TIE: Topological Information Enhanced Structural Reading Comprehension on Web Pages

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1808 - 1821 July 10-15, 2022 ©2022 Association for Computational Linguistics TIE: Topological Information Enhanced Structural Reading Comprehension on Web Pages Zihan Zhao, Lu Chen∗, Ruisheng Cao, Hongshen Xu, Xingyu Chen and Kai Yu∗ X-LANCE Lab, Department of Computer Science and Engineering MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China Shanghai Jiao Tong University, Shanghai, China State Key Laboratory of Media Convergence Production Technology and Systems zhao_mengxin@sjtu.edu.cn, chenlusz@sjtu.edu.cn {211314, xuhongshen, galaxychen, kai.yu}@sjtu.edu.cn Abstract Recently, the structural reading comprehen- sion (SRC) task on web pages has attracted increasing research interests. Although previ- ous SRC work has leveraged extra informa- tion such as HTML tags or XPaths, the in- formative topology of web pages is not ef- fectively exploited. In this work, we pro- pose a Topological Information Enhanced model (TIE), which transforms the token-level task into a tag-level task by introducing a two-stage process (i.e. node locating and answer refining). Based on that, TIE inte- grates Graph Attention Network (GAT) and Pre-trained Language Model (PLM) to lever- age the topological information of both logical structures and spatial structures. Experimen- tal results demonstrate that our model outper- forms strong baselines and achieves state-of- the-art performances on the web-based SRC benchmark WebSRC at the time of writing. The code of TIE will be publicly available at https://github.com/X-LANCE/TIE. 1 Introduction With the rapid development of the Internet, web pages have become the most common and rich source of information (Dong et al., 2014). There- fore, the ability to understand the contents of struc- tured web pages will guarantee a rich and diverse knowledge source for deep learning systems. Each web page is mainly rendered from the correspond- ing HyperText Markup Language (HTML) codes. In other words, the understanding of a structured web page can be achieved by the comprehension of its HTML codes. One of the commonly used tasks to verify the model’s ability of comprehension is Question An- swering (QA). However, previous QA models only focus on the comprehension of plain texts (Ra- jpurkar et al., 2016; Yang et al., 2018; Reddy et al., 2019; Zeng et al., 2020), tables (Pasupat and Liang, *The corresponding authors are Lu Chen and Kai Yu. Figure 1: An example of web pages in WebSRC and its corresponding Document Object Model (DOM) tree and Node Positional Relation (NPR) graph in WebSRC. The colored HTML tag in (a) is corresponding to the bounding box with the same color in (a) and the node with the same color in (b) and (c). 2015; Chen et al., 2020c, 2021b), or knowledge bases (KBs) (Berant et al., 2013; Talmor and Be- rant, 2018). These sources have either no topo- logical structure or fixed-form structures. On the contrary, the topological structures of web pages are complex and flexible, which are less investi- gated in previous QA works. Specifically, HTML codes can be viewed as mul- tiple semantic …

Content truncated to 500 words