Universal Sentence Encoder

The Universal Sentence Encoder provides TensorFlow.js implementations for converting text into high-dimensional embeddings. It includes both the standard USE model that generates 512-dimensional embeddings for general text similarity and clustering tasks, and the USE QnA model that creates 100-dimensional embeddings specifically optimized for question-answering applications.

Package Information

• Package Name: @tensorflow-models/universal-sentence-encoder
• Package Type: npm
• Language: TypeScript
• Installation: npm install @tensorflow/tfjs @tensorflow-models/universal-sentence-encoder

Core Imports

import * as use from '@tensorflow-models/universal-sentence-encoder';

For CommonJS:

const use = require('@tensorflow-models/universal-sentence-encoder');

Basic Usage

import * as use from '@tensorflow-models/universal-sentence-encoder';

// Load the model
const model = await use.load();

// Embed sentences
const sentences = [
  'Hello.',
  'How are you?'
];

const embeddings = await model.embed(sentences);
// embeddings is a 2D tensor with shape [2, 512]
embeddings.print();

Architecture

Universal Sentence Encoder is built around several key components:

• Main USE Model: Generates 512-dimensional embeddings using the Transformer architecture
• USE QnA Model: Specialized 100-dimensional embeddings for question-answering tasks
• Tokenizer: SentencePiece tokenization with 8k word piece vocabulary using Trie data structure
• Model Loading: Supports custom model and vocabulary URLs for flexibility
• TensorFlow.js Integration: Built on tfjs-converter and tfjs-core for browser and Node.js compatibility

Capabilities

Standard Text Embeddings

Core Universal Sentence Encoder functionality for generating 512-dimensional embeddings from text. Ideal for semantic similarity, clustering, and general NLP tasks.

function load(config?: LoadConfig): Promise<UniversalSentenceEncoder>;

interface LoadConfig {
  modelUrl?: string;
  vocabUrl?: string;
}

class UniversalSentenceEncoder {
  embed(inputs: string[] | string): Promise<tf.Tensor2D>;
}

Standard Embeddings

Question-Answering Embeddings

Specialized Universal Sentence Encoder for question-answering applications, generating 100-dimensional embeddings optimized for matching questions with answers.

function loadQnA(): Promise<UniversalSentenceEncoderQnA>;

class UniversalSentenceEncoderQnA {
  embed(input: ModelInput): ModelOutput;
}

interface ModelInput {
  queries: string[];
  responses: string[];
  contexts?: string[];
}

interface ModelOutput {
  queryEmbedding: tf.Tensor;
  responseEmbedding: tf.Tensor;
}

Question-Answering

Text Tokenization

Independent tokenizer functionality using SentencePiece algorithm for converting text into token sequences. Can be used separately from the embedding models.

function loadTokenizer(pathToVocabulary?: string): Promise<Tokenizer>;
function loadVocabulary(pathToVocabulary: string): Promise<Vocabulary>;
function stringToChars(input: string): string[];

class Tokenizer {
  constructor(vocabulary: Vocabulary, reservedSymbolsCount?: number);
  encode(input: string): number[];
}

class Trie {
  constructor();
  insert(word: string, score: number, index: number): void;
  commonPrefixSearch(symbols: string[]): Array<[string[], number, number]>;
}

Tokenization

Types

// TensorFlow.js tensors
import * as tf from '@tensorflow/tfjs-core';

// Core types
type Vocabulary = Array<[string, number]>;

// Version information
const version: string;