查看文章

To tackle this problem, we rely on deep learning techniques to (1) read a new product’s textual description,(2) find similar products to it based on semantic similarity in their descriptions, and (3) use them to predict the new product’s taxonomy. Since the product taxonomy follows a hierarchical structure, the prediction task is a hierarchical multi-class classification task, so for each product, the model predicts one class for each level of the hierarchy. By learning a similarity function instead of training a classifier from scratch, at inference time we view the product taxonomy prediction as a one-shot learning problem [2], as it does not require that every class in the taxonomy be present in the training data. This is especially important as product catalogs evolve over time and we do not wish to retrain the model every time there is a change to the catalog, such as when new products are introduced. Moreover, after we train the model, we use it to precompute embedding vectors for all product descriptions in our catalog. In this way, we can use these precomputed vectors when finding similar products and performing taxonomy classification in a very efficient manner, at scale.

In summary, here’s what we set out to accomplish:• Train a product similarity model using a Transformer-based deep Siamese network based on a pre-trained Microsoft DeBERTa [1] model, which is fine-tuned using textual product descriptions• Use the model to generate contextual embeddings that capture the semantic information in product descriptions• Use the generated embeddings for a downstream task consisting of one-shot product taxonomy classification• Design and conduct …