如何使用 deeplearning4java 训练 Word2vec model 后更新词汇表

Question

I used deeplearning4j library to do vectorization task using word2vec. I need to vectorize new word after training the model using a specific corpus. so how to add this new word and update the training to get a new weight vector for the new word? my code as folows:

Test word2Vec = new Test();
   word2Vec.train();
    
    //test the generated trained file
    Word2Vec word2VecModel = WordVectorSerializer.readWord2VecModel(new File(word2Vec.modelFilePath));
   
    
    double cosi=word2VecModel.similarity("httpdbpediaorgresourcethe_terminator", "httpdbpediaorgresourceterminator_salvation");
    System.out.println(cosi);

  }
  
  public  void train() throws IOException {
        SentenceIterator sentenceIterator = new FileSentenceIterator(new File(inputFilePath));

        TokenizerFactory tokenizerFactory = new DefaultTokenizerFactory();
        tokenizerFactory.setTokenPreProcessor(new CommonPreprocessor());

        Word2Vec vec = new Word2Vec.Builder()
                .layerSize(100)
                .windowSize(5)
                .epochs(5) //3-50 https://arxiv.org/pdf/1301.3781.pdf
                .elementsLearningAlgorithm(new SkipGram<VocabWord>())
                .iterate(sentenceIterator)
                .tokenizerFactory(tokenizerFactory)
                .build();
        vec.fit();
      
        WordVectorSerializer.writeWordVectors(vec, modelFilePath);
      
    }

Answer 1

Deeplearning4j is capable of updating the weights but won't add new words. Generally it's not recommended to do that. Adding only singular sentences doesn't really allow for good training of good word vectors.

The training doesn't normally take too long anyways. I would recommend just training the weights with new words.