Application for training the pretrained transformer model DeBERTaV3 (see paper DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing) on an Aspect Based Sentiment Analysis task.

Aspect Based Sentiment Analysis is a Sequence Labeling task where product reviews are labeled with their aspects as well as the detected sentiments towards each of these aspects. Aspects in the context of product reviews are N-Grams explicitly mentioning specific functionalities, parts and related services around the product, with the part of speech being limited to nouns, noun phrases or verbs.

Example of an annotated product review

Training data source for the model were 1.570 sampled product reviews (5.872 sentences) from the Amazon Review Dataset - specifically from the five product categories Laptops, Cell Phones, Mens Running Shoes, Vacuums, Plush Figures - which I manually annotated for my bachelor's thesis following a modified version of the SemEval2014 Aspect Based Sentiment Analysis guidelines and the annotation tool Universal Data Tool.

The model was trained for 10 epochs on the combined dataset from all five categories (training time: 02h:05m:03s on NVIDIA GeForce GTX 1660 Ti). Model training, evaluation and inference is implemented using the wrapper simpletransformers which uses huggingface. Since it requires word tokenized and sentence tokenized inputs, the raw text is first pre-processed using SpaCy.

The frontend and routing is implemented in Flask, using Jinja as Template Engine for rendering the HTML and Bootstrap for the frontend design.

• pytorch==2.6.0
• cudatoolkit=12.6

• simpletransformers
• spacy
• pandas
• openpyxl
• tqdm
• flask

• en_core_web_lg

The uploaded versions of the training data in this repository are cut off after the first 1.000 rows of each file, the real training data contains a combined ~90.000 rows. The trained model file pytorch_model.bin is omitted in this repository.