Abstract:
Sentiment analysis is often used in NLP to understand people’s subjective opinions. However, the analysis results may be biased if people use sarcasm in their statements. In
order to correctly understand people’s true intentions, being able to detect sarcasm is
critical. Detection of sarcasm is the crucial step in sentiment analysis due to the inherent
ambiguous nature of sarcasm. It also involves the complexity of language which makes
sarcasm detection much harder, even for humans too. Therefore, sarcasm detection has
gained importance in many Natural Language Processing applications. So, any progress
in sarcasm detection, is a positive step towards pushing the boundaries of Natural Language Processing.
Many previous models have been developed to detect sarcasm based on the utterances in
isolation, meaning only the reply text itself. Several researches have used both context
and reply texts to detect sarcasm in reply and showed great improvement in performance.
BERT, short for Bidirectional Encoder Representations from Transformers has been used
for this specific task of sarcasm detection before and it has shown surprisingly good results
for this task.
So, in this project, we aim to improve the accuracy of sarcasm detection by further
exploring the role of contextual information in detecting sarcasm. We will also investigate the performances of different BERT based models where we have utilized the BERT
embedding with different networks and compare the models for different data setups.