Skip to main content

๐ŸŽฏ Outputs

The likelihood of a word represents how likely this word is given previous words, according to the model. For instance, the likelihood of Paris in "The capital of France is Paris" is very high, whereas the likelihood of "London" would be lower, and that of "book" would be even smaller.

The log-probability is a representation of the likelihood, ranging from minus infinity to zero. Words with log-probability close to zero have high likelihood, whereas words with large negative log-probability (e.g. -10, -50, etc.) are more unlikely. Log-probabilities are also called log-probs or scores as an abbreviation.

They are useful because log-probabilities can simply be added to evaluate the log-probability of a combination of words. For instance, the log-prob of "New York" in the sentence "I love New York" is the log-prob of "New" in "I love New" plus the log-prob of "York" in "I love New York". This can be used to score entire sentences, and evaluate multiple pre-defined options according to their likelihood.

See our entries about โค๏ธ Reviews Classification or ๐Ÿ”Ž๏ธ Content Marketing and Search Engine Optimization with Muse to learn more. For the French version, you can visit โค๏ธ Classification de Critiques.

Embeddingsโ€‹

Status: Alpha

Embeddings are a numerical representation of a given text, built by the model internally to make predictions. They are a vector (i.e. list) of numbers that encodes information about the input text, its context, as well as general knowledge derived from the training data in a computer-understandable format. They can be used by machine learning algorithms as a representation of the input text, to compare different sentences and documents, classify samples, or cluster texts.