I have achieved 93 % accuracy using RoBERTa for Ptacek et al(2014) dataset but for SemEval 2018 Task 3a it is 69% due to being a small dataset compared to Ptacek et al. However I would like to see if it is possible to achieve better accuracy. Mozafari et al(2019) proposed adding CNN after the BERT encoder and reported good accuracy on the hate speech detection task. My question is Do you have any proposition for text classification with transformers so that I can propose something new and get better accuracy at the same time?
Your answers will be appreciated.
Dear Nazmus Sakib
These papers seems to be useful, have a look:
1. https://link.springer.com/article/10.1007/s00521-020-05102-3
2. https://towardsdatascience.com/sarcasm-detection-with-nlp-cbff1723f69a
3. https://aclanthology.org/2020.figlang-1.11.pdf
4. Article Detecting sarcasm in multi-domain datasets using convolution...
5. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0252918
6. Article Multi-feature fusion framework for sarcasm identification on...
7. https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9383219
Kind Regards
Qamar Ul Islam
Thank you Qamar Ul Islam for the suggestions. So far i have found out that researchers used Bi-LSTM (Potamias et al, 2020)or a CNN (Mozafari et al,2019) after the transformer layer and reported good results. Text classification can be cross-domain. Although my research topic is Sarcasm detection, i am also reading papers in hate speech detection or fake news classification. So far i did not find any architecture that is concrete other than (potamias et al,2020). Should i be happy with 93% accuracy that i achieved using RoBERTa, or should i dig more to propose a novel architecture?
- Badges
- Science topic
- Similar topics
- Biological Science
- Bioecology
- Systems Ecology
- Landscape Ecology
- Connectivity