Lamees Mohee

R&D AI Intern

• Job Details:Multimodal e-commerce product attribute value extraction from image and text (graduation project). Explored various models from past research, including opentag, su-opentag, and aveqa, to understand their effectiveness for product attribute extraction. Performed data preprocessing, including handling missing or unlabeled data, which increased the dataset size by approximately 30,000 additional samples. Enhanced approximately 18,000 noisy product descriptions using deepseek api to improve attribute extraction accuracy. Created a category prediction model using s-bert and neural networks, achieving an impressive accuracy of 98%. Implemented the su-opentag model for 12 attributes, achieving an overall f1 score of 85%. Fine-tuned bert for the named entity recognition (ner) problem, where the entities correspond to product attributes, achieving an enhanced f1 score of 91%. Investigated various fusion techniques for combining two pretrained text encoders, bert and roberta, with vision transformer (vit) embeddings, and utilized a blip decoder to generate attribute values, achieving an enhanced f1 score of 92.9%.