CheMapBERT

Project Description

In this project, we present a novel approach for ingredient matching in cosmetic products using a knowledge-infused language model. We fine-tuned a large language model that is pretrained on domain-specific corpora in order to generate (or match) the label, i.e intended cosmetic use given a list of cosmetic ingredients. An additional step that we did is to incorporate an external source of knowledge - list of possible matches into the model. We show how introducing the external knowledge can affect the performance of the model on this downstream task, quantitatively and qualitatively.

Technical Skills

Dependencies

1) PyTorch

  https://pytorch.org/get-started/locally/

2) TQDM

  pip install tqdm

3) Transformers

  pip install transformers

4) Pandas

  pip install pandas

5) Numpy

  pip install numpy

6) Requests

  pip install requests

7) BeautifulSoup

  pip install beautifulsoup4

Datasets

Domain Specific (Pre-training)

• Amazon Product Dataset : https://cseweb.ucsd.edu/~jmcauley/datasets/amazon_v2/

Task Specific (Fine tuning)

• Cosmetic Product : https://www.kaggle.com/datasets/kingabzpro/cosmetics-datasets
• Sephora Dataset : https://www.kaggle.com/datasets/raghadalharbi/all-products-available-on-sephora-website

Files

• BERT_Domain_Training.ipynb: Pretraining of BERT using MLM
• BERT_Classification_Task_Training.ipynb: Fine tuning of BERT Model after Weight Transfer from MLM BERT
• BERT_Task_Training.ipynb: Downstream Task Finetuning on MLM BERT (Not Recommended)
• CheMapBERT.ipynb: Implementation of BERT fine tuning w/ External Knowledge
• External_Knowledge_Retrieval.ipynb: Contains code for external knowledge retrieval
• Label_Merging_591.xlsx: Future Work Label Merging (Under Work)

How to Run

1. Download all the ipynb files from the Repository
2. Download all the datasets using the links provided
3. Run BERT_Domain_Training.ipynb for the pretraining and then store the weights (in the same folder)
4. Run CheMapBERT/BERT_Classification_Task_Training files to fine tune the pretrained model on the downstream task with and without External Knowledge Incorporation

Future Works:

• Try to use other Faster models (less parameters) - such as DistilBERT and Larger Models - such as RoBERTa
• Query to filter information into various types: Filter down the information provided through external knowledge by using FAISS or ElasticSearch
• Merging of Labels in order to have more samples for each class/ Introduce more datasets in fine-tuning