erkanbalaban / supervised-product-matching

⚖️ Neural network for product matching, aka classifying whether two product titles represent the same entity

Product Matching Neural Network

This project aims to create a model using CharacterBERT (and added Transformers in some models) that is able to classify two product titles as representing the same entity or not. This project train a model to specifically discern between electronics titles.

Example 1:

Title 1: ASUS VivoBook Thin and Lightweight FHD WideView Laptop, 8th Gen Intel Core i5-8250U, 8GB DDR4 RAM, 128GB SSD+1TB HDD, USB Type-C, NanoEdge, Fingerprint Reader, Windows 10 - F510UA-AH55

Title 2: ASUS Laptop 15.6, Intel Core i5-8250U 1.6GHz, Intel HD, 1TB HDD + 128GB SSD, 8GB RAM, F510UA-AH55

Using these two titles, the model should output a 1

Example 2:

Title 1: AMD Ryzen 5 5600X 6-core, 12-Thread Unlocked Desktop Processor with Wraith Stealth Cooler

Title 2: AMD Ryzen 7 5800X 8-Core 3.8 GHz Socket AM4 105W 100-100000063WOF Desktop Processor

Using these two titles, the model should output a 0

Project Overview

data/base contains data that is going to be transformed into training data.

data/train contains data used to actually train.

data/test contains data used to validate the models trained.

torch_train_model.py is where to train the model.

test_model.py allows you to use the validation script on a specific model.

create_data.py uses functions under src/data_creation to transform data found in base

The supervised_product_matching directory contains code associated with the model.

The src directory are the functions that create data.

The models directory contains the different models trained so far and also the fastText model (if you want to use the ).

The src/data_scrapers directory contains scripts to scrape data for creating training data.

The pretrained-models directory is where the user should put the bert and character_bert models.

• The CharacterBERT model can be downloaded using the author's repository here
• The BERT model can be downloaded using HuggingFace Transformers

The Data

All the data can be found in the repository's latest release.

Source Code (Under src)

The data_creation directory contains scripts that transforms data in base into usable training data.

The data_scrapers directory uses web scraping scripts to get raw data (like product titles for laptops off of different retailers) to be processed into training data.

common.py and data_preprocessing.py are functions used throughout the other scripts

Package (Under supervised_product_matching)

The model_architectures directory contains different neural network architectures to use for training (all written using pytorch). They include:

• BERT
• CharacterBERT
• CharacterBERT with my custom Transformer added on top
• CharacterBERT that concatenates word embeddings together as opposed to adding and averaging

config.py just contains variables needed to define the model architectures.

model_preprocessing contains code to format data to feed into the model.

The reason for the seperate folder (which is really a package) is to make the model more portable. First, install Character BERT using:

pip install -e git+https://github.com/Mascerade/character-bert#egg=character_bert

Then, install this package using:

pip install -e git+https://github.com/Mascerade/supervised-product-matching#egg=supervised_product_matching