Online reviews are market signals, not opinions. Reading them rigorously takes a real analytical stack, and the barrier was never the statistics: it was the tooling. Running the analysis through Google Colab and Gemini removes the coding barrier, so the attention goes where it belongs, to why a method exists, how to read its output, and what the result does and does not license.
This page is the chapter by chapter map of the book by Jeong-Yeol Park, Ph.D., written for graduate students, researchers, and working professionals who need rigorous quantitative analysis of review data without a coding background. Between the parts are three analyses you can run today, on your own exported reviews or on the public companion file. The program page has the pitch and the instructor companion; this page has the curriculum.
Four chapters and one lab, before a single estimate. They set what a review actually measures, what the platform already selected away, the working environment you will use for the rest of the book, and the vocabulary every later chapter assumes. The part ends with an analysis you run end to end, so the first model you meet arrives after you have already handled data yourself. The chapter one lesson is free to watch now, with no account.
Export your reviews as a CSV, or take the public companion file from the book's data repository. Open a Google Colab notebook with Gemini enabled, upload the file, and paste the request below. Nothing to install, nothing to configure. This is a chapter 3 environment doing chapter 2 work: before you model anything, find out what you are holding.
Companion file and code: github.com/jeong-yeol/AOR_with_AI. Maple City v3 is 11,010 synthetic reviews across 200 restaurants, 2019 to 2022, and every number printed in the book is computed from it.
1. Check the row count against what you exported. A gap means the platform, the export, or the load dropped something, and you want to know which before anything else.
2. Look at reviews per business. If the median business carries a handful of reviews, no method in Part II will rescue that, and the honest move is to change the unit of analysis rather than the estimator.
3. Read the date range as a history of the platform, not of the market. Does your window cover the change you care about, with enough on both sides of it to see a before and an after?
4. It named columns it would not trust. Take one and decide for yourself: is it merely noisy, or is it measuring something other than what its name promises? Those two problems have different fixes.
Nothing on this page shows what any of these requests returned, and that is deliberate. The output that teaches you anything is the one your own file produces; a printed answer would turn a piece of work back into a piece of reading.
Every methods chapter holds one shape. WHY the method exists, and what question it answers that the others cannot. How to INTERPRET the output, line by line, including the parts that look like results and are not. How to APPLY the finding, in a paper and in a decision. The shape is the point: it is what lets a reader who has never estimated a model before arrive at a defensible reading of one.
Same notebook, same file. This is the chapter 5 move that every later chapter depends on, run as an audit rather than a demonstration. You are not looking for a good score. You are looking for the cases where the score and the star rating disagree, because that is where you find out what your new variable is measuring.
1. Read the 15 disagreements and label them yourself before you accept the score. If most of them turn on sarcasm, negation, or a mixed review with one loud complaint, your variable is tracking wording rather than evaluation.
2. Correlation with the star rating is a sanity check, not a validation. Very near one means you have rebuilt the star rating slowly; near zero means at least one of the two is not measuring what you assumed.
3. Find the thin business-months. Every panel, time-series, and spatial method in Part II inherits them, so decide now whether to aggregate up, restrict the sample, or narrow the question.
4. Say what the number measures in one sentence without using the word sentiment. If you cannot, you have a column, not yet a construct, and the chapters after this one will amplify that gap rather than close it.
Real problems do not arrive labeled with a method. The last two chapters put the methods back together on complete questions, then look at where review analytics is going and how to judge a new capability before adopting it. The second of those matters more than it sounds: most of what arrives next will be easy to run and hard to validate.
This is the chapter 13 discipline in its smallest form. The request deliberately asks for labels and nothing else: no summary, no accuracy figure. The scoring is yours to do, because a measure you did not check is a measure you cannot report.
1. Label the same 100 yourself first, ideally with a second person, and settle your own disagreements before you look at the machine labels. How often two humans disagree is the ceiling on what any classifier can be scored against.
2. Compare row by row and count the two kinds of error separately. A single overall figure hides the direction of failure, and the direction is what determines whether your downstream estimate is biased up or down.
3. Reread the cases where you and the machine disagreed. If your category was ambiguous on those, fix the category definition and rerun, rather than blaming the classifier.
4. Write the sentence you would put in a manuscript, with the sample size and the agreement statistic in it. If that sentence needs a hedge, the hedge belongs in the manuscript too.
Analyzing Online Reviews with AI: A No-Code Guide from Raw Data to Research Findings, by Jeong-Yeol Park, Ph.D. Every number in the book is computed from a single companion corpus, Maple City v3: 11,010 synthetic reviews across 200 restaurants, 2019 to 2022. One dataset across thirteen chapters means a reader can rerun any example, compare against what is printed, and see exactly where their own run diverged.
The corpus is public, so nothing here has to be taken on trust. Download it, work the chapters in the order your own question needs, and check the book against itself.
Available as paperback and Kindle. Data and code: github.com/jeong-yeol/AOR_with_AI.
The mini-project is the entry point for every reader, whether the output is headed for a manuscript or for Monday. After that the order is yours: text measurement first, then whichever of panel regression, causal design, forecasting, spatial diagnostics, theory testing, or prediction your data can actually support. The mini-project is a free video lesson, and the program page carries it in written form alongside the practice files and the instructor companion for programs adopting the book as a methods text.