rid17pawar/semantic-search-model-experiments

Experiments in the field of Semantic Search using BM-25 Algorithm, Mean of Word Vectors, along with state of the art Transformer based models namely USE and SBERT.

Semantic-Search-Model-Experiments

Dataset Used For Semantic Search/ Information Retrieval:

CISI Dataset - Kaggle

Experiments:

Experiment-1. Using BM-25 Algorithm and Parameter Tuning For Semantic Search

BM-25 Algorithm variations used:

• BM25Okapi
• BM25L
• BM25Plus
  
  

Result:

BEST MODEL: BM25Plus

Experiment-2. Using Mean of Word Vectors (MWV) with Pretrained Embeddings For Semantic Search

BM-25 Algorithm variations used:

• word2vec
• GloVe
• FastText
  
  

Result:

BEST MODEL: word2vec

Experiment-3. Using LDA Topic Modelling For Semantic Search

Result:

Performs worst than BM-25

Experiment-4. Using Universal Sentence Encoder (USE) For Semantic Search

USE Model variations used:

• Transformer Encoder
• Deep Averaging Network(DAN) Encoder
  
  

Result:

BEST MODEL: USE-Transformer

Experiment-5. Using Pretrained and Finetuned Sentence Transformers (SBERT) For Semantic Search

Result:

BEST MODEL: Finetuned SBERT

Final Result:

Overall Best Model: Finetuned SBERT