Sarcasm is a term that refers to the use of words to mock, irritate, or amuse someone. It is commonly used on social media. The metaphorical and creative nature of sarcasm presents a significant difficulty for sentiment analysis systems based on affective computing. The technique and results of our team, UTNLP, in the SemEval-2022 shared task 6 on sarcasm detection are presented in this paper. We put different models, and data augmentation approaches to the test and report on which one works best. The tests begin with fundamental machine learning models and progress to transformer-based and attention-based models. We employed data augmentation based on data mutation and data generation. Using RoBERTa and mutation-based data augmentation, our best approach achieved an F1-score of 0.38 in the competition’s evaluation phase. After the competition, we fixed our model’s flaws and achieved an F1-score of 0.414.

All the models are available in Models directory. In addition to models reported in the paper, we have implented XLNet and Electra models, which are available in in the models directory.

To run the models, first make sure to install the requirements using below command:

pip install -r requirements.txt

Then run each models by using calling python for the model name like:

python ./Models/model-name/model-name.py

We mostly used the Isarcasm (Oprea and Magdy, 2019) dataset in this study. In specific experiments, we integrated the primary dataset with various secondary datasets, including the Sarcasm Headlines Dataset (Misra and Arora, 2019) and Sentiment140 dataset (Go et al., 2009) to increase the quantity of data and compensate for the lack of sarcastic data.

Data directory includes our dataset which are in the following fomat:

|
____ Data
    |
    ___ Cleaned Dataset/
    |
    ___ Foreign Dataset/
    |
    ___ Main Dataset/
    |
    ___ Mutant Dataset/
    |
    ___ Train_Dataset.csv
    |
    ___ Test_Dataset.csv

Cleaned Dataset includes main dataset with preprocessing. Foregin Dataset includes our secondary datasets like Sentiment 140, they available in this link. Main Dataset is the task train dataset. Mutant Dataset is our mutated version of main task dataset. Each dataset name has a three bit in it. First bit shows using word elimination. Second bit is for word replacement and third one is for shuffling. Train_Dataset.csv is the concatation of main task dataset and mutant100 which we find it the best dataset for the models. Test_Dataset.csv is the task test dataset with labels.

Here is the dataset mapping:

In this directoy, we have our data augmentor models.

Generative Pre-trained Transformer 2 (GPT-2) is an open-source artificial intelligence created by OpenAI in February 2019. GPT-2 translates text, answers questions, summarizes passages, and generates text output on a level that, while sometimes indistinguishable from that of humans, can become repetitive or nonsensical when generating long passages. It is a general-purpose learner; it was not specifically trained to do any of these tasks, and its ability to perform them is an extension of its general ability to accurately synthesize the next item in an arbitrary sequence. GPT-2 was created as a "direct scale-up" of OpenAI's 2018 GPT model, with a ten-fold increase in both its parameter count and the size of its training dataset.

We used three distinct ways to change the data in Mutation-based data augmentation: eliminating, replacing with synonyms, and shuffling. These processes were used in the following order: shuffling, deleting, and replacing. The removal and replacement were carried out systematically. We used the words’ roots to create a synonym dictionary. When a term was chosen to be swapped with its synonyms, we chose one of the synonyms uniformly at random(Figure 1). We tried each combination of these processes to find the best data augmentation combination (a total of seven).

Here is the code which we used for preprocessing. We have checked different methods like Stemming, Lemmaitization, Link removation, and etc.

Includes our task paper. For more information, we suggest you to read this paper.

F1-score and accuracy for different data augmentation methods on SVM model with BERT word embedding and no preprocessing.

Best results for each model using iSarcasm dataset and mutation-based data augmentation.

We want to convey our heartfelt gratitude to Prof. Yadollah Yaghoobzadeh from the University of Tehran, who provided us with invaluable advice during our research. We would also like to thank Ali Edalat from the University of Tehran, who provided us with the initial proposals for resolving the dataset’s imbalance problem.

@inproceedings{abaskohi-etal-2022-utnlp,
    title = "{UTNLP} at {S}em{E}val-2022 Task 6: A Comparative Analysis of Sarcasm Detection using generative-based and mutation-based data augmentation",
    author = "Abaskohi, Amirhossein  and
      Rasouli, Arash  and
      Zeraati, Tanin  and
      Bahrak, Behnam",
    booktitle = "Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022)",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.semeval-1.135",
    pages = "962--969",
    abstract = "Sarcasm is a term that refers to the use of words to mock, irritate, or amuse someone. It is commonly used on social media. The metaphorical and creative nature of sarcasm presents a significant difficulty for sentiment analysis systems based on affective computing. The methodology and results of our team, UTNLP, in the SemEval-2022 shared task 6 on sarcasm detection are presented in this paper. We put different models, and data augmentation approaches to the test and report on which one works best. The tests begin with traditional machine learning models and progress to transformer-based and attention-based models. We employed data augmentation based on data mutation and data generation. Using RoBERTa and mutation-based data augmentation, our best approach achieved an F1-score of 0.38 in the competition{'}s evaluation phase. After the competition, we fixed our model{'}s flaws and achieved anF1-score of 0.414.",
}