search

DOsinga / deep_learning_cookbook

Deep Learning Cookbox
Apache License 2.0
689 stars 338 forks source link

Chapter 6 Model Goal #14

Closed WillKoehrsen closed 6 years ago

WillKoehrsen commented 6 years ago

In the code for Chapter 6, it seems to me that the model is trying to learn how to associate a title of a question with the body and not with the answer. Is this the correct interpretation? I'm a little confused because there are numerous places in the text where it looks like it's stated the model is learning to associate the questions and answers, such as here:

As in the previous recipe, we are going to train our model to distinguish between a question title and the correct answer (body) versus the answer to another random question.

This is somewhat confusing because in most cases, the body is itself a question. I think the terms title and body should be used instead of question and answer.

It seems like it would be possible to associate the questions with the accepted answers since these are also included in the dataset.

DOsinga commented 6 years ago

Yeah, this is sloppily formulated. Here's the full(er) story. Originally we wanted to create a question answer model like you are suggesting. But we couldn't quite make it work with a model the size of a notebook (or even beyond that). So we settled for something less ambitious, finding similar questions. We do this by learning the association of a title and a body, which gives you an embedding space for the titles and now you can search for nearest neighbors. Sorry about the confusion!

WillKoehrsen commented 6 years ago

No problem! I enjoyed the example and thought it to be useful, just was a little confused by the wording. Thanks for explaining the full backstory (maybe you could add the explanation into the book) .

It would be an interesting project to try and take on that question and answer model!