A Cross-lingual Annotation Projection-based Self-supervision Approach for Open Information Extraction

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 741–748, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP A Cross-lingual Annotation Projection-based Self-supervision Approach for Open Information Extraction Seokhwan Kim, Minwoo Jeong†, Jonghoon Lee, Gary Geunbae Lee Department of Computer Science and Engineering, Pohang University of Science and Technology {megaup, stardust, jh21983, gblee}@postech.ac.kr Abstract Open information extraction (IE) is a weakly supervised IE paradigm that aims to extract relation-independent informa- tion from large-scale natural language documents without significant annotation efforts. A key challenge for Open IE is to achieve self-supervision, in which the training examples are automatically ob- tained. Although the feasibility of Open IE systems has been demonstrated for En- glish, utilizing such techniques to build the systems for other languages is problem- atic because previous self-supervision ap- proaches require language-specific knowl- edge. To improve the cross-language portability of Open IE systems, this paper presents a self-supervision approach that exploits parallel corpora to obtain training examples for the target language by pro- jecting the annotations onto the source lan- guage. The merit of our method is demon- strated using a Korean Open IE system developed without any language-specific knowledge. 1 Introduction The objective of information extraction (IE) is to generate structured information representing se- mantic relationships among a set of arguments from natural language documents. Although many supervised machine learning approaches have been successfully applied to IE tasks, appli- cations of these approaches are still limited be- cause large amounts of training data are required † Now at Microsoft Bing to achieve good extraction results. Because man- ual annotation for training examples is very expen- sive, weakly-supervised techniques to learn the IE system without significant annotation efforts have been sought (Zhang, 2004; Chen et al., 2006). Open IE is an alternative weakly-supervised IE paradigm (Banko et al., 2007). The goal of Open IE is to yield both domain-independent and relation-independent extractions from a large amount of natural language text without requiring hand-crafted rules or hand-annotated training ex- amples. A key challenge to implementing Open IE is to learn extractors without manually anno- tated training examples. Self-supervised learning approaches have allowed Open IE systems such as TextRunner (Banko et al., 2007) and WOE (Wu and Weld, 2010) to extract relations from large- scale English text with automatically annotated training examples obtained using external knowl- edge. However, applying the self-supervision ap- proaches adopted by previously reported Open IE systems to build a new system is problematic in languages other than English, because these approaches mainly depend on language-specific knowledge for English. For example, TextRun- ner obtains training examples from the English Penn Treebank by triggering a set of hand-written heuristics denoting syntactic structural constraints to decide whether or not a given instance has a semantic relationship. To learn an extrac- tor for a new language, this approach requires a syntactically annotated corpus and language- specific heuristics for the target language. WOE achieves self-supervised learning of Open IE by using heuristic matches between attribute values in Wikipedia …

Content truncated to 500 words