How chmorfop/sentiment_analysis_llm works
Project Type and Purpose
• This is an analysis and visualization project using a Jupyter Notebook (Greek_Sentiment_Analysis.ipynb) that processes sentiment data extracted from movie reviews.
• The prime goal is to perform sentiment classification (positive, negative, neutral) across a large dataset (150,000 reviews) using pretrained large language models.
• Additionally, the project integrates data exploration and visualization (e.g., Wordclouds and performance charts).
Main Components
• Data Sources:
• External Datasets: Reviews from Athinorama and Kaggle.
• Data Preprocessing & Ingestion:
• Data Cleaning and Preparation: Steps that transform raw review text into input suitable for the models.
• Model Fine-tuning:
• Three finetuned models (BERT, T5, GPT2) – each can be represented as separate components or sub-modules, showing training and inference.
• Analysis & Visualization:
• Visualization Module: Including the generation of Wordclouds and performance charts (images in the “images” directory).
• Reporting & Output:
• Jupyter Notebook serving as the central computational document.
• Supplementary Presentation (Presentation.pptx) and Report (Report.pdf) for summarizing results.
Relationships and Data Flow
• Pass this data into the Preprocessing module, which cleans and prepares the text.
• The cleaned data then feeds into the three model fine-tuning modules, each processing the input to classify sentiment.
• Model outputs lead to the Analysis module, which generates visualization outputs (e.g., Wordclouds and performance metrics) as showcased in the images folder.
• Finally, the results are collated and documented in the Jupyter Notebook, with additional summarization provided in the Presentation and Report files.
Architectural Patterns and Design Elements
• Although this is not a full-stack web application, you still have a clear pipeline architecture (ETL and Model Training/Evaluation pipeline).
• Modular separation (data ingestion, preprocessing, model fine-tuning, evaluation, and reporting) to keep components decoupled.
Relevant Technologies
• Jupyter Notebook as the main orchestration document.
• Pretrained models (BERT, T5, GPT2) – these may be used via frameworks like Hugging Face’s Transformers.
• Data visualization libraries (e.g., matplotlib, seaborn, or wordcloud packages) that created the images.
• Underlying data source tools or scripts managing external data ingestion.