Long text

The model was designed for short, informal text: a sentence or a few. On long text the normalized compound score drifts toward ±1 as more words pile up, so a long review with a few strong words can look more extreme than it is, and a mixed article can hide its structure.

For anything longer than a short paragraph, analyze it per sentence and average.

Per-sentence averaging

use Risan\Sentiment\Analyzer;

function averageCompound(Analyzer $analyzer, string $text): float
{
    $sentences = preg_split('/(?<=[.!?])\s+/u', trim($text), -1, PREG_SPLIT_NO_EMPTY);

    if ($sentences === false || $sentences === []) {
        return 0.0;
    }

    $total = 0.0;

    foreach ($sentences as $sentence) {
        $total += $analyzer->analyze($sentence)->compound;
    }

    return $total / count($sentences);
}

$review = 'The room was spotless. The staff were lovely! Breakfast was cold, though.';

averageCompound(new Analyzer(), $review);

The average is a float from -1 to 1, so you can compare it with the same thresholds as a normal compound score.

Why it helps

• Each sentence gets its own negation, contrast and emphasis handling, so a not in one sentence cannot reach into the next.
• Every sentence has equal weight, so one long angry paragraph does not drown out five calm ones.
• You get the individual scores for free. Keep them to show which sentences drive the result.

Weighting and ranking

A plain average treats Thanks. and a detailed complaint the same. If that matters, weight by length, drop very short sentences, or pick the most extreme sentences:

$scores = array_map(
    fn (string $sentence): float => $analyzer->analyze($sentence)->compound,
    $sentences,
);

$worst = min($scores);
$best = max($scores);

A note on splitting

Sentence splitting is a text problem of its own. The regular expression above is fine for typical prose, but it will split after abbreviations such as Dr. and e.g.. If your text needs better splitting, use IntlBreakIterator from the intl extension or a splitter that suits your language, and pass the sentences to the analyzer.