What is the current state-of-the-art NLP Question Answering model that has no need to receive [passage, question] as inputs, receiving only [question] and delivering the [answer] in unstructured text?
In my researches, I just have found [passage, question] -> model -> [answer], like R-NET (Microsoft) and QANet (Google). I would like to find something like: [question] -> model -> [answer]. Is there anything like this?
Thanks!
https://arxiv.org/pdf/1704.00051.pdf
It's not state-of-the-art. But it does include a passage retrieval step. So you have [question] > model > [passage, question] > [answer]
Asterios Chardalias, as you said, it includes a passage retrieval. It would be great if it was not necessary, but it is a start. I will try to understand and use something of the paper ideas. Thank you!
You'd have to feed the model input somehow. Whatever its knowledge base may be, it has to look up for its answers somewhere. Even if you throw an inference machine in the mix, that could combine inputs from disparate sources, you'd still need to define where to retrieve them from. And if we're talking NLP, these sources will be textual by definition. So, no matter how deep 'under the hood' the passages are, they need to be there to begin with.
- Badges
- Science topic