Reviewed by Petra Gieselmann, University of Karlsruhe

Question answering (QA) is an upcoming research topic: the rise of the internet and the success of document retrieval has generated customer demand for good QA engines. Intelligent technologies are therefore used so that modern QA systems can combine complex natural language processing techniques, sophisticated linguistic representations, and advanced machine learning methods to find exact answers to a variety of natural language questions. The eighteen papers in this volume focus on present QA research and provide an extensive overview of this field. The volume is divided into six sections, which address different viewpoints and various approaches.

Part 1 introduces various QA approaches: the opening chapter by Dan Moldovan, Marius Paşca, and Mihai Surdeanu presents the Language Computer Corporation’s QA system, which was one of the most successful participants in the Text Retrieval Conference evaluation (TREC; an annual event sponsored by the United States National Institute of Standards and Technology). Relying on deep linguistic processing combined with logic theorem, the authors illustrate how multiple resources can be used to support a wide range of questions. In ‘A statistical approach for open-domain question answering’, Abraham Ittycheriah describes a completely different approach that features machine learning techniques such as maximum entropy modeling. Jose Vicedo and Antonio Ferrández present a detailed analysis of the problem of coreference resolution and its impact on QA.

Part 2 explores question processing. In ‘Questions and intentions’, Sanda Harabagiu explains the difficulties of question understanding, which includes the implications of user intentions. Using metonymies and the implications of user intentions to generate predictive and implied questions may reveal ways to enhance the efficiency, efficacy, and user satisfaction of QA results. Tomek Strzalkowski, Sharon Small, Hilda Hardy, Paul Kantor, Wu Min, Sean Ryan, Nobuyuki Shimizu, Liu Ting, Nina Wacholder, and Boris Yamron present their high-quality interactive QA, in which the user and the system first negotiate the scope and shape of the information and then cooperate in finding the answer within a large repository of unstructured data. Finally, Brigitte Grau, Olivier Ferret, Martine Hurault-Plantet, Christian Jacquemin, Laura Monceaux, Isabelle Robba, and Anne Vilnat examine linguistic variation within questions. Since paraphrases can take place either at the term or sentence level, an extensive use of natural language processing (NLP) components is essential.

Part 3 focuses on QA as a kind of information retrieval. The approach taken by Laszlo Grunfeld and Kui-Lam Kwok uses pattern matching as well as some information retrieval techniques, which work well for factoid questions, even without a deep linguistic analysis. However, Grunfeld and Kwok’s technique is inferior to sophisticated systems with knowledge inference and deep NLP. Failures are often due to the fact that answers are part of longer sentences and cannot be extracted correctly. Charles Clarke, Gordon Cormack, Thomas Lynam, and Egidio Terra present their passage selection approach to QA: the MultiText QA system first retrieves passages within a corpus that are related to the topic of the question and might, therefore, contain the answer. Then an answer is selected out of these passages by considering answer type and candidate redundancy. In ‘Query modulation for web-based QA’, Dragomir Radev, Hong Qi, Zhiping Zheng, Sasha Blair-Goldensohn, Zhu Zhang, Weiguo Fan, and John Prager evaluate QA using statistical models, which learn the best transformations to paraphrase a natural language question in preparation for a query to be sent to a search engine.

Part 4 deals with answer extraction. John Prager, Jennifer Chu-Carroll, Eric Brown, and Krzysztof Czuba present their approach to QA through predictive annotation: they index a database with extended named entities that might become answers in factoid questions. The next chapter by Rohini Srihari, Wei Li, and Xiaoge Li describes how information extraction can be used to support QA. Using multiple levels of information extraction, they generate a precise answer to a factoid question. Finally, Abdessamad Echihabi, Ulf Hermjakob, Eduard Hovy, Daniel Marcu, Eric Melz, and Deepak Ravichandran address the issue of how to select an answer string. Evaluating knowledge-based, pattern-based, and statistics-based answer selection, they claim to improve the individual answer selection modules and combine their outputs in a maximum entropy model.

Part 5 explains methods and techniques employed to evaluate the performance of QA systems. Ellen Voorhees looks at the evaluation methods used within the QA track of TREC, the first large-scale evaluation of open-domain QA systems. William Hersh describes two evaluation initiatives: the interactive QA track on TREC and user studies in the medical domain. He argues that studies with users are essential for understanding the efficacy of QA systems. The last chapter of this section by William Ogden, James McDonald, Philip Bernick, and Roger Chadwick explores habitability in QA systems.

Finally, Part 6 deals with perspectives on QA. Steven Maiorano discusses QA as a technology for intelligence analysis. He believes that QA can help various analysts to get answers for specific questions in a dialog-based system. The next article by Ellen Riloff, Gideon Mann, and William Phillips explains the need for a large collection of question answer pairs to train statistical models for QA. Therefore, they developed a method to automatically generate question answer pairs from normal text. Finally Mark Maybury deals with new directions in QA. He discusses the need for answering broader and deeper questions, extracting answers from varied sources, and supporting users with different backgrounds.

This volume provides a good overview of current research in QA. It will be interesting for students just beginning to look at QA as well as for researchers wanting to gain insights into interactive QA.