Porting an Open Information Extraction System from English to German

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 892–898, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Porting an Open Information Extraction System from English to German Tobias Falke† Gabriel Stanovsky‡ Iryna Gurevych† Ido Dagan‡ †Research Training Group AIPHES and UKP Lab Computer Science Department, Technische Universit¨at Darmstadt ‡Natural Language Processing Lab Department of Computer Science, Bar-Ilan University Abstract Many downstream NLP tasks can benefit from Open Information Extraction (Open IE) as a semantic representation. While Open IE sys- tems are available for English, many other languages lack such tools. In this paper, we present a straightforward approach for adapt- ing PropS, a rule-based predicate-argument analysis for English, to a new language, Ger- man. With this approach, we quickly obtain an Open IE system for German covering 89% of the English rule set. It yields 1.6 n-ary extrac- tions per sentence at 60% precision, making it comparable to systems for English and readily usable in downstream applications.1 1 Introduction The goal of Open Information Extraction (Open IE) is to extract coherent propositions from a sentence, each represented as a tuple of a relation phrase and one or more argument phrases (e.g., born in (Barack Obama; Hawaii)). Open IE has been shown to be useful for a wide range of semantic tasks, including question answering (Fader et al., 2014), summariza- tion (Christensen et al., 2013) and text comprehen- sion (Stanovsky et al., 2015), and has consequently drawn consistent attention over the last years (Banko et al., 2007; Wu and Weld, 2010; Fader et al., 2011; Akbik and L¨oser, 2012; Mausam et al., 2012; Del Corro and Gemulla, 2013; Angeli et al., 2015). Although similar applications of Open IE in other languages are obvious, most previous work focused 1Source code and online demo available at https://github.com/UKPLab/props-de on English, with only a few recent exceptions (Zhila and Gelbukh, 2013; Gamallo and Garcia, 2015). For most languages, Open IE systems are still missing. While one could create them from scratch, as it was done for Spanish, this can be a very laborious pro- cess, as state-of-the-art systems make use of hand- crafted, linguistically motivated rules. Instead, an alternative approach is to transfer the rule sets of available systems for English to the new language. In this paper, we study whether an existing set of rules to extract Open IE tuples from English de- pendency parses can be ported to another language. We use German, a relatively close language, and the PropS system (Stanovsky et al., 2016) as examples in our analysis. Instead of creating rule sets from scratch, such a transfer approach would simplify the rule creation, making it possible to build Open IE systems for other languages with relatively low ef- fort in a short amount of time. However, challenges we need to address are differences in syntax, dis- similarities in the corresponding dependency rep- resentations as well as language-specific phenom- ena. Therefore, the existing rules cannot be directly mapped to the German part-of-speech and depen- …

Content truncated to 500 words