I've updated the example 08_Sentiment_Classification_Using_Word_Vectors.knwf, replacing the deprecated old nodes Word Vector Learner and Vocabulary extractor with the new ones. The accuracy with the new ones is much worse, I don't understand why. I upload the new example.
What are the differences in the new Doc2Vec that can impact accuracy? The new one has no Basic Token Preprocessing, and has two Sequence Learning Algorithm to choose from.
In the new workflow I've left the old nodes too, so that by just changing a connection you can compare the two versions.
I did some tests and found out the difference in accuracy most likely results from the added 'sampling' parameters. You could try to set both to 0.0 (the default value before). Generally, there are no new parameters compared to the old version. We just always used the DL4J default values. Now they have just been made configurable. Regarding default values in KNIME, I took the values from https://radimrehurek.com/gensim/models/doc2vec.html (you'll find further explanation of the parameters there too).
Furthermore, you need to enable Hierarchical Softmax. This will disable the 'negative sampling' field, however, there is a bug that the value will be used nevertheless. Therefore, set the value to 0.0 and then check the Hierarchical Softmax box. We'll fix that issue.