As far as I can tell, most NLP tasks today use word embeddings and recurrent networks or transformers.
Are there any examples of state-of-the-art NLP applications that are still n-gram based and use Naive Bayes?
$\begingroup$
$\endgroup$
3
-
1$\begingroup$ There exists a field in text forensics thats about identifying an author based on their writing style. Pan - pan.webis.de, is the most active conference about this topic. And as far as I know, each year the winner submissions tends to be handcrafted features that uses word and character n-grams amongst other handcrafted stylometric features, they are usually trained with SVMs or other simple methods. I believe the reason for this is that when the feature vectors are embeddings of the whole input, it takes time to learn style, it tends to learn topic instead, which is very misleading. $\endgroup$ – Isbister Sep 30 '20 at 15:56
-
1$\begingroup$ Also, have a look at this paper arxiv.org/pdf/2009.03116.pdf, here we evaluate different text representations on the task of sentence similarity, as you can see in Table 2, a svm with char n-grams are very competitive to state of the art if you consider the computation time and the slight drop in performance ;) $\endgroup$ – Isbister Sep 30 '20 at 16:00
-
$\begingroup$ @Isbister thanks! I'll take a look $\endgroup$ – chessprogrammer Sep 30 '20 at 18:41
Add a comment
|
