Enhanced Feedback Analysis System

Agentic LLM RAG for NLP analysis including Named Entity Recognition, sentiment quintuple extraction, coreference resolution & comparative opinion mining for comprehensive customer feedback insights

Comment

Enhanced Feedback Analysis System

Inspiration

Traditional feedback analysis systems often fail to capture the nuanced relationships between entities, sentiments, and comparative opinions in customer reviews. We were inspired to create a comprehensive NLP solution that goes beyond simple sentiment analysis to provide multi-dimensional insights into customer feedback.

What it does

Our system performs five key NLP tasks:

  1. Text Cleaning & Preprocessing
  2. Named Entity Recognition (NER) - Identifies entities like people, organizations, locations, dates
  3. Sentiment Quintuple Extraction - Extracts Target Objects, Features, Sentiments, Opinion Holders, and Time references
  4. Coreference Resolution - Resolves pronouns and references to their actual entities
  5. Comparative Opinion Analysis - Identifies and analyzes comparative statements between different entities

How we built it

  • Backend: Flask API with Server-Sent Events for real-time streaming
  • AI/ML Stack: Mistral LLM with Phidata Agentic Framework for few-shot learning
  • Vector Database: FAISS with all-MiniLM-L6-v2 embeddings (384-dimensional)
  • Frontend: Modern responsive UI with Bootstrap and Material Icons
  • Data: Annotated Air India flight reviews from Kaggle for domain-specific training

The system uses agentic RAG (Retrieval-Augmented Generation) with cosine similarity search to retrieve relevant annotated examples, providing context to the LLM for accurate analysis.

What we learned

  • Advanced NLP techniques including quintuple extraction and coreference resolution
  • Implementation of agentic frameworks for complex AI workflows
  • Vector embeddings and similarity search optimization
  • Real-time streaming with Server-Sent Events
  • Domain-specific model adaptation using few-shot learning

Challenges we faced

  • Context Management: Implementing effective few-shot learning with limited context windows
  • Real-time Processing: Ensuring smooth streaming of analysis results across multiple NLP tasks
  • Complex Data Structures: Handling nested sentiment analysis and comparative opinion extraction

What's next

  • Expand to multiple domains (restaurants, products, healthcare)
  • Add multilingual support
  • Implement advanced visualization dashboards
  • Scale to handle larger datasets and real-time social media feeds

🔗 Domain Adaptability

While currently optimized for airline reviews, our methodology is domain-agnostic. The same approach can be extended to any industry by simply replacing the training data with domain-specific annotated feedback.

Built With

  • bootstrap
  • cosine-similarity-search
  • cosinesimilarity
  • faiss-(facebook-ai-similarity-search)
  • flask
  • html/css
  • javascript
  • langchain
  • mistral-llm
  • phidata-agentic-framework
  • python
  • rest-api
  • sentence-transformers-(all-minilm-l6-v2)
  • server-sent-events-(sse)
  • vector-embeddings-(384-dimensional)
Share this project:

Updates