Mining Health-related Cause-Effect Statements with High Precision at Large Scale

Proceedings of the 29th International Conference on Computational Linguistics, pages 1925–1936 October 12–17, 2022. 1925 Mining Health-related Cause–Effect Statements with High Precision at Large Scale Ferdinand Schlatt ⋆ Dieter Bettin ‡ Matthias Hagen ⋆ Benno Stein † Martin Potthast § ⋆Martin-Luther-Universität Halle-Wittenberg †Bauhaus-Universität Weimar ‡Westfälische Wilhelms-Universität Münster §Universität Leipzig Abstract An efficient assessment of the health related- ness of text passages is important to mine the web at scale to conduct health sociological anal- yses or to develop a health search engine. We propose a new efficient and effective termhood score for predicting the health relatedness of phrases and sentences, which achieves 69% re- call at over 90% precision on a web dataset with cause–effect statements. It is more effective than state-of-the-art medical entity linkers and as effective but much faster than BERT-based approaches. Using our method, we compile the Webis Medical CauseNet 2022, a new resource of 7.8 million health-related cause–effect state- ments such as “Studies show that stress induces insomnia” in which the cause (‘stress’) and ef- fect (‘insomnia’) are labeled. 1 Introduction Health sociology studies the interaction of soci- ety with health. An important subject is how consumers obtain and perceive health information. Since the web and search engines are among the most important sources, the quality of online health information has been studied so frequently in re- cent decades that three systematic reviews have been conducted (see Table 1). However, many of the individual studies address only a single medical condition and almost all were conducted manually. On average, only about 50–100 web pages were analyzed per study, sometimes only a single hand- picked one, and never more than 1,524 pages. Studying larger portions of the online health do- main requires the automation of various acquisition tasks: (1) the discovery of health-related websites and web pages, (2) the identification and extraction of health-related statements from these pages, and (3) the attribution of health-related statements to au- thoritative sources (e.g., for fact-checking). While the first and third steps have been and continue to be the subject of ongoing research, the second step has received much less attention. Systematic Review Studies Websites / Web Pages Min Max Mean Stddev Total Eysenbach et al. (2002) 79 3 1,147 100.5 157.7 7,836 Zhang et al. (2015) 165 3 388 78.5 73.4 12,870 Daraz et al. (2018) 157 1 1,524 50.3 133.9 7,891 Table 1: Sizes of systematic reviews of online health in- formation studies. Some studies are part of multiple re- views; most do not distinguish websites and web pages. To minimize manual data cleansing, we view the extraction of health-related statements as a precision-oriented task. And since a significant por- tion of consumers’ health information needs ask about causes and effects (Bondarenko et al., 2022), we focus on health-related cause–effect statements (e.g., ‘smoking causes cancer’). Extracting cause– effect statements in general has been thoroughly investigated in the past and several approaches ex- tract them from web text efficiently and effectively (Yang et al., 2022). However, the extracted state- ments …

Content truncated to 500 words