How to pass input to deep learning models for Multiple choice question answering problem?

Question

I'm currently working on a multiple-choice question answering system. The training set consists of a question, answer and 4 options and I need to predict the correct answer among 4 options. Sometimes there is one paragraph too, For example :

1.Which among the following is measured using a Vernier Caliper?

[A] Dimensions
[B] Time
[C] Sound
[D] Temperature

Answer : A [Dimensions]

Chapter text: [Book chapter related to Dimension, time, sound and temperature ]

How to feed this input to any of deep learning models? I thought two approaches :

1. Using tokens

and correct and as one hot encoding => [1, 0, 0, 0 ]

2. Using concatenation

Generating fix sized word embedding for each text :

 - Chapter text = [1,1024] 
 - Text         = [1,1024] 
 - option_a     = [1,1024] 
 - option_b     = [1,1024] 
 - option_c     = [1,1024] 
 - option_d     = [1,1024]

final_input = concat( [ Chapter text, Text, option_a, option_b, option_c, option_d] ) ==> [1,6144]

and correct and as one hot encoding => [1, 0, 0, 0 ]

Is it good representation for understanding and reasoning over text for mcqa task?

Answer 1

The basic idea of most of the current question answering architectures is:

• get a common representation of the question and the input text (e.g., using BERT)
• get a representation of the answers
• do sort of attention over the answers: compute a scalar score for each of the answers (using a dot-product, linear layer, multilayer-perceptron) and normalize the scores using softmax

The architectures are typically based on the Bidirectional Attention Flow Model, although it was designed for a slightly different task and although the pre-trained word embeddings and RNNs in the model are today usually replaced with BERT-like models. In 2018, there was a competition in question answering at SemEval where many interesting ideas on this problem were presented.