ZeroShotCeres: Zero-Shot Relation Extraction from Semi-Structured Webpages

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8105–8117 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8105 ZeroShotCeres: Zero-Shot Relation Extraction from Semi-Structured Webpages Colin Lockard University of Washington lockardc@cs.washington.edu Prashant Shiralkar Amazon shiralp@amazon.com Xin Luna Dong Amazon lunadong@amazon.com Hannaneh Hajishirzi University of Washington, Allen Institute for AI hannaneh@cs.washington.edu Abstract In many documents, such as semi-structured webpages, textual semantics are augmented with additional information conveyed using visual elements including layout, font size, and color. Prior work on information ex- traction from semi-structured websites has re- quired learning an extraction model specific to a given template via either manually la- beled or distantly supervised data from that template. In this work, we propose a solu- tion for “zero-shot” open-domain relation ex- traction from webpages with a previously un- seen template, including from websites with little overlap with existing sources of knowl- edge for distant supervision and websites in en- tirely new subject verticals. Our model uses a graph neural network-based approach to build a rich representation of text fields on a web- page and the relationships between them, en- abling generalization to new templates. Exper- iments show this approach provides a 31% F1 gain over a baseline for zero-shot extraction in a new subject vertical. 1 Introduction Semi-structured websites offer rich sources of high- quality data across many areas of knowledge (Dong et al., 2014). These websites present information via text that is accompanied by rich visual and layout features that can be generalized beyond a single website. However, most prior work on infor- mation extraction (IE) from websites has largely ignored most of these features, instead relying only on HTML features specific to an individual web- site (Ferrara et al., 2014). This requires training data for every website targeted for extraction, an approach that cannot scale up if training data must be manually created. To circumvent manual data annotation, previous work used a distant supervision process requiring a knowledge base aligned to the website targeted Figure 1: Our zero-shot open-domain information ex- traction process learns generalizable graph-based rep- resentations of how relations are visually presented on semi-structured websites, allowing for training on one vertical (such University sites) and extraction from an- other (such as Movie sites). for extraction (Gentile et al., 2015; Lockard et al., 2018), including for OpenIE extraction (Banko et al., 2007; Bronzi et al., 2013; Lockard et al., 2019). These methods, however, can only learn a website-specific model based on seed knowledge for the site, but cannot be generalized to the major- ity of websites with knowledge from new verticals, by long-tail specialists, and in different languages. In this paper, we introduce the task of zero-shot relation extraction from semi-structured websites, in which a learned model is applied to extract from a website that was not represented in its train- ing data (Figure 1). Moreover, we introduce ZE- ROSHOTCERES, a graph neural network model that encodes semantic textual and visual patterns common across different training websites and can generalize …

Content truncated to 500 words