dair-iitd / openie6

OpenIE6 system
GNU General Public License v3.0
119 stars 36 forks source link

Roadmap towards openie7 #2

Closed LifeIsStrange closed 3 years ago

LifeIsStrange commented 3 years ago

Context: I am a software engineer working on pushing the boundaries of natural language semantic parsing. To do such a task I critically need the state of the art of information extraction.

So I skimmed through your paper, it seems to have great ideas and seems to be the new state of the art. Do you think there is a new paper that outperform openie6?

You should probably update your paper by adding the results of predpatt https://github.com/hltcoe/PredPatt Predpatt seems like a paradigm shift in information extraction, by exploiting Universal Dependencies Predpatt didn't get much human resources and is now in maintenance mode but I do believe that a future openie7 could gain to use universal dependencies insights to some extent. Especially by making use of the Enhanced Universal Dependencies which predpatt does not make use of! So this is my main idea toward improving openie.

The second one can already be done today and has 100% chance of success: I believe that openie6 is making use of BERT, if that's the case then you should switch to XLnet instead which is the real state of the art pretrained language model. I'd advise the lib Transformers to do that https://github.com/huggingface/transformers It could yield significant accuracy improvments.

I and humanity really need better information extraction, please improve the SOTA once again!

@SaiKeshav

LifeIsStrange commented 3 years ago

https://huggingface.co/transformers/model_doc/xlnet.html

SaiKeshav commented 3 years ago

Thank you for your interest in our work and OpenIE in general. PredPatt is a useful reference that I was not aware of.

Currently, I am working on another problem. We will look into this whenever we get back to improving OpenIE.

Thank you for sharing your thoughts on this.

LifeIsStrange commented 3 years ago

@Saikeshav Good to know! Out of curiosity, what problem are you currently working on?

SaiKeshav commented 3 years ago

I am currently looking at Open Link Prediction, which is related to OpenIE. Reference: https://madoc.bib.uni-mannheim.de/55724/

LifeIsStrange commented 3 years ago

@Saikeshav very interesting, there is some topic overlap with what I currently work on :} If you wanna share some of your thoughts/strategies/difficulties, feel free !

LifeIsStrange commented 2 years ago

It's interesting to note that similarly to predpatt, the currently known state of the art in semantic role labeling is leveraging dependency parsing: https://paperswithcode.com/paper/semantic-role-labeling-as-dependency-parsing Any update on your research BTW? :)

LifeIsStrange commented 2 years ago

@SaiKeshav Hi, unrelated to my previous comments, I have a problem and I would like your opinion on the topic: Socrates is a man men are mortal Is Socrates mortal? : Yes


The car is red Red is a color Is the car a color? : No

Those syllogisms have the same structure. Yet the latter conclusion is false. Why is that ?