Incorporating Multimodal Information in Open-Domain Web Keyphrase Extraction

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1790–1800, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1790 Incorporating Multimodal Information in Open-Domain Web Keyphrase Extraction Yansen Wang∗, Zhen Fan∗, Carolyn P. Ros´e Language Technologies Institute, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh PA, 15213 {yansenwa, zhenfan, cp3a}@andrew.cmu.edu Abstract Open-domain Keyphrase extraction (KPE) on the Web is a fundamental yet complex NLP task with a wide range of practical applica- tions within the field of Information Retrieval. In contrast to other document types, web page designs are intended for easy navigation and information finding. Effective designs encode within the layout and formatting signals that point to where the important information can be found. In this work, we propose a model- ing approach that leverages these multi-modal signals to aid in the KPE task. In particular, we leverage both lexical and visual features (e.g., size, font, position) at the micro-level to enable effective strategy induction, and meta- level features that describe pages at a macro- level to aid in strategy selection. Our evalu- ation demonstrates that a combination of ef- fective strategy induction and strategy selec- tion within this approach for the KPE task out- performs state-of-the-art models. A qualitative post-hoc analysis illustrates how these features function within the model. 1 Introduction We present a novel multi-modal approach to KeyPhrase Extraction (KPE), which is the task of automatically extracting salient phrases from a given document. The KPE task is a founda- tional task, which plays a facilitating role in many Information Retrieval (IR) tasks, including clas- sification, summarization, and document index- ing (Hasan and Ng, 2014). Specifically, the KPE task requires accurate selection of the phrases that best capture the web document’s topic. Well- performing approaches take advantage of docu- ment structure and entity co-occurrences. Over the history of work in this area, there have been a variety of benchmarks (Medelyan and Wit- ten, 2002; Nguyen and Kan, 2007; Wan and Xiao, ∗ Equally contributed. 2008; Meng et al., 2017) and an equally wide variety of both non-neural (Grineva et al., 2009; Liu et al., 2009, 2010) and neural modeling ap- proaches (Meng et al., 2017; Zhang et al., 2017; Chen et al., 2018). The earliest KPE approaches were mainly limited to domain-specific keyphrase extraction. The recent release of OpenKP (Xiong et al., 2019), a large-scale feature-rich dataset specifically developed for open-domain web-page keyphrase extraction, has encouraged further re- search related to the KPE task. A novel charac- teristic of this data set is the inclusion of features related to visual properties. Visual properties of words and web page layout offer a KPE model utility in at least two respects. First, micro-level features operating at the word level, including lexical features as well as features related to size, font, color, and position of words, signal relative importance of words within ex- tended texts. Intuitively, texts that are highlighted with colored, bold or bigger font or placed in more obvious places within a web page are more …

Content truncated to 500 words