Abstract
Extracting product attribute values is essential for e-commerce applications like product search, retrieval, and recommendation. While existing methods excel in extracting values from product descriptions, little attention has been given to extracting values from product profiles, which may contain coarse-grained information consisting of multiple attributes at the same time. In this work, we propose QPAVE, a novel multi-task question answering approach for fine-grained attribute value extraction. QPAVE treats each fine-grained attribute as a question, identifying values from the context of a coarse-grained attribute. To capture dependencies, we employ a hypernetwork to parameterize the decoding layer of the question answering model with the embeddings of the coarse-grained attribute. Additionally, the model predicts the product category in a multi-task training setup to learn category-aware token embeddings. We conducted extensive experiments on a large distantly annotated dataset and a gold human-annotated test set showing superior performance over several state-of-the-art methods. Our code and data are available at https://github.com/kassemsabeh/qpave. We also release our model as a demo at https://bit.ly/3U8rbuI.