The disparity between the statement and its intended meaning is referred to as irony. Detecting this disparity in Arabic tweets is a challenging task. We presented three approaches namely deep learning using transformers, deep learning using Recurrent Neural Networks (RNN) and a features-based approach for detecting the irony in Arabic tweets. Among these approaches, the deep learning approach using transformers scores better F1-score of 0.816 than the deep learning approach using RNN which has 0.719 and the features-based approach which scores 0.709 on the IDAT@FIRE-2019 dataset.
Irony is a complex linguistic phenomenon widely studied in philosophy and
linguistics. Irony can be defined as an incompatibility between the literal meaning
and its conveyed meaning. Irony detection has gained relevance recently, due
to its importance in various NLP applications such as sentiment analysis, hate
speech detection, author profiling, fake news detection, and crisis management.
The irony is detected in various language namely Italian [
We propose two deep learning approaches and a features-based approach for detecting irony in Arabic tweets. 3.1
Deep learning approach using transformers architecture
In this method, Tensorflow based bi-directional transformers [
In general, BERT predicts whether the given two sentence are adjacent sentences (i.e. Is the next sentence are not). The architecture includes the parameters such as number of layers, hidden nodes and attention heads. For binary classification, BERT is fine tuned by adding extra parameters to a classification layer W–>(K*H), where ‘K’ is the number of classifier labels and ‘H’ is the number of final hidden states which is at the top. The input data for the BERT model is prepared by removing the extra lines and special characters. The preprocessed data is given to the model for training. The output of the classifier for the test data is submitted as SSN_NLP Run 1. 3.2
Deep learning approach using Recurrent Neural Network
architecture
In this method, sequence to sequence model [
Initially, the given training set is split into training and testing set and then
analysed. During that analysis, training set had 3500 instances and testing had
524 instances. The tweets along with their classes are given to the model for
1 https://github.com/tensorflow/nmt
training. The tweets are encoded into an intermediate representation which are
then filtered by deciding whether they are important or not by attention and
then decoded. Different models are constructed by varying the recurrent units
and attention. In our work, two attentions namely Scaled Luong [
Recurrent Unit
LSTM LSTM GRU GRU GNMT GNMT
Attention
Scaled Luong Normed Bahdanau
Scaled Luong Normed Bahdanau
Scaled Luong Normed Bahdanau From the Table 1, it is clear that the model with GRU as recurrent unit and Scaled Luong attention performs well than the other architectures and it is submitted as SSN_NLP Run 2. 3.3
Features-based Approach
In this method, the full corpus given are pre-processed and vectorized using
AraVec-1.0 [
Classifiers
After vectorization, every tweet is represented as a single vector by either considering the minimum vector of 300 dimension, maximum of 300 dimension and minimum-maximum vectors of 600 dimension. Then the vectors are classified using different machine learning classifiers namely Multi-Layer Perceptron(MLP), Support Vector Classifier, Decision Tree, KNN, Random Forest, Adaboost and Quadratic DA.
Different traditional classifiers are trained by considering various forms of vector representation. Three vector representations namely minimum, maximum and minimum-maximum are considered whose results are shown in the Table 2. From the table, MLP with minimum vector seems to perform better than other classifiers with different vector representation. Thus, it was submitted as SSN_NLP Run 3. 4
Approaches
Deep learning approach using transformers architecture
Deep learning approach using Recurrent Neural Network architecture
Features-based Approach
F1-score
0.816
0.793
0.709
Word embeddings have been used in the deep learning approach for detecting the
irony in tweets. The significant phrases or words present in the tweets of ironic
class may fall under the non-ironic class in the training set or vice versa, which
may lead to misclassification [
Irony detection in Arabic tweets have gained a good relevance nowadays in which the incompatibility between the comment and considered statement can Teams and Runs SSN_NLP Run 1 SSN_NLP Run 2 SSN_NLP Run 3
YOLO Run 1 chiyu_zhang_ubc Run 2
BENHA Run 3
rgcl Run1 Ali Allaith Run1
PITS Run1 Tha3aroon-ft Run1 kinmokusu Run1 Amrita_CEN Run1 be found. We have presented three different approaches in which a deep learning approach using transformers have achieved 0.816 as F1-score for the test data which is better than deep learning approach of using LSTM and features-based approach of MLP classifiers, with 0.793 and 0.709 as F1-scores respectively. The performance can further be improved by changing the value of parameters in transformers architecture.The probability scores of different deep learning architectures and traditional classifiers can also be ensembled to improve the performance.
We would like to thank the Science and Engineering Research Board (SERB), Department of Science and Technology for funding the GPU system (EEQ/2018/000262) where this work is being carried out.