I will be using BERT to get word embeddings before performing cosine similarity analysis on my data. According to this paper the accuracy of word embeddings can be improved by updating the model with domain specific textbooks. They do not provide any method to actually do that. Could anyone let me know the method? And how many textbooks might suffice to enrich the model - will 10-20 textbooks be fine?
Asked
Active
Viewed 28 times
