NLContextualEmbeddingResultMBS.averageToken as NLContextualEmbeddingResultVectorMBS

const lang = "en"
Var n As NLContextualEmbeddingMBS = NLContextualEmbeddingMBS.contextualEmbeddingWithLanguage(lang)

var e as NSErrorMBS
var r1 as NLContextualEmbeddingResultMBS = n.embeddingResultForString("Hello World", lang, e)
Var r2 As NLContextualEmbeddingResultMBS = n.embeddingResultForString("Hallo World", lang, e)

Var average1 As NLContextualEmbeddingResultVectorMBS = r1.averageToken
Var average2 As NLContextualEmbeddingResultVectorMBS = r2.averageToken

Var compare As Double = average1.cosineSimilarity(average2)

MessageBox compare.ToString // 0.93

The "Hello World" has six tokens. We sum them up and later divide by count to get the average.

NLContextualEmbeddingResultMBS.Constructor   Private

NLContextualEmbeddingResultMBS.tokenVectorAtIndex(characterIndex as Integer) as NLContextualEmbeddingResultVectorMBS

characterIndex: The index to get the token vector at.

NLContextualEmbeddingResultMBS.tokenVectors as NLContextualEmbeddingResultVectorMBS()

const lang = "en"
Var n As NLContextualEmbeddingMBS = NLContextualEmbeddingMBS.contextualEmbeddingWithLanguage(lang)

var e as NSErrorMBS
var r as NLContextualEmbeddingResultMBS = n.embeddingResultForString("Hello World", lang, e)

Var tokenVectors() As NLContextualEmbeddingResultVectorMBS = r.tokenVectors
Break // inspect array in debugger

Use this method to access the individual (subword) token embeddings. You can apply pooling or combination techniques to aggregate these subword vectors into a single representation for a word, phrase, or entire input.

Common pooling techniques include

• Mean pooling to take the average of subword vectors.
• Max pooling for finding the element-wise maximum across tokens.
• Use the embeddings of the first or last subword tokens to represent the entire input.