AdaTag: Multi-Attribute Value Extraction from Product Profiles with Adaptive Decoding

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4694–4705 August 1–6, 2021. ©2021 Association for Computational Linguistics 4694 AdaTag: Multi-Attribute Value Extraction from Product Profiles with Adaptive Decoding Jun Yan1∗, Nasser Zalmout2, Yan Liang2, Christan Grant3, Xiang Ren1, Xin Luna Dong2 University of Southern California1, Amazon.com2, University of Oklahoma3, {yanjun, xiangren}@usc.edu, {nzalmout, ynliang, lunadong}@amazon.com, cgrant@ou.edu Abstract Automatic extraction of product attribute val- ues is an important enabling technology in e-Commerce platforms. This task is usu- ally modeled using sequence labeling archi- tectures, with several extensions to handle multi-attribute extraction. One line of pre- vious work constructs attribute-specific mod- els, through separate decoders or entirely sep- arate models. However, this approach con- strains knowledge sharing across different at- tributes. Other contributions use a single multi- attribute model, with different techniques to embed attribute information. But sharing the entire network parameters across all at- tributes can limit the model’s capacity to cap- ture attribute-specific characteristics. In this paper we present AdaTag, which uses adap- tive decoding to handle extraction. We param- eterize the decoder with pretrained attribute embeddings, through a hypernetwork and a Mixture-of-Experts (MoE) module. This al- lows for separate, but semantically correlated, decoders to be generated on the fly for differ- ent attributes. This approach facilitates knowl- edge sharing, while maintaining the specificity of each attribute. Our experiments on a real- world e-Commerce dataset show marked im- provements over previous methods. 1 Introduction The product profiles on e-Commerce platforms are usually comprised of natural texts describing prod- ucts and their main features. Key product features are conveyed in unstructured texts, with limited impact on machine-actionable applications, like search (Ai et al., 2017), recommendation (Kula, 2015), and question answering (Kulkarni et al., 2019), among others. Automatic attribute value extraction aims to obtain structured product fea- tures from product profiles. The input is a textual ∗ Most of the work was done during an internship at Amazon. Figure 1: An example of the product profile on an e- Commerce platform. It consists of a title, several infor- mation bullets, and a product description. sequence from the product profile, along with the required attribute to be extracted, out of potentially large number of attributes. The output is the corre- sponding extracted attribute values. Figure 1 shows the profile of a moisturizing cream product as an example, which consists of a title, several infor- mation bullets, and a product description. It also shows the attribute values that could be extracted. Most existing studies on attribute value extrac- tion use neural sequence labeling architectures (Zheng et al., 2018; Karamanolakis et al., 2020; Xu et al., 2019). To handle multiple attributes, one line of previous contributions develops a set of “attribute-specific” models (i.e., one model per attribute). The goal is to construct neural networks with (partially) separate model parameters for dif- ferent attributes. For example, one can construct an independent sequence labeling model for each attribute and make predictions with all …

Content truncated to 500 words