Search pairwise ranking implementation

[kmalik22]

Summary:
Implements a pairwise ranking model for search.

Training Datum: {search query, positive response, negative response}
Loss function: pairwise ranking hinge loss: https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/35662.pdf

Differential Revision: D13166103

[hikushalhere]

@kmalik22 Thanks a lot for the PR. It looks like you want a two-tower model that can represent a query and a document and learns by using pairwise ranking loss. Have you checked out PairClassificationTask: https://github.com/facebookresearch/pytext/blob/master/pytext/task/tasks.py#L162? I wonder if you can just add a new output layer that implements the pairwise loss (that you have provided a link to) and then add it to PairClassificationTask as an option for its output_layer. Pls let me know what you think.

[kmalik22]

@hikushalhere, thanks for your comment.
I did look at PairClassification, it solves a different problem. For PairClassification, the inputs are {label, doc1, doc2}. The goal is to classify the two documents as a 'pair'. Eg: the documents "agree", or "contradict", or neither agree nor contradict.

For SearchPairwiseRanking, the input is three documents: {query, response1, response2}. We want to find which response is better for the query in question. In this case (unlike PairClassification), the query is a document and needs to go through an embedding-->representation(LSTM, eg)-->MLP layers.

Fundamentally, PairClassification is solving a classification problem, computing something like similarity(doc1, doc2) and mapping that to a label (eg., via a Softmax layer)
SearchPairwiseRanking is solving a relevance ranking problem, computing s1=similarity(query, doc1); s2=similarity(query, doc2), and then using f(s1, s2) as a loss function.

Hope this makes sense!