Sentiment Analysis Pipeline

Overview

An NLP pipeline for automated sentiment classification of text into positive, neutral, and negative categories, built as a focused exercise in text preprocessing, feature representation, and evaluation methodology.

Scientific Context

Sentiment classification is a well-studied problem, which is exactly what makes it useful as a testbed: with the modelling question mostly settled, it becomes a place to be rigorous about the parts that are easy to shortcut: data cleaning, evaluation design, and honestly characterizing where a model fails, rather than just reporting an aggregate accuracy number.

Problem Statement

How much of a sentiment classifier’s performance comes from the model versus from the quality of preprocessing and feature representation upstream of it? And where, specifically, does a reasonably simple pipeline still get things wrong?

Objectives

• Build a complete pipeline from raw text to classified sentiment, treating each stage as something to evaluate independently.
• Compare feature representations rather than assuming one is obviously correct.
• Characterize failure modes, not just aggregate accuracy.

Methodology

The pipeline follows a standard structure (preprocessing, vectorization, training, evaluation) with deliberate attention at each stage: text normalization choices were tested for their effect on downstream accuracy rather than applied by default, and multiple vectorization approaches were compared before settling on one.

Architecture

Raw text moves through a preprocessing stage (normalization, tokenization, noise removal), a vectorization stage that converts cleaned text into numerical features, and a classification stage that assigns a sentiment label. Each stage is a separate, swappable component, which made it straightforward to isolate how much a given preprocessing choice actually mattered.

Implementation

Evaluation was built around more than a single accuracy number: confusion matrices and per-class performance were used throughout to catch cases where a model looked good in aggregate but was systematically weak on, for instance, neutral-sentiment text, which is typically the hardest class to classify correctly.

Challenges Encountered

Neutral sentiment was consistently the hardest category, both for the model and for defining ground truth in the first place: human-labeled “neutral” text is often genuinely ambiguous, which puts a ceiling on any model’s achievable accuracy that has nothing to do with the model itself. Balancing preprocessing aggressiveness was another tension: more aggressive text cleaning simplified the input but occasionally removed signal that mattered for sentiment.

Results

The pipeline achieves solid separation between clearly positive and clearly negative text, with the expected drop in reliability for neutral and ambiguous cases. Its main value has been as a methodology exercise: a clean, well-evaluated reference pipeline rather than a novel modelling contribution.

Lessons Learned

Comparing preprocessing and feature-representation choices explicitly, instead of defaulting to whatever’s conventional, surfaced real differences in downstream performance that would have otherwise gone unnoticed. Evaluating per-class performance rather than only aggregate accuracy is what actually revealed where the pipeline was weak.

Future Improvements

Extending the evaluation to out-of-domain text (text stylistically different from the training data) would be the most informative next step, since that is where preprocessing and feature choices tend to matter most.

← Back to all projects