AspectWise
End-to-end automated pipeline for review collection, analysis and report creation
Q Agentur für Forschung offers a service for automated analysis of customer reviews in any product category in any European language. I’m the inventor and lead developer. We build on our research in aspect-based sentiment analysis (ABSA). Since our paper on the SemEval benchmark, we have re-thought ABSA from the ground up and developed a labeling scheme and model that allows for fine-grained analysis of customer reviews. See this example:
This model is the core of our pipeline. There are two other steps: review collection and reporting.
Everything’s automated in Dagster with data quality checks at every step. Here’s the full stack:
| Category | Technology |
|---|---|
| Data | Web scraping services accessed via API |
| Data Warehouse | MotherDuck |
| Data Transformation | dbt, polars |
| Orchestrator | Dagster |
| Models | Fine-tuned LLMs + Prompt engineering |
| Experiment Tracking | Weights & Biases |
| Labeling Tool | Custom Shiny for Python app |
| Reporting | Quarto |
The project leverages several machine learning models for ABSA, text classification, translation, and summarization.
See the AspectWise website and a case study for more details.