MEANINGMATCH NOTES
How to read a sentence similarity score
A semantic signal, not a verdict.
Two sentences can share a meaning without sharing many words. An embedding model turns text into numerical vectors, and cosine similarity measures the direction those vectors have in common.
What the number means
Meaningmatch normalizes 384-dimensional vectors and takes their dot product. Scores are bounded from −1 to 1. Identical text should be very close to 1. Higher values suggest a closer meaning within this model, but a value of 0.8 is not an 80% probability of equivalence.
Use a small comparison set
Try one reference sentence and several alternatives. Inspect whether the closest match is useful, rather than treating a single universal threshold as correct. A dog running and a puppy running usually match more closely than a sentence about stock prices.
Watch for the important mistakes
“The system is safe” and “The system is not safe” can look close in embedding space despite a crucial opposite meaning. Names, numbers, dates and factual errors also need human review. The model does not establish plagiarism, truth, authorship or legal similarity.
Short English input
The chosen English model uses at most 256 wordpieces per input. A wordpiece is not the same as a word or character. Longer input is truncated. Use short sentences or short paragraphs; this tool is not a long-document comparison service.
Try the local similarity checker ↗