“PageRank” for Argument Relevance

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1117–1127, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics “PageRank” for Argument Relevance Henning Wachsmuth and Benno Stein and Yamen Ajjour Faculty of Media, Bauhaus-Universität Weimar, Germany {henning.wachsmuth, benno.stein, yamen.ajjour}@uni-weimar.de Abstract Future search engines are expected to de- liver pro and con arguments in response to queries on controversial topics. While ar- gument mining is now in the focus of re- search, the question of how to retrieve the relevant arguments remains open. This pa- per proposes a radical model to assess rel- evance objectively at web scale: the rel- evance of an argument’s conclusion is de- cided by what other arguments reuse it as a premise. We build an argument graph for this model that we analyze with a recur- sive weighting scheme, adapting key ideas of PageRank. In experiments on a large ground-truth argument graph, the resulting relevance scores correlate with human av- erage judgments. We outline what natural language challenges must be faced at web scale in order to stepwise bring argument relevance to web search engines. 1 Introduction What stance should I take? What are the best argu- ments to back up my stance? Information needs of people aim more and more at arguments in favor of or against a given controversial topic (Cabrio and Villata, 2012b). As a result, future information systems, above all search engines, are expected to deliver pros and cons in response to respective queries (Rinott et al., 2015). Recently, argument mining has become emerging in research, also be- ing studied for the web (Al-Khatib et al., 2016a). Such mining finds and relates units of arguments (i.e., premises and conclusions) in natural lan- guage text, but it does not assess what arguments are relevant for a topic. Consider the following arguments (with implicit conclusions) for a query “reasons against capital punishment”: Example a1. “The death penalty legitimizes an irreversible act of violence. As long as human jus- tice remains fallible, the risk of executing the in- nocent can never be eliminated.” Example a2. “Capital punishment produces an unacceptable link between the law and violence.”1 While both arguments are on-topic, a1 seems more clear, concrete, and targeted, potentially ma- king it more relevant. First approaches to assess argument quality exist (see Section 2). However, they hardly account for the problem that argument quality (and relevance in particular) is often per- ceived subjectively. Whether a3, e.g., is more rele- vant than a1 or less depends on personal judgment: Example a3. “The death penalty doesn’t deter people from committing serious violent crimes. The thing that deters is the likelihood of being caught and punished.” In this paper, we study from a retrieval perspec- tive how to assess argument relevance objectively, i.e., without relying on explicit human judgments. Following argumentation theory, we see relevance as a dialectical quality that depends on how benefi- cial all participants of a discussion deem the use of an …

Content truncated to 500 words