Analyze Coupang and Olive Young Reviews Without Knowing Korean

Korean e-commerce reviews are long and full of slang, and a single review often mixes several complaints. For teams that don't read Korean, even after machine translation, you still have to read every review one by one to extract the information — feasible on small samples, but efficiency collapses once the review volume grows. A more practical approach: scrape the reviews, tag each one with structured English labels, then count and chart the labels rather than translating the text.

Core Idea: Label Counting Instead of Review-by-Review Translation

The logic of this method isn't complicated. First, use a review scraping tool to export the dataset, then have the classifier tag each review across three dimensions: complaint type, sentiment, and purchase motive. The labels are English key-value pairs with accompanying probability values, so you can filter, sort, and aggregate directly.

According to sample data published by the tool provider, the output looks like this: a review saying "it stopped charging in under a month, and no one responded to my exchange request" would be tagged quality_defect (0.95) and customer_service (0.92), with sentiment negative; a review saying "delivery took a week and the box was torn to shreds" would be tagged packaging (0.97) and delivery (0.96). There are nine complaint categories: delivery, quality/defect, size/fit, price/value, customer service, packaging, no effect, skin problems, and other. Each category has both an English key and a corresponding Korean name (e.g., quality_defect / 품질·불량), so the same table can be shown directly to Korean colleagues.

Purchase motives cover five categories: price/discount, reviews and word of mouth, brand, gifting, and repurchase. The value of this labeling system lies in cross-comparability: you can put your own product and competitors side by side and compare differences in complaint composition without understanding a single word of the original text.

A Five-Minute Setup Process

The first step is exporting reviews. Simply run the review scraper for Coupang or Olive Young on the Apify Store — the original article names the Coupang Review Scraper (zen-studio), Olive Young Review Scraper (cocomoon, Korean site), and Olive Young Review Scraper (solidcode, Olive Young Global).

The second step is opening the Korean Review Classifier tool. It doesn't handle scraping — only tagging. The text, rating, and ID fields are detected automatically, and you can also define up to 10 custom yes/no labels, in either Chinese or English. The official configuration example is to pass in a dataset ID, specify additional text fields (such as title), and add two custom labels, e.g., "mentions the smell" and "says they will buy again".

The third step is filtering and counting. Sort by complaint type to see the real reasons behind one-star reviews, or group by product to compare the complaint structure of your own brand versus competitors. The tool can also be paired with Apify Schedules to automatically run against new reviews every day.

For those who prefer to operate in natural language, you can use the Apify MCP server to let Claude, Cursor, or Claude Code call it directly — for example, "classify the dataset abc123 from this run of the Coupang scraper and summarize the complaints".

The Apify Platform and the Actor Mechanism

Apify, which appears repeatedly throughout the article, is a cloud-based web scraping and automation platform. Its core concept is the "Actor" — a packaged, reusable scraping or processing program. Users don't need to write any code; they simply select and run one in the Apify Store. Both the review scrapers and the Korean Review Classifier are Actors. Each Actor run produces a "dataset" containing structured results organized by field, along with a unique dataset ID. This is precisely why the classifier can read the previous scraping results directly by passing in a dataset ID, forming a "scrape → tag → filter and count" pipeline. Apify Schedules is the platform's built-in scheduled task feature, enabling the entire pipeline to run automatically on a daily basis; the MCP server is the external interface layer that lets AI clients supporting the MCP protocol — such as Claude and Cursor — call these Actors directly.

Cost and Accuracy

Pricing is $0.50 per 1,000 reviews, with no prompt writing and no need for your own LLM key. Ten thousand reviews cost $5, and Apify's monthly $5 credit roughly covers ten thousand reviews. On speed, about 100 reviews take 5 seconds; rows without review text (such as certain summary rows included by some scrapers) are skipped and not billed.

On accuracy, the provider says that on 100 human-annotated Korean reviews (from Coupang, Naver Shopping, and Naver Place, of which 47 were 1–4 stars), complaint type match rate was 88% and sentiment match rate was 96%. The tool is built and tested for Korean, and can also handle English reviews. That said, automated labels will always have errors, and the provider also advises sampling and verifying before making major decisions.

What the Classification Metrics Mean

The article's "88% complaint type match rate and 96% sentiment match rate" are evaluation metrics for classification tasks. Match rate usually refers to the proportion of model-predicted labels that agree with human-annotated labels, roughly equivalent to accuracy. But in multi-label scenarios, note the following: a single review may belong to multiple complaint types at once. If evaluation judges each label individually for whether it was hit, that's closer to precision/recall; if it only checks whether the "primary label was guessed correctly", it is insensitive to missed labels. Moreover, an overall match rate of 88% does not mean every category is equally reliable — categories with small sample sizes (such as skin problems) will fluctuate more. That's why the provider recommends sampling and verifying before major decisions: essentially, using manual spot checks to estimate error in real-world scenarios and avoid extrapolating an optimistic number from a single test directly onto your own data.

Three Alternatives and Their Limitations

Machine translation followed by manual reading: fine for a few dozen reviews, but once the volume grows you're still reading line by line, and translation flattens the slang that carries the complaint semantics.

Star ratings only: fast, but a four-star review may still mention skin problems, and star ratings can't tell you whether the problem lies in delivery or the product itself.

Writing prompts for a general-purpose chatbot: it can summarize a batch of pasted reviews, but you have to maintain the prompt yourself and then convert the free text back into a table; whereas the classification tool returns the same set of labels with probabilities on every run.

It should be noted that this is an independent tool and is not affiliated with Coupang, CJ Olive Young, NAVER Corp., or the scraper authors mentioned above. For teams doing business in the Korean market but lacking Korean-language staff, turning "reading reviews" into "counting labels" is the lowest-cost path available today.

One-line Summary

Tag Korean e-commerce reviews with structured English labels instead of translating them one by one, enabling low-cost batch analysis.

Paragraph Summary

The article addresses the pain points of Korean e-commerce reviews being long, slang-heavy, and mixing multiple complaints, and proposes using structured English labels instead of review-by-review translation. Specifically, first export a dataset using the Coupang or Olive Young review scraper on Apify, then use the Korean Review Classifier to tag each review across three dimensions — complaint type, sentiment, and purchase motive — with probability values, enabling direct filtering, sorting, and aggregation. There are nine complaint categories (delivery, quality/defect, size, price, customer service, packaging, no effect, skin problems, other) and five purchase motives, with labels provided in both English keys and Korean names. Cost is $0.50 per 1,000 reviews, with no prompts or own LLM key required, and about 5 seconds per 100 reviews; the provider's tests show an 88% complaint type match rate and 96% sentiment match rate. The article also compares the limitations of three alternatives: machine translation, star-rating statistics, and general-purpose chatbots.