/LLM-knowledge-graph

Primary LanguagePython

LLM-Driven Knowledge Graph for Customer Review Analysis and Product Recommendation

DE & Ingestion

ETL using Airflow (get reviews from online reddit/some website real time) Amazon product review -> data contract -> store in S3/sqlite/AWS-RDS/GCP-BigQuery Store in relation or graph database

LLM Fine tuning

Fine tune LLM (BERT, GPT-3) on customer review data -> Entity, relation and sentiment extraction.

  1. Extract entities & Extract relations b/w entities (GPT-3.5)
  2. sentiment of sentence (fine tune BERT/LLM model to get sentiment)
  3. store entities (products, customers, features) and their relation in a relational database/CSV
  4. knowledge graph that captures relation b.w customers, products, features and sentiments.

Knowledge graph

  1. Entities (Nodes)

    • Customer: Represents the individuals who are providing reviews or feedback.
    • Product: Represents the items that customers are reviewing.
    • Feature: Represents specific aspects or attributes of a product (e.g., battery life, camera quality).
    • Sentiment: Represents the emotional tone or sentiment associated with a customer's review of a feature (e.g., positive, negative, neutral).

  2. Relationships (Edges)

    • REVIEWED: Connects a Customer to a Product, indicating that the customer has reviewed the product.
    • HAS_FEATURE: Connects a Product to a Feature, indicating that the product has a particular feature.
    • ASSOCIATED_WITH: Connects a Feature to a Sentiment, indicating the sentiment associated with a particular feature based on customer feedback.

UI & Neo4J

Docker

Deploy to EC2

Monitoring and Logging

Unit Testing

Flask

  1. Upload a file with specific data contract or create one and store it in DB
    • In the background compute all things like extract entities, relations and sentiment. (With custom function - outputs in JSON format)
    • Read the JSON format and provide it to Neo4J for visualization