Information Extraction across Linguistic Barriers

Information extraction (IE) systems have been tailored to extract fixed target information from documents in a fixed language. In order to be truly useful for information analysts, the target information must be user-definable and the source documents should cover multiple languages. We will map out the path toward such open-target multilingual IE systems, identifying necessary technological breakthroughs along the path. We also discuss a Japanese-English named entity extraction system under development, which represents a case of the next step along the path. Introduction: Toward Multilingual Information Extraction Systems The natural language processing field has witnessed a rapid development of the information extraction (IE) technology since the early 90’s, driven by the series of Message Understanding Conferences (MUC’s) the government-sponsored TIPSTER program. 1 This technology enables a rapid, robust, and automatic extraction of certain predefined target information from real-world on-line texts or speech transcripts accessible through computer networks. Information analysts, whose task is to keep track of changing states of affairs about particular topics such as microelectronic products and international terrorist activities, can use the IE technology for accomplishing their tasks more efficiently and effectively. IE systems, however, have so far been tailored to extract fixed target information from documents in a fixed language. In order for the IE technology to be truly useful for information analysts, the target information must be user-definable, or ’open,’ and it should also obtain information from documents in multiple languages. To appear in the Working Notes for the Workshop on Cross-Language Text and Speech Retrieval, AAAI Spring Symposium Series, Stanford, CA, 1997. 1 The TIPSTER web page is at http://www.tipster.org In this paper, we will map out the path toward such open-target multilingual IE systems, identifying necessary technological breakthroughs along the path. We also discuss a Japanese-English named entity extraction system under development, which represents an IE system that lies in the immediate future along the path. Information Extraction Technology Given a set of source documents, the input to an IE system is a description of the target information type, and the output is a set of target information instances found in the source documents. The target information, typically of the form "who did what to whom where when," is extracted from natural language sentences or formatted tables, and fills parts of predefined template data structures with slot values. Partially filled template data objects about the same entity or event instances are then merged to create a network of related data objects. These template data objects depicting instances of the target information are the raw output of information extraction, ready for a wide range of applications such as database updating or summary generation. An IE output can also take the form of SGML or other types of markups on the source documents, which need not go through template data objects. IE systems, however, have so far been tailored to extract fixed target information from documents in a fixed language, through a short but intense customization work by IE system experts. We can characterize these first-generation IE systems as systems for monolingual information extraction with closed target information. See Figure 1. Analysis of MUC evaluation results has led to a clearer understanding of the strengths and weaknesses of the technology. The MUC-6 evaluation results (Sundheim 1995) showed that name recognition is largely a solved problem, with a humanlike performance in the same task (higher F-measure around 95%). This IE subtask is thus ready for real-world applications. Extracting template entities (higher F-measure around 75%) and recognition 111 From: AAAI Technical Report SS-97-05. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved.