Fintech Review ABSA
A structured review data workflow for risk, trust, and service signals, with applied NLP as the downstream layer.
The Data Problem
The Operating Context
"The repository is a curated public version of a larger thesis and experimentation workspace, designed to keep the full data preparation, evaluation, and delivery workflow visible without bundling private datasets or heavy checkpoints."
A single overall sentiment score was too coarse for fintech reviews because the same review can talk about billing risk, platform trust, and service quality at the same time. The first requirement was a clean aspect schema and reproducible text preparation before downstream modeling.
- The Reliability Trade-offs: Each workflow balanced data preparation effort, downstream usefulness, and measurable reliability under practical constraints.
- Constraint 01: Public release scope: the repo had to keep its data and evaluation workflow understandable without shipping raw private assets, large checkpoints, or machine-local experiment folders.
- Constraint 02: Noisy Indonesian review text: preprocessing and dataset reconciliation had to be explicit before model comparison became meaningful.
The Data Flow
Pipeline Rationale
Data Flow
Built the repository around a practical data-to-output loop: collect and preprocess noisy Google Play reviews, reconcile the active dataset, run downstream aspect-based inference, compare baseline and PEFT tracks, and surface outputs in Streamlit.
Downstream Outcome
The public repo reads as a complete data-to-output workflow instead of a model demo, with evaluation artifacts, a live dashboard surface, and reproducible entry points for inference and comparison.
Source and preparation
Google Play review collection, preprocessing, and dataset reconciliation for the active ABSA setup.
Downstream inference and training
baseline and PEFT experiment tracks including LoRA, DoRA, AdaLoRA, and QLoRA for risk / trust / service prediction.
Validation Results
Data schema evidence: The workflow produces three domain-specific outputs - risk, trust, and service - which is more useful than a single sentiment label.
Experiment evidence: Baseline and PEFT tracks are both included, so downstream comparison is built into the repository instead of implied.
Delivery evidence: The dashboard makes prepared artifacts and live inference readable, turning the project into a usable analysis surface rather than a notebook-only experiment.
Pipeline Delivery & Reflections
Schema design: One review can contain multiple business signals, so the preparation step must preserve risk, trust, and service rather than collapse everything into one polarity label.
Workflow packaging: A public repository does not need every checkpoint bundled if preprocessing, evaluation entry points, and summary artifacts are documented clearly.
Downstream usability: The dashboard is part of the technical story because it shows how validated multi-aspect outputs become inspectable for non-model users.