Thank you for your appreciation of my work: the dataset consists of 900k questions (link: https://www.kaggle.com/c/quora-question-pairs/data).

One way to use the trained model for duplicate detection can be as follows: consider you have a database of questions, and you want to check whether any newly added questions has already been asked. For that matter, you can come up with the following processing steps:

1) Use your model to generate an embedding of each question of your DB using the encoder part of the network I present in this article.

2) Everytime a new question is added, you generate its embedding representation.

3) Then, on this representation, you can run an algorithm such as k-NN to retrieve the k closest questions already in your DB.

4) Finally, for each of these questions, you apply the distance classifier trained alongside of the encoder.

You can then find wether or not your newly added question is a duplicate of any existing question. Hope I was helpful :)