Understanding the scores

Every call to analyze() returns a Result with a label and four numbers. This page explains where they come from.

$result = Sentiment::analyze('VADER is smart, handsome, and funny.');

$result->compound; // 0.8316
$result->positive; // 0.746
$result->neutral;  // 0.254
$result->negative; // 0.0
$result->label;    // Label::Positive

Compound

compound is the one number to use when you want a single score. It runs from -1 (extremely negative) to +1 (extremely positive), rounded to four decimals.

It is computed in three steps:

1. Every word found in the lexicon contributes its valence (-4 to +4), adjusted by the rules below.
2. The valences are summed.
3. The sum is normalized so it always lands between -1 and +1. Long, strongly worded text approaches the limits but never passes them.

Because of that last step, a compound of 0.5 does not mean “50% positive”. Read it as a position on a scale: close to 0 is neutral, close to ±1 is emphatic.

Positive, neutral and negative

positive, neutral and negative are ratios between 0 and 1, rounded to three decimals. They show how the text splits between positive signal, negative signal and everything else, and they add up to about 1.

For VADER is smart, handsome, and funny., 74.6% of the signal is positive, none is negative, and the remaining 25.4% is neutral (words with no sentiment, such as VADER and is).

These ratios are not probabilities. A text that is mostly neutral words with one mildly positive word can have a high neutral and still a clearly positive compound. Use them to see how a text is built, and use compound (or the label) to decide.

Label

The Label is derived from compound and the threshold:

The default threshold is 0.05, the value recommended by the VADER authors.

What changes a score

The model does not just add up word values. These rules shape the result:

For example, these English texts are in increasing order of positivity: not good, good, very good, VERY good!!!.

Indonesian has the same rules plus intensifiers that come after the word (bagus banget), contrast words such as tapi and namun, and informal negations such as gak. See Languages.

Examples

Tips

• Classify with the label, rank with compound.
• Empty or whitespace-only text gives all zeros and a neutral label.
• Text with no known words is neutral. The model only knows what its lexicon knows; add your own words for domain slang.
• Long text is better analyzed sentence by sentence. See Long text.