How to obtain word embedding vector for each word in the sentence using pre-trained BERT in MATLAB

Hello,

I have a question on How to obtain word embedding vector for each word in the sentence using pre-trained BERT in MATLAB. I successfully loaded bert and tokenized the words in the sentence, but I didn't find example code in MathWork website to get each word's embedding vector, like word2vec.

[net,tokenizer] = bert;
str = "Bidirectional Encoder Representations from Transformers";
words = wordTokenize(tokenizer,str)

% Then what...?

I would thank you if anyone can help this.

Answer 1

[Ganesh]

Hi @Yuhong,

I understand that you want to generate Word Embeddings for BERT Model using MATLAB. To achieve this, you can use the "encode()" function, implemented similar to your own implementation.

[net,tokenizer] = bert;
str = "Bidirectional Encoder Representations from Transformers";
words = encode(tokenizer,str);

In case of BERT, "embedding" and "encoding" can be used interchangeably. Further, you can use the "decode()" function to decode the "encodings".

Kindly refer to the documentation below to know more on these functions:

"encode()" function - https://in.mathworks.com/help/textanalytics/ref/berttokenizer.encode.html

"decode()" function - https://in.mathworks.com/help/textanalytics/ref/berttokenizer.decode.html

Kindly note that using "bert" model in MATLAB requires the Text Analytics Toolbox.

Hope this helps