Multilingual Pornography Detection on Twitter: A Comparative Machine Learning Approach

    ## TL;DR
    This research addresses the critical challenge of identifying pornographic content on Twitter across multiple languages (Indonesian and English). By benchmarking **Decision Trees**, **Naive Bayes**, and **Support Vector Machines (SVM)** using a TF-IDF framework, the study reveals that while Naive Bayes is highly effective for Indonesian text (92.78% accuracy), SVM with a Linear Kernel is superior for handling the complexities of English and mixed-language environments.

    ## Problem & Motivation
    The proliferation of adult content on social media presents a significant risk to the morale and safety of minors. Prior works often relied on static **URL Blocking** or simple **Keyword Filtering**. However, Twitter poses a unique challenge:
    *   **Dynamic Language**: Natural language on Twitter is irregular, filled with slang, and frequently skips standard grammar.
    *   **Multilingualism**: Users often switch between Indonesian and English, complicating traditional monolingual filters.
    *   **Evasion**: Content creators use creative text variations to bypass simple keyword-based blocks.

    The authors' insight was to move beyond simple filters toward a **Machine Learning classification** approach that analyzes the statistical significance of words (Unigrams) within documents to differentiate between "Positive" (Safe) and "Negative" (Pornographic) content.

    ## Methodology
    The system follows a classic NLP pipeline optimized for the noisy nature of social media data:

    1.  **Pre-processing**: This involves aggressive cleaning (removing links, hashtags, and special characters) followed by **Porter Stemming** (adapted for Indonesian) to reduce words to their base forms.
    2.  **Feature Engineering**: The team utilized **TF-IDF (Term Frequency-Inverse Document Frequency)** to weight the importance of words. They focused on **Unigrams** (single-word features) like "sex" or "porn."
    3.  **Classification**: Three main algorithms were compared:
        *   **Decision Tree**: A flowchart-like structure for attribute testing.
        *   **Naive Bayes**: A probabilistic classifier based on the assumption of independence between variables.
        *   **Support Vector Machines (SVM)**: A high-dimensional mapping technique aimed at finding the optimal hyperplane for classification.

    ![System Architecture](https://cdn.atominnolab.com/wisdoc/images/20260603-fe8e25a9-15a8-4c9a-9d7a-a02639e26f5b/page_002_block_004.png)

    ## Experiments & Results
    The researchers used a dataset of 600 tweets (equally split between pornographic and non-pornographic across three categories: Indonesian, English, and Combined).

    ### Performance Breakdown
    The results highlighted a fascinating discrepancy between language performance:
    *   **Indonesian Dataset**: **Naive Bayes** dominated with an average accuracy of **92.78%**. The authors noted that Indonesian grammar, while complex, showed high regularity in the specific pornographic corpus used.
    *   **English Dataset**: **SVM** took the lead with **83.33%**.
    *   **Combined Dataset**: Performance dropped to **72.72%** (SVM), illustrating the difficulty of "code-switching" where users mix languages in a single post.

    ![Accuracy Comparison Table](https://cdn.atominnolab.com/wisdoc/tables/20260603-fe8e25a9-15a8-4c9a-9d7a-a02639e26f5b/page_004_block_005.png)

    ### The Kernel Selection (SVM)
    A deep dive into SVM kernels for the Indonesian dataset proved that the **Linear Kernel** is most effective for text classification (87.60% average). More complex kernels like Radial or Polynomial actually degraded performance, suggesting that text features are often linearly separable in a high-dimensional TF-IDF space.

    ## Critical Analysis & Conclusion
    ### Takeaway
    The study confirms that a "one-size-fits-all" algorithm is rarely optimal for multilingual NLP. **Naive Bayes** remains a powerful, lightweight tool for specific linguistic structures (like Indonesian), whereas **SVM** provides the necessary robustness for the more fragmented English-speaking "Twitter-verse."

    ### Limitations
    1.  **Dataset Size**: 600 tweets is relatively small for a modern machine learning task; larger datasets would likely reveal more edge-case failures.
    2.  **Context**: Unigrams cannot capture sarcasm or nuanced meanings that depend on word order (Bigrams/Trigrams).

    ### Future Outlook
    With the rise of Large Language Models (LLMs), the next step for this research would be to apply **transformers** to capture the semantic context of tweets, potentially closing the gap in the "Multilingual Combined" category where accuracy currently struggles.