Spot the mood behind customer reviews with Hugging Face's free tool

Find the sentiment tool on Hugging Face

In your browser's address bar, type huggingface.co and press Enter. Once the page loads, look for the search bar at the top of the screen. Click inside it and type sentiment analysis, then press Enter. A list of AI models appears. Look for one near the top called distilbert-base-uncased-finetuned-sst-2-english (the name looks technical — that's normal). Click on that result.

What you'll see: a model page with a description and an interactive area. Somewhere on that page there should be a text box where you can type or paste text. Look for placeholder text like "Type something…" or "Input text here." Above or beside the box, you'll see a button to make the tool run.

What happens after: the page is ready to accept text. You don't need to read the rest of the page — just stay focused on the input box.

If you don't see a search bar at the top, look for a magnifying glass icon 🔍 instead. If the model page doesn't show an input box, scroll down — it's sometimes lower on the page than you'd expect.

Next step ↓

Sign in so the tool will accept longer text

Click the button on the tool page without typing anything in the box first. A small message usually appears saying something like "You must be logged in" or "Authentication required." That's normal. Look for a Login or Sign in option at the top-right corner of the page. Enter your email and password, or create a free account (it takes about a minute: email, password, confirm).

What you'll see: after signing in, the page reloads and your profile picture or initials appear in the top-right corner where the Login button was.

What happens next: the tool is now ready to process text from your account. Some browsers stay logged in across visits, so you may not see the prompt at all — just move on.

Next step ↓

Paste one review and run the tool

Open your text file of reviews. Copy just the first review — one or two sentences. Click into the text box on the model page and paste the review. Then click the button underneath the box to submit it.

What you'll see: a short wait (usually 1–3 seconds), then a result panel appears. It shows a label — usually POSITIVE or NEGATIVE in capital letters — followed by a percentage like 98.67%. That percentage is the model's confidence: how sure it is about its answer.

What happens next: the original review stays in the box and the result sits below it. You can now record the result somewhere and move on to the next review.

That's normal; it means the model is very sure. Just read the bigger number.

Anything below 80% confidence is a "soft" result — the model isn't sure, so that review needs a human read.

Next step ↓

Record the result somewhere you'll find it

This is where the time savings actually show up. Open a free tool — Google Sheets, Excel, or even a plain text file. For each review, write the review text in one column (or one line) and the model's label and confidence in the next column.

What you'll see: a growing list that looks like a table. As you go, you'll start to see patterns — most reviews positive, a clear cluster of negatives, a few mixed-feeling ones with low confidence.

What happens next: when you're done, you can sort the list by the label column. In Google Sheets, click the column header, then look for a sort option (the exact menu name varies — try Data → Sort range, or right-click the column and pick "Sort"). The negative reviews jump to the top automatically. That's your shortlist for follow-up.

Just write the review and the label next to it, like "Broken after a week — NEGATIVE 99%."

Next step ↓

Repeat for each review, then spot the soft results

Go back to Step 3 and process the next review: paste it, submit, copy the result into your table, move on. After you've run a batch — say, ten reviews — scroll through your table and look for any result where confidence is below 80%. Those are the ones the model wasn't sure about.

What you'll see: roughly 10–20% of your reviews fall into this "unsure" zone. That's expected.

What happens next: open each soft result and read it yourself. These often contain sarcasm ("Oh sure, great service"), mixed feelings ("Shipping was great but the product is awful"), or very short comments ("meh") that confuse simple models. You decide what to do with each one — positive, negative, or neutral — based on your own judgment. The tool saved you from re-reading the easy 80%; you only focus on the tricky ones.

The English-only model struggles with both. In that case, the tool isn't useful for this batch — try a different sample.