Scaling up Open Tagging from Tens to Thousands: Comprehension Empowered Attribute Value Extraction from Product Title

Abstract

Supplementing product information by extracting attribute values from title is a crucial task in e-Commerce domain. Previous studies treat each attribute only as an entity type and build one set of NER tags (e.g., BIO) for each of them, leading to a scalability issue which unfits to the large sized attribute system in real world e-Commerce. In this work, we propose a novel approach to support value extraction scaling up to thousands of attributes without losing performance: (1) We propose to regard attribute as a query and adopt only one global set of BIO tags for any attributes to reduce the burden of attribute tag or model explosion; (2) We explicitly model the semantic representations for attribute and title, and develop an attention mechanism to capture the interactive semantic relations in-between to enforce our framework to be attribute comprehensive. We conduct extensive experiments in real-life datasets. The results show that our model not only outperforms existing state-of-the-art N-ER tagging models, but also is robust and generates promising results for up to 8, 906 attributes.

Learning to Extract Attribute Value from Product via Question Answering: A Multi-task Approach

“…If we remove the distilled MLM and the no-answer classifier by setting both and to 0, our model degenerates to the standard question answering model with BERT [7]. If we further replace the BERT contextual layer of the QA component with the BiLSTM layer, our model is regressed to the sequence tagging model in [50]. Moreover, if we also remove the question (attribute) from the QA model, our model is degenerates to the attribute-dependent OpenTag method [54], which is not able to scale to large attribute set.…”

“…The most recent attribute value extraction model [50] employs two separate LSTM-based contextual layers for the context and the question respectively, followed by a cross-attention layer to join the outputs of the two layers. Different from them, we utilize one unique contextual encoder with self-attention mechanism developed in BERT [7].…”

“…The new concatenated embedding is then used for finding the best end index. This beginend dependency modeling is similar to the usage of CRF layer in the open tagging models [50,54].…”

“…Several other approaches such as [29,35] formulate the attribute value extraction as an instance of named entity recognition (NER) problem [30], and build extraction models to identify the entities/values from the input text. With the recent advance in natural language understanding, sequence tagging [50,54] based approaches have been proposed, which achieve promising results. However, these techniques suffer from two major limitations:…”

“…• Scalability -they do not scale to millions of attributes that are necessary for real world applications. For instance, the AliExpress taxonomy contains thousands of product categories, and a single category, Sports & Entertainment, has over 8.9k unique attributes [50]. Existing methods treat each attribute independently and build one separate model for each of them, which are not suitable for large scale attribute systems.…”