irony-detection-machine-learning

URL: https://github.com/ShereenMamdouh/irony-detection-machine-learning

Description: It’s a Python code that works on 3 features (bag of words, semantic analysis, and word embedding) and 4 classifiers (SVM, K nearest neighbor, Naive Bayes, and decision trees) to classify tweets as ironic or not, supervised by a Twitter dataset.

Methods

• Features Used:
  • Bag of Words: Counts word frequencies in tweets to identify irony.
  • Semantic Analysis: Analyzes the sentiment polarity of the tweet, checking for contradictions between sentiments.
  • Word Embedding: Converts words into numerical vectors to identify patterns in ironic tweets.
• Classifiers Used:
  • Naive Bayes Classifier: A probabilistic classifier assuming independence between features.
  • SVM: A supervised learning algorithm aiming to increase the gap between data to separate two categories.
  • Random Forest: An algorithm generating multiple decision trees.
  • K-Nearest Neighbor (K-NN): Classifies tweets based on the k closest data points, with k=1 for best results.
• Text Preprocessing:
  • Tokenization: Splits sentences into words.
  • Lemmatization: Reduces words to their root forms.
  • Stop Words Removal: Removes irrelevant words to improve performance.
  • Data Filtering: Filters out unnecessary information from tweets.

Results

• Best Accuracy:
  • Bag of Words with Naive Bayes classifier yielded the best accuracy.
  • Semantic analysis with K-NN performed best.
  • Word embedding with Decision Trees gave the best performance.
• Overall Performance: The best accuracy achieved so far is approximately 70% using Naive Bayes with Bag of Words, given that the problem is mostly defined with 80% ironic tweets.
• Next Steps: The project is under implementation for further improving results by using Sentence2Vec, LSTM, and combining all features into one matrix for better classification performance.

Dataset

• Dataset Used:
  • The dataset used is from SemEval-2018 Task 3, consisting of 3842 tweets.
  • Distribution:
    • 604 tweets are classified as Non-Ironic.
    • 2396 tweets are classified as Ironic.
  • Format:
    • The dataset contains 3 attributes: Index, Class (Ironic or Non-Ironic), and Tweet.
• Data Processing:
  • The data is processed by splitting the dataset into vectors for Index, Labels, and Tweets.
  • Text preprocessing is applied, including filtering stop words, tokenization, and lemmatization, followed by feature extraction (bag of words, sentiment analysis, and word embedding).