Inducing Information Extraction Systems for New Languages via Cross-language Projection

Inducing Information Extraction Systems for New Languages via Cross-Language Projection Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 riloff@cs.utah.edu Charles Schafer and David Yarowsky Department of Computer Science Johns Hopkins University Baltimore, MD 21218 {cschafer,yarowsky}@cs.jhu.edu Abstract Information extraction (IE) systems are costly to build because they require development texts, pars- ing tools, and specialized dictionaries for each ap- plication domain and each natural language that needs to be processed. We present a novel method for rapidly creating IE systems for new lan- guages by exploiting existing IE systems via cross- language projection. Given an IE system for a source language (e.g., English), we can transfer its annotations to corresponding texts in a target lan- guage (e.g., French) and learn information extrac- tion rules for the new language automatically. In this paper, we explore several ways of realizing both the transfer and learning processes using off-the- shelf machine translation systems, induced word alignment, attribute projection, and transformation- based learning. We present a variety of experiments that show how an English IE system for a plane crash domain can be leveraged to automatically cre- ate a French IE system for the same domain. 1 Introduction Information extraction (IE) is an important appli- cation for natural language processing, and recent research has made great strides toward making IE systems easily portable across domains. However, IE systems depend on parsing tools and specialized dictionaries that are language specific, so they are not easily portable across languages. In this re- search, we explore the idea of using an information extraction system designed for one language to au- tomatically create a comparable information extrac- tion system for a different language. To achieve this goal, we rely on the idea of cross- language projection. The basic approach is the fol- lowing. First, we create an artificial parallel cor- pus by applying an off-the-shelf machine translation (MT) system to source language text (here, English) to produce target language text (here, French). Or conversely, in some experiments we generate a par- allel corpus by applying MT to a French corpus to produce artificial English. We then run a word alignment algorithm over the parallel corpus. Next, we apply an English IE system to the English texts and project the IE annotations over to the corre- sponding French words via the induced word align- ments. In effect, this produces an automatically an- notated French corpus. We explore several strate- gies for transferring the English IE annotations to the target language, including evaluation of the French annotations produced by the direct projec- tion alone, as well as the use of transformation- based learning to create French extraction rules from the French annotations. 2 Information Extraction The goal of information extraction systems is to identify and extract facts from natural language text. IE systems are usually designed for a specific do- main, and the types of facts to be extracted are de- fined in advance. In this paper, we will focus on the domain of plane crashes …

Content truncated to 500 words