I'm working on a thing at the moment and I'm trying to do what's basically written on the title of this post.

Disclaimer: I'm not trying to get my job done by others. Just share ideas and gathering material (if any).

Do you know if there's some study/example/paper on something like that? I'm thinking about dependency trees and part-of-speech tagging to try not to focus on the train set, so that my model is not too "word" dependent. I mean, if I train it on some text, it won't work with something different while, maybe, the structure of agreement and disagreement may show similar patterns event with different subjects.

Let me know and thank you in advance.

  • Exactly what method are you going to use to train your models? In many cases it’s not necessary to POS tag the data. – Atamiri Mar 9 '20 at 14:31
  • Initially I was only dealing with the tfidfvectorizer, and use the tfidf matrix of title and headline as input for the model. I forgot to say that this is related to news. So if I train my model on past week news, this week terms may not the same used to train the model. From this comes the idea to try with dependency tree and/or POS tag (I think the first one could be better) – loricelli Mar 9 '20 at 14:57
  • OK, for TFIDF you do need the lemmata. How does the output look like ("stance")? Dependencies might be useful but they only capture syntactic relations while you may be interested primarily in semantics. – Atamiri Mar 9 '20 at 15:06
  • The labels are agree,disagree,unrelevant and discuss. I was also trying to get Named Entities and embeddings involved. So, my actual idea, is to have as input: [title NE, body NE, embeddings of title words, embeddings of body words]. The question is: could this be a reasonable approach? – loricelli Mar 9 '20 at 15:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.