Similarly, I was talking to someone and suggested we do a simple word embedding + LSTM as a baseline, but they suggested why not to just skip that and use only BERT as a baseline. Not wrong I suppose…. but also takes much longer to train/feedback.
-
-
-
TBH I actually think go straight to BERT and ignore simpler baselines is the right call for most tasks. Its super easy to set up with
or allennlp, pretty data efficient, and its not *that slow*, if you have >= 1 GPU - Još 5 drugih odgovora
Novi razgovor -
-
-
(We had a good laugh afterwards). Crazy to see how our methods have changed in just a short time.
Prikaži ovu nitHvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Presenter at a CL conference (PhD student) received the following question: Q: How did you preprocess your data, did you use lemmatisation? Presenter: ...? Lemmatisation? No, no, it could introduce bias in the data. O_o
-
It is interesting to see how people from different disciplines dive into NLP but miss basic notions of linguistics or any knowledge about anything before the BERT era. Can we fill this gap?
- Još 3 druga odgovora
Novi razgovor -
-
-
Full credit to him for asking, lots of people might not have said anything!
-
Absolutely
Kraj razgovora
Novi razgovor -
-
-
Turns out TFIDF works as good as any dep learning method on domains like medical text, for example. And you get this with almost zero training overhead and interpretable features.
- Kraj razgovora
Novi razgovor -
-
-
Seems the
#SustaiNLP workshop will be pretty exciting in this respect - Još 1 odgovor
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.