Briefly Noted

This book is a diverse collection of ten presentations given at an International Summer School on Information Extraction in Rome, 1997. The goal of information extraction (IE) is selective, task-driven interpretation of text narrative in order to fill out templates with information about a particular scenario. I was disappointed to find that only five of the articles were actually about IE research. The other half of the articles addressed issues peripheral to IE, such as information retrieval (IR) and text classification. The first two articles are by Yorick Wilks and by Ralph Grishrnan, who are prominent IE researchers in the UK and the US, respectively. Each gives a high-level discussion of IE, its successes and limitations. Wilks makes the observation that IE's strength comes from its modular architecture. Individual modules such as part-of-speech tagging or morphology analysis can be constructed and optimized independently and reused in a variety of applications. He sees the primary limitation of IE to be the template representation that restricts the type of information that can be extracted. Grishman describes the typical architecture of IE systems whose modules include lexical analysis, name recognition, shallow syntactic parsing, task-specific pattern matching, coreference analysis, event merging, and finally template generation. He identifies the main challenges to IE as the cost of adapting a system to a new domain or scenario and a ceiling on performance, which is closely related to the issue of knowledge acquisition and difficulty handling complex syntactic structures. Three other articles deal with more specialized topics within IE. Robert Gaizauskas, Kevin Humphreys, Saliha Azzam, and Yorick Wilks describe a system for multilingual IE with some language-independent modules that are indexed by language-specific lexicons. Roberto Basili and Maria Teresa Pazienza discuss corpus-driven lexical acquisition, in particular for the "foreground" lexicon of words that support a particular IE task. Branimir Boguraev and Christopher Kennedy present work in technicalterm recognition and how this can be a step towards document summarization. The remaining articles concern IR, text dassification, or heterogeneous database techniques, and are only tangentially related to IE. Gregory Grefenstette presents an NLPbased strategy for suggesting additional IR query terms to a user. Alan Smeaton gives a tutorial on uses of NLP in IR. Nicola Guarino discusses formal ontologies and how these can enhance IR with semantic matching. Filippo Neri and Lorenza Saitta give a tutorial on machine learning that briefly touches on text classification. Sophie Cluet describes database techniques for querying semi-structured Web pages.--Stephen Soderland, Children's Hospital, Seattle