Exploring generative frameworks for product attribute value extraction

Abstract

E-commerce platforms rely heavily on the attribute values of their products as they play a crucial role in various retail functions such as product search, recommendations, and question answering. Therefore, identifying the attribute values from unstructured product information is critical for any e-commerce retailer. This problem is challenging due to the diversity of product types and their attributes and values. Attribute value extraction deals with extracting the values of attributes from the product profile. Previous approaches for this task have formulated the attribute value extraction as a Named Entity Recognition task (NER) or a Question Answering (QA) task. In this paper, we propose to tackle the attribute value extraction task using generative frameworks. In the first task, the attribute name and the product title are used to generate the value of the attribute. In the second task, only the product title is utilized to extract both the attributes and the values jointly. The value extraction is formulated as a text-infilling task and an answer-generation task. For joint attribute value extraction, we present two types of generative paradigms, namely, word sequence-based paradigm and positional sequence-based paradigm. The pre-trained language models such as GPT-2, BART, T5, and FLAN-T5 are leveraged to perform these tasks. Our experiments show that a single general model effectively performs value extraction task over a broad set of product attributes. Experiments conducted on two datasets depict that the generative approaches achieve new state-of-the-art results, which indicates that the proposed frameworks are helpful for attribute value extraction tasks without additional tagging.