Apply transformer models for Question Answering task

Data used

Data downloaded from here https://www.kaggle.com/c/tweet-sentiment-extraction

Problem

Given a tweet and sentiment extract the 'phrase' leading to a sentiment.

input = {'sentence': 'soo sad I missed you here in san diego', 'sentiment':'negative'} expected output = {'target':'soo sad'}

Evaluation

Jaccard similarity between phrases of target and prediction.

    a = set(str1.lower().split()) 
    b = set(str2.lower().split())
    c = a.intersection(b)
    return float(len(c)) / (len(a) + len(b) - len(c))

Approach

Straight forward way to approach this problem is to use a recurrent neural network on char-level emebedding to predict the characters which appear in output.

NER approach: Formulate this task as named entity recognition task. All the words in the output are marked as entities to be recognized from the input. Instead of characters we predict words here.

QA approach: We can formulate this problem Question answering problem with

1. sentiment as question.
2. text as context.
3. selected_text as answer

and apply BERT-based models like RoBERTa, ALBERT models. (Current state-of-art models)

Judging by the notebooks in the competition page, BERT-based models seem to outperform other approaches. Based

Model Implemented

• Apply ALBERT model and record cv-score
• Apply BERT model and record cv-score