Questions about purity model inference notebook

hlab-repo / purity-and-danger

Explorations of frame overlap in viral discourse and xenophobic discourse.

MIT License

0 stars 0 forks source link

Questions about purity model inference notebook #5

Closed textpotential closed 2 years ago

textpotential commented 3 years ago

What is BERT doing with the hashtags with some split tokens?
"with" seems to be metaphorical often?
@jobarber , can you clarify what you think is not working in your model?
- do you feel like it is overfitting?
- the model seems to do pretty well differentiating literal from metaphorical use of infect - we tested with some combinations around "my sister getting an infection" or "my sister infecting the house with negativity"
"disease" seems to not get read as metaphor very often - is this an indication that our fine tuning corpus needs more metaphorical language?
does the absolute value of score as it compares to other words in sentence matter or just whether it is near 1?

textpotential commented 3 years ago

We took a look at the corpus used for fine tuning and I think what you are saying @jobarber is that in this metaphor corpus (vuamc), the uses of viral language are not often metaphorical.

We tried this sample

TERRORISM is A VIRUS which can SPREAD and INFECT THE BODY of AMERICA, a virus against the invasion of which PROPHYLACTIC MEASURES have to be taken and which, once identified, has to be EXPELLED from the body politic

And virus and infect were not flagged as metaphor.