Automated Complaint Analysis: Navigating Economic Activities with Machine Learning

    ## TL;DR
    The Portuguese Economic and Food Safety Authority (ASAE) receives over 20,000 complaints a year. Managing this manually is a resource drain. This paper introduces an NLP-driven system that automatically categorizes these complaints into economic sectors (e.g., Restoration, Retail, Industry) using a **Linear SVM**, achieving **81.6% accuracy** and a staggering **94.7% accuracy** when providing a top-3 recommendation list.

    ## The Administrative Bottleneck: Why This is Hard
    Governments are digitizing, but "digital" usually just means free-form text fields in contact forms. For ASAE, this results in:
    - **High Volume**: 20k+ complaints annually.
    - **Noisy Data**: Complaints include HTML tags, meta-data, mixed languages (Portuguese/English), or references to previous attachments.
    - **Class Imbalance**: Over 40% of complaints fall into "Restoration and Beverages," while others have almost zero data.
    - **Resource Constraints**: Portuguese is a "low-resource" language in the NLP world compared to English, making standard tools less reliable.

    ## Methodology: The Search for the Best Pipeline
    The authors didn't just pick a model; they conducted a comprehensive "battle royale" of NLP libraries and algorithms.

    ### 1. Preprocessing Excellence
    They compared **NLTK**, **spaCy**, and **StanfordNLP**. While NLTK showed slight numerical leads, the authors opted for **StanfordNLP** due to its superior Part-of-Speech (PoS) tagging and punctuation identification, which proved critical for cleaning official documents.

    ### 2. The Model Architecture
    The system uses a classical but robust pipeline:
    - **Feature Engineering**: TF-IDF (Term Frequency-Inverse Document Frequency) to weigh the importance of words.
    - **Decision Engine**: Evaluation of 8+ classifiers including Random Forests, Naive Bayes, and K-Nearest Neighbors.

    ![Model Performance Comparison](https://cdn.atominnolab.com/wisdoc/tables/20260527-48caa062-5bd8-436b-bfa6-c6cba40aeafd/page_004_block_006.png)
    *Table: Cross-comparison of NLP libraries and ML classifiers. Linear SVM consistently dominates.*

    ## Key Insights: When "Less" is "More"
    The paper provides several counter-intuitive findings that challenge common ML wisdom:
    1.  **LDA/PCA Failed**: Attempting to reduce dimensionality via Latent Dirichlet Allocation (LDA) or Principal Component Analysis (PCA) significantly *decreased* accuracy. The raw sparsity of TF-IDF was more informative.
    2.  **Balancing Hurdles**: Neither Random Over-Sampling (ROS) nor Random Under-Sampling (RUS) improved the SVM. The model naturally handled the distribution better without synthetic intervention.
    3.  **The Power of Rankings**: In an administrative context, a model doesn't need to be 100% right on the first try. By offering a "Top-3" list, the system becomes a high-reliability assistant for human operators.

    ![Top-K Accuracy Result](https://cdn.atominnolab.com/wisdoc/tables/20260527-48caa062-5bd8-436b-bfa6-c6cba40aeafd/page_006_block_003.png)
    *Table: The jump from Acc@1 (80.7%) to Acc@3 (94.7%) demonstrates the model's utility as a decision-support tool.*

    ## Error Analysis: The Human Element
    The authors performed a manual deep-dive into the failures. They found that the model often gets confused by:
    - **Semantic Overlap**: "Industry" vs. "Retail" often share the same vocabulary in a complaint about a faulty product.
    - **Empty Content**: Some "complaints" are just citizens asking for status updates on previous files, containing no sector-specific keywords.
    - **Human Error**: Interestingly, the ML model identified complaints where the *human* operators had originally assigned the wrong category.

    ## Conclusion & Future Outlook
    The study concludes that for Portuguese administrative text, **Linear SVM with TF-IDF** is the champion. It’s fast, explainable, and highly accurate. 

    **What's next?** The authors plan to move into the era of **Deep Learning** and **Word Embeddings** (like BERT or FastText). However, this paper serves as a vital benchmark: if a sophisticated Neural Network can't beat a 94.7% Top-3 accuracy, the classical approach remains the optimal choice for production-grade government tools.

    ### Takeaway for Practitioners
    If you are dealing with imbalanced, noisy, domain-specific text in a non-English language, don't rush to use a massive Transformer. A well-tuned SVM with a solid NLP preprocessing pipeline might just be your most cost-effective and accurate solution.