresolved cases visible and distinguish machine-proposed labels from final reviewed decisions.

Any summary count needs a denominator and exclusions. State whether it counts comments, unique authors or aspect mentions, and disclose how missing context and promotional material were handled. A share of retained comments is a description of that selected evidence set, not a percentage of customers. Likes, sentiment and repeated wording do not establish sales, market share or defect rates.

Turn an objection into a testable decision

Return to the hypothetical umbrella. If a future, genuine study retained questions about wet storage, a bounded interpretation could be that this scenario deserves investigation. It would not establish a widespread unmet need.

The product team could verify dimensions and storage instructions, then ask eligible interview participants to walk through what they do with a wet umbrella. If that work supports a communication test, compare clear explanations of the storage arrangement. Do not promise a sales improvement from the comment labels.

Keep observation, interpretation and next action separate in the handoff. Include contrary material and explain what would change the recommendation. Our consumer research methods comparison helps select follow-up evidence; the research-report checklist explains how to preserve the reasoning in a deliverable.

Agree access, retention and reuse arrangements before handling real material. The Association of Internet Researchers' ethics resources offer context-sensitive guidance. Public visibility alone should not determine whether identifying quotations belong in a commercial report; minimize unnecessary personal information and assess potential harm from reuse.

For a source-linked English brief, tell Whatson.red Intelligence your category, intended customer and decision. We can discuss a focused scope that distinguishes documented objections, uncertain interpretations and the evidence needed before acting.