AI ᱥᱮᱱᱴᱤᱢᱮᱱᱴ ᱮᱱᱟᱞᱟᱭᱤᱡᱚᱨ

What it tells you

Paste a piece of text, press Analyze, and the page reports whether it reads as positive or negative, with a score for each side. There is no neutral option: whichever score is higher becomes the headline, and the progress bar leans toward that side.

This is a polarity classifier, not a full sentiment system. It does not extract aspects, does not label emotions such as anger or joy, and does not explain which words drove the result. It answers one narrow question — does this text sound favourable or unfavourable? — and leaves the interpretation to you.

The model

The bundled model is a sentence-level polarity classifier in the SST-2 tradition: short English sentences, mostly film reviews, labelled positive or negative. That origin matters. It is good at opinions phrased in ordinary, evaluative language and weaker at factual text, where there is no opinion to detect at all.

The two scores normally add up to roughly one, because the underlying classifier assigns a probability to each class. The page shows both, but the large label and the bar follow only the higher one.

How the result is drawn

When positive wins, the panel shows a smiling icon, the word Positive in green, and a bar whose width is 50 percent plus half of the positive score. A 0.90 positive score therefore draws an 85 percent bar, not a 90 percent one. The bar is a visual lean away from the midpoint, not a direct plot of the score.

When negative wins, the same maths applies to the negative score, and the label turns red. If the scores are exactly equal, the code falls into the negative branch, so a perfect tie is displayed as Negative.

Below the bar, both percentages are printed side by side, labelled Negative and Positive. They are the real numbers; the bar is only a summary. The Clear button empties the textarea, resets the character count and hides the result.

Step by step

1. Paste or type text into the box. The character counter updates as you type.
2. The Analyze button stays disabled until there are at least three characters.
3. On the first run, the browser downloads the text runtime and the BERT model. The panel names each stage.
4. Press Analyze Sentiment. The result replaces the empty state.
5. Edit the text and analyze again to compare; the model stays loaded.

The Load AI Model button preloads everything if you want to avoid the first-run wait. Once loaded it changes to a green "Model Ready" state.

Reading the scores honestly

A score is the classifier's confidence in a label, not a measured intensity of feeling. 0.95 positive means the model is confident the sentence belongs to the positive class; it does not mean the sentence is extremely positive. A mildly favourable review and an ecstatic one can both land near 0.99.

Small differences are noise. A result of 0.52 positive against 0.48 negative is close to a coin flip, and rewording a single word can flip it. Treat anything in the 0.45 to 0.55 band as unresolved rather than as a weak positive.

Long text is a special case. The model has a fixed input window, so a long passage is effectively truncated or summarised internally. A review that praises five features and criticises one may be classified on whichever part dominates the opening, and the displayed score will not warn you that most of the text was ignored.

Where it works

• Short opinion sentences. "The battery lasts all day" or "The app crashes constantly" are exactly the shape the model was trained on.
• Review triage. Sorting a pile of comments into broadly positive and broadly negative before reading them.
• Quick tone checks. Sanity-checking a headline, a tagline or a customer reply.
• Teaching. A compact demonstration of how a text classifier turns words into two probabilities.
• Filtering. Flagging likely complaints in a support inbox for a human to review.

Where it fails

Sarcasm is the classic failure. "Great, another delay" contains positive vocabulary and negative meaning, and the model often reads the words rather than the intent. Irony, backhanded compliments and rhetorical questions are similarly unreliable.

ᱵᱟᱝ ᱞᱮᱠᱷᱟᱜ ᱫᱚ ᱵᱟᱝ ᱥᱚᱢᱟᱱ ᱛᱮ ᱪᱟᱞᱟᱜ ᱠᱟᱱᱟ᱾ "ᱵᱟᱝ ᱠᱷᱟᱨᱟᱯ" ᱟᱨ "ᱵᱟᱝ ᱱᱟᱯᱟᱭ" ᱟᱹᱲᱟᱹ ᱛᱮ ᱥᱩᱨ ᱥᱩᱯᱩᱨ ᱨᱮ ᱢᱮᱱᱟᱜᱼᱟ ᱠᱤᱱ, ᱢᱮᱱᱠᱷᱟᱱ ᱢᱮᱱᱮᱛ ᱛᱮ ᱵᱤᱯᱨᱤᱛ᱾ ᱟᱨ ᱢᱤᱫ ᱵᱟᱝ ᱞᱮᱠᱷᱟᱜ ᱫᱚ ᱱᱟᱥᱮ ᱡᱤᱞᱤᱧ ᱟᱹᱭᱟᱹᱛ ᱨᱮ ᱟᱫ ᱫᱟᱲᱮᱭᱟᱜᱼᱟ᱾ ᱢᱮᱥᱟᱞ ᱯᱷᱤᱞᱤᱝ ᱦᱚᱸ ᱥᱚᱢᱟᱱ ᱜᱚᱫᱚᱜ ᱠᱟᱱᱟ: ᱢᱤᱫ ᱯᱮᱨᱮᱜᱨᱟᱯᱷ ᱡᱟᱦᱟᱸ ᱨᱮ ᱯᱨᱚᱰᱟᱠᱴ ᱨᱮᱱᱟᱜ ᱥᱟᱨᱦᱟᱶ ᱟᱨ ᱰᱮᱞᱤᱵᱷᱟᱨᱤ ᱨᱮᱱᱟᱜ ᱱᱤᱱᱫᱟᱹ ᱢᱮᱱᱟᱜᱼᱟ, ᱚᱱᱟ ᱢᱤᱫ ᱞᱮᱵᱮᱞ ᱦᱩᱭᱩᱜ ᱠᱟᱱᱟ, ᱵᱟᱱᱟᱨ ᱢᱚᱛ ᱢᱮᱱᱟᱜ ᱠᱟᱱᱟ ᱢᱮᱱᱛᱮ ᱵᱟᱝ ᱵᱟᱰᱟᱭᱚᱜᱼᱟ᱾

ᱥᱟᱹᱨᱤᱛᱮ ᱟᱨ ᱱᱤᱴᱨᱟᱞ ᱚᱱᱚᱞ ᱫᱚ ᱵᱟᱭᱱᱟᱨᱤ ᱨᱮ ᱛᱮᱭᱟᱨ ᱦᱩᱭᱩᱜ ᱠᱟᱱᱟ᱾ ᱢᱤᱫ ᱥᱯᱮᱥᱤᱯᱷᱤᱠᱮᱥᱟᱱ ᱥᱤᱴ, ᱢᱤᱫ ᱱᱤᱣᱡᱽ ᱦᱮᱰᱞᱟᱭᱤᱱ ᱥᱮ ᱢᱤᱫ ᱟᱭᱤᱱᱤ ᱠᱞᱚᱡᱽ ᱨᱮ ᱯᱚᱞᱟᱨᱤᱴᱤ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱤᱠᱷᱟᱹᱱ ᱠᱞᱟᱥᱤᱯᱷᱟᱭᱟᱨ ᱫᱚ ᱟᱹᱨᱤᱛᱮ ᱢᱤᱫ ᱡᱤᱛᱠᱟᱹᱨᱤᱭᱟᱹ ᱮ ᱨᱩᱣᱟᱹᱲ ᱮᱢᱟ᱾ ᱚᱱᱟ ᱚᱨᱡᱚ ᱠᱚ ᱢᱤᱫ ᱦᱟᱞᱠᱟ ᱥᱮᱦᱮᱫ ᱞᱮᱠᱟᱛᱮ ᱵᱟᱝ, ᱢᱮᱱᱮᱛ ᱵᱟᱹᱱᱩᱜ ᱞᱮᱠᱟᱛᱮ ᱢᱟᱱᱟᱣ ᱢᱮ᱾

ᱰᱚᱢᱮᱱ ᱥᱤᱯᱷᱴ ᱫᱚ ᱢᱟᱱᱮ ᱫᱚᱭ ᱫᱚᱦᱚᱭᱟ᱾ ᱢᱚᱰᱮᱞ ᱫᱚ ᱯᱷᱤᱞᱢ-ᱨᱤᱵᱷᱤᱭᱩ ᱯᱟᱹᱨᱥᱤ ᱠᱷᱚᱱ ᱥᱮᱬᱟᱣ ᱟᱠᱟᱱᱟ; ᱯᱨᱚᱰᱟᱠᱴ ᱨᱤᱵᱷᱤᱭᱩ, ᱪᱟᱴ ᱢᱮᱥᱮᱡᱽ, ᱴᱣᱤᱴ, ᱠᱞᱤᱱᱤᱠᱟᱞ ᱱᱚᱴ ᱟᱨ ᱵᱟᱝ-ᱤᱝᱞᱤᱥ ᱚᱱᱚᱞ ᱡᱚᱛᱚ ᱵᱷᱮᱜᱟᱨ ᱞᱮᱠᱟᱛᱮ ᱪᱟᱞᱟᱜ ᱠᱟᱱᱟ᱾ ᱤᱢᱚᱡᱤ, ᱥᱞᱮᱝ ᱟᱨ ᱠᱷᱟᱴᱚᱬ ᱫᱚ ᱱᱚᱶᱟ ᱞᱟᱹᱜᱤᱫ ᱟᱹᱰᱤ ᱜᱮ ᱵᱟᱝ ᱧᱮᱞᱚᱜᱼᱟ᱾

ᱡᱟᱦᱟᱸ ᱫᱚ ᱵᱟᱭ ᱠᱟᱹᱢᱤᱭᱟ

ᱱᱤᱴᱨᱟᱞ ᱠᱞᱟᱥ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱤᱢᱚᱥᱚᱱ ᱞᱮᱵᱮᱞ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱟᱥᱯᱮᱠᱴ-ᱵᱮᱥᱰ ᱥᱮᱱᱴᱤᱢᱮᱱᱴ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱯᱟᱨ-ᱥᱮᱱᱴᱮᱱᱥ ᱵᱨᱮᱠᱰᱟᱣᱩᱱ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱠᱤᱣᱟᱨᱰ ᱦᱟᱭᱞᱟᱭᱤᱴ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱵᱤᱵᱨᱚᱬ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱯᱟᱹᱨᱥᱤ ᱪᱤᱱᱦᱟᱹ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱛᱚᱨᱡᱚᱢᱟ ᱵᱟᱹᱱᱩᱜᱼᱟ ᱟᱨ ᱵᱮᱪ ᱮᱱᱟᱞᱟᱭᱥᱤᱥ ᱵᱟᱹᱱᱩᱜᱼᱟ᱾ ᱱᱚᱶᱟ ᱫᱚ ᱢᱤᱫ ᱥᱠᱮᱞ ᱨᱮ ᱛᱤᱵᱽᱨᱚᱛᱟ ᱵᱟᱭ ᱥᱠᱚᱨ ᱮᱢᱟ, ᱟᱨ ᱢᱤᱫ ᱚᱱᱚᱞ ᱪᱮᱫ ᱞᱮᱠᱟᱛᱮ ᱞᱮᱵᱮᱞ ᱦᱩᱭᱮᱱᱟ ᱚᱱᱟ ᱦᱚᱸ ᱵᱟᱭ ᱞᱟᱹᱭ ᱮᱢᱟ᱾

ᱯᱨᱟᱭᱵᱷᱮᱥᱤ

ᱚᱱᱚᱞ ᱫᱚ ᱣᱮᱵᱽᱮᱥᱮᱢᱵᱽᱞᱤ ᱨᱮ ᱞᱚᱠᱟᱞ ᱞᱮᱠᱟᱛᱮ ᱯᱨᱚᱥᱮᱥ ᱦᱩᱭᱩᱜ ᱠᱟᱱᱟ᱾ ᱥᱟᱨᱵᱷᱟᱨ ᱛᱮ ᱪᱮᱫ ᱦᱚᱸ ᱵᱟᱝ ᱠᱩᱞ ᱦᱩᱭᱩᱜᱼᱟ, ᱟᱨ ᱯᱮᱡᱽ ᱵᱚᱸᱫᱽ ᱠᱟᱛᱮ ᱛᱟᱭᱚᱢ ᱪᱮᱫ ᱦᱚᱸ ᱵᱟᱝ ᱫᱚᱦᱚ ᱦᱩᱭᱩᱜᱼᱟ᱾ ᱨᱟᱱᱴᱟᱭᱤᱢ ᱟᱨ ᱢᱚᱰᱮᱞ ᱫᱚ ᱯᱩᱭᱞᱩ ᱵᱮᱵᱷᱟᱨ ᱨᱮ ᱥᱟᱫᱷᱟᱨᱚᱱ CDN ᱠᱷᱚᱱ ᱰᱟᱣᱩᱱᱞᱚᱰ ᱦᱩᱭᱩᱜ ᱠᱟᱱᱟ, ᱡᱟᱦᱟᱸ ᱫᱚ ᱚᱱᱟ ᱩᱫᱩᱜᱟ ᱡᱮ ᱴᱩᱞ ᱫᱚ ᱠᱷᱚᱞᱟᱣ ᱞᱮᱱᱟ ᱢᱮᱱᱠᱷᱟᱱ ᱟᱢ ᱪᱮᱫ ᱴᱟᱭᱤᱯ ᱠᱮᱫᱟ ᱚᱱᱟ ᱵᱟᱝ᱾ ᱚᱱᱟ ᱫᱚ ᱟᱢᱟᱜ ᱱᱤᱡᱮᱨᱟᱜ ᱰᱮᱴᱟ ᱱᱤᱭᱚᱢ ᱞᱮᱠᱟᱛᱮ ᱵᱷᱤᱛᱨᱤ ᱱᱚᱴ ᱟᱨ ᱠᱚᱥᱴᱚᱢᱟᱨ ᱢᱮᱥᱮᱡᱽ ᱞᱟᱹᱜᱤᱫ ᱵᱮᱵᱷᱟᱨ ᱫᱟᱲᱮᱭᱟᱜᱼᱟ᱾

ᱥᱤᱢᱟᱹᱠᱚ

ᱢᱤᱫ ᱫᱷᱟᱣ ᱢᱤᱫ ᱴᱚᱴᱷᱟ ᱚᱱᱚᱞ, ᱯᱷᱟᱭᱤᱞ ᱟᱯᱞᱚᱰ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱵᱮᱪ ᱢᱳᱰ ᱵᱟᱹᱱᱩᱜᱼᱟ, ᱯᱟᱹᱨᱥᱤ ᱵᱟᱪᱷᱟᱣ ᱵᱟᱹᱱᱩᱜᱼᱟ ᱟᱨ ᱥᱟᱡᱟᱣ ᱫᱟᱲᱮᱭᱟᱜ ᱛᱷᱨᱮᱥᱦᱳᱞᱰ ᱵᱟᱹᱱᱩᱜᱼᱟ᱾ ᱢᱚᱰᱮᱞ ᱨᱮᱱᱟᱜ ᱤᱱᱯᱩᱴ ᱣᱤᱱᱰᱳ ᱫᱚ ᱢᱤᱫ ᱡᱤᱞᱤᱧ ᱰᱚᱠᱩᱢᱮᱱᱴ ᱠᱷᱚᱱ ᱛᱤᱱᱟᱹᱜ ᱜᱟᱱ ᱥᱟᱹᱨᱤᱛᱮ ᱵᱤᱪᱟᱹᱨ ᱦᱩᱭᱩᱜᱼᱟ ᱚᱱᱟ ᱥᱤᱢᱟᱹ ᱫᱚᱭ ᱫᱚᱦᱚᱭᱟ᱾ ᱯᱩᱭᱞᱩ ᱞᱳᱰ ᱨᱮ ᱣᱮᱵᱽᱮᱥᱮᱢᱵᱽᱞᱤ ᱟᱨ ᱱᱮᱴᱣᱟᱨᱠ ᱮᱠᱥᱮᱥ ᱥᱟᱶ ᱢᱤᱫ ᱟᱫᱷᱩᱱᱤᱠ ᱵᱽᱨᱟᱣᱡᱟᱨ ᱞᱟᱹᱠᱛᱤᱜᱼᱟ᱾

ᱥᱟᱫᱷᱟᱨᱚᱱ ᱵᱷᱩᱞᱠᱚ

• ᱢᱤᱫ ᱱᱤᱴᱨᱟᱞ ᱚᱨᱡᱚ ᱮᱥᱯᱮᱠᱴ ᱠᱟᱜ ᱾ ᱢᱚᱰᱮᱞ ᱨᱮ ᱥᱩᱢᱩᱝ ᱵᱟᱨᱭᱟ ᱠᱞᱟᱥ ᱢᱮᱱᱟᱜᱼᱟ ᱟᱨ ᱟᱹᱨᱤᱛᱮ ᱢᱤᱫᱴᱟᱹᱝ ᱮ ᱵᱟᱪᱷᱟᱣᱟ᱾
• ᱥᱠᱚᱨ ᱫᱚ ᱛᱤᱵᱽᱨᱚᱛᱟ ᱞᱮᱠᱟᱛᱮ ᱢᱟᱱᱟᱣ ᱾ ᱱᱚᱶᱟ ᱫᱚ ᱢᱤᱫ ᱞᱮᱵᱮᱞ ᱨᱮ ᱯᱟᱹᱛᱭᱟᱹᱣ, ᱵᱩᱡᱷᱟᱹᱣ ᱨᱮᱱᱟᱜ ᱫᱟᱲᱮ ᱫᱚ ᱵᱟᱝ᱾
• ᱢᱤᱫ ᱯᱩᱨᱟᱹ ᱞᱮᱠᱷᱟᱱ ᱮᱱᱟᱞᱟᱭᱤᱥ ᱠᱟᱜ ᱾ ᱡᱤᱞᱤᱧ ᱚᱱᱚᱞ ᱫᱚ ᱠᱷᱟᱴᱚ ᱦᱩᱭᱩᱜ ᱠᱟᱱᱟ, ᱟᱨ ᱚᱨᱡᱚ ᱫᱚ ᱚᱱᱟ ᱨᱮᱱᱟᱜ ᱥᱩᱢᱩᱝ ᱢᱤᱫ ᱦᱟᱹᱴᱤᱧ ᱜᱮ ᱩᱫᱩᱜ ᱫᱟᱲᱮᱭᱟᱜᱼᱟ᱾
• ᱵᱷᱮᱸᱜᱟ ᱪᱤᱱᱦᱟᱹ ᱨᱮ ᱯᱟᱹᱛᱭᱟᱹᱣ ᱾ ᱵᱤᱫᱽᱨᱩᱯ ᱫᱚ ᱟᱹᱨᱤᱛᱮ ᱥᱚᱫᱚᱨᱚᱜ ᱯᱚᱞᱟᱨᱤᱴᱤ ᱮ ᱩᱞᱴᱟᱹᱣᱟ᱾
• ᱥᱟᱹᱨᱤᱛᱮ ᱚᱱᱚᱞ ᱨᱮ ᱱᱚᱶᱟ ᱵᱮᱵᱷᱟᱨ ᱾ ᱥᱯᱮᱥᱤᱯᱷᱤᱠᱮᱥᱟᱱ ᱟᱨ ᱱᱤᱣᱡᱽ ᱨᱮ ᱞᱮᱵᱮᱞ ᱞᱟᱹᱜᱤᱫ ᱪᱮᱫ ᱢᱚᱛ ᱦᱚᱸ ᱵᱟᱹᱱᱩᱜᱼᱟ᱾
• ᱕᱒ ᱯᱟᱨᱥᱮᱱᱴ ᱚᱨᱡᱚ ᱢᱤᱫ ᱧᱟᱢ ᱞᱮᱠᱟᱛᱮ ᱯᱟᱲᱦᱟᱣ ᱾ ᱥᱩᱨ ᱥᱩᱯᱩᱨ ᱥᱚᱢᱟᱱ ᱫᱚ ᱵᱟᱝ ᱛᱷᱤᱨ ᱟᱨ ᱟᱞᱜᱟᱛᱮ ᱩᱞᱴᱟᱹᱣ ᱫᱟᱲᱮᱭᱟᱜᱼᱟ᱾

ᱰᱷᱮᱨ ᱠᱩᱠᱞᱤ ᱟᱠᱟᱱ ᱠᱩᱠᱞᱤᱠᱚ