The repository contains a method for classifying and cleaning text using NLP transformers.

The input data are web-scraped product names gathered from various e-shops. The products are either monitors or printers. Each product in the dataset has a scraped name containing information about the product brand, and product model name, but also unwanted noise - irrelevant information about the item. Additionally, only some records are relevant, meaning that they belong to the correct category: monitor or printer, while other records belong to unwanted categories like accessories or TVs.

The goal of the tasks is to preprocess web-scraped data by removing noisy records and cleaning product names. Preliminary experiments showed that classic machine learning methods like tf-idf vectorization and classification struggled to achieve good results. Instead NLP transformers were employed:

• First, DistilBERT was utilized for removing irrelevant records. The available data are monitors with annotated labels where the records are classified into three classes: "Monitor", "TV", and "Noise".
• After, T5 was applied for cleaning product names by translating scraped name into clean name containing only product brand and product model name. For instance, for the given input "monitor led aoc 24g2e 24" ips 1080 ..." the desired output is "aoc | 24g2e". The available data are monitors and printers with annotated targets.

The datasets are split into training, validation and test sets without overlapping records.

The results and details about training and evaluation procedure can be found in the Jupyter Notebooks, see Content section below.

The repository contains Jupyter Notebooks for training and evaluating NNs:

• 01_data_exploration.ipynb - The notebook contains an exploration of the datasets for sequence classification and translation. It includes visualization of distributions of targets, and overview of available metadata.
• 02a_classification_fine_tuning.ipynb - The notebook fine-tunes a DistilBERT classifier using training and validation sets, and saves the trained checkpoint.
• 02b_classification_evaluation.ipynb - The notebook evaluates classification scores on the test set. It includes: a classification report with precision, recall and F1 scores; and a confusion matrix.
• 03a_translation_fine_tuning.ipynb - The notebook fine-tunes a T5 translation network using training and validation sets, and saves the trained checkpoint.
• 03b_translation_evaluation.ipynb - The notebook evaluates translation metrics on the test set. The metrics are: Text Accuracy (exact match of target and predicted sequences); Levenshtein Score (normalized reversed Levenshtein Distance where 1 is the best and 0 is the worst); and Jaccard Index.
• 04_benchmarking.ipynb - The notebook evaluates GPU memory and time needed for running inference on DistilBERT and T5 models using various values of batch size and sequence length.

The method were developed using Python=3.7 with transformers=4.8 framework that uses PyTorch=1.9 machine learning framework on a backend. Additionally, the repository requires packages: numpy, pandas, matplotlib and datasets.

To install required packages with PyTorch for CPU run:

pip install -r requirements.txt

For PyTorch with GPU run:

pip install -r requirements_gpu.txt

The requirement files do not contain jupyterlab nor any other IDE. To install jupyterlab run

pip install jupyterlab

Rail Chamidullin - chamidullinr@gmail.com - Github account