The vast amounts of information readily available on the World Wide Web can be effectively used for question answering in two fundamentally different ways. In the federated approach, techniques for handling semistructured data are applied to access Web sources as if they were databases, allowing large classes of common questions to be answered uniformly. In the distributed approach, largescale text-processing techniques are used to extract answers directly from unstructured Web documents. Because the Web is orders of magnitude larger than any human-collected corpus, question answering systems can capitalize on its unparalleled-levels of data redundancy. Analysis of real-world user questions reveals that the federated and distributed approaches complement each other nicely, suggesting a hybrid approach in future question answering systems.

To address the problem of information overload in today's world, we have developed Start,anatural language question answering system that provides users with high-precision multimedia information access through the use of natural language annotations. To address the di#culty of accessing large amounts of heterogeneous data, we have developed Omnibase, which assists Start by integrating structured and semistructured Web databases into a single, uniformly structured "virtual database." Our ultimate goal is to develop a computer system that acts like a "smart reference librarian," and we believe we have laid a firm foundation for achieving our goal. This paper describes our current implemented system and discusses future research directions.

AnswerFinder combines lexical, syntactic, and semantic information in various stages of the question answering process. The candidate sentences are preselected on the basis of (i) the presence of named entity types compatible with the expected answer type, and (ii) a score combination of the overlap of words, grammatical relations, and flat logical forms. The candidate answers, in turn, are extracted from (i) the set of compatible named entities and (ii) the output of a logical-form pattern matching algorithm. 1

Most question answering systems narrow down their search space by issuing a boolean IR query on a keyword indexed corpus. This technique often proves futile for definitional questions, because they only contain one keyword or name. Thus, an IR search for only that term is likely to produce many spurious results; documents that contain mentions of the keyword, but not in a definitional context. An alternative approach is to glean the corpus in pre-processing for syntactic constructs in which entities are defined. In this thesis, I describe a regular expression language for detecting such constructs, with the help of a part-of-speech tagger and a named-entity recognizer. My system, named CoL. ForBIN, extracts entities and their definitions, and stores them in a database. This reduces the task of definitional question answering to a simple database lookup.

In this paper we present an evaluation of overlap-based measures of similarity for sentences in the same language.

Synonymy in word and expression is a problem in many natural language analysis tasks. While single word resources for synonymy exist such as thesauri and Wordnet, there are few resources for multiple word synonyms or paraphrases. We attempt to implement an unsupervised technique for automatically extracting paraphrases from aligned mono-lingual corpora. By comparing relations between similar words in two aligned sentences, we can extract paraphrases. Paraphrases which occur often and in di#erent contexts are scored higher. While the final results were encouraging, the low density of paraphrases necessitates a much larger set of data. Such data is hard to obtain, because of the strict requirements of the alignment tool.

are stored in PDF, with the report number as filename. Alternatively, reports are available by post from the above address. Orientador: Júri:

To address the problem of information overload in today's world, we have developed Start, a natural language question answering system that provides users with multimedia information access through the use of natural language annotations. In order to harness the potential of knowledge sources on the World Wide Web, we have developed Omnibase, a virtual database that provides uniform access to Web resources. Our ultimate goal is to develop a computer system that acts like a "smart reference librarian," and to a large extent, we have accomplished our goal. However, expanding our system's domain of knowledge is a time-consuming task that requires trained individuals. This paper describes several research directions aimed at overcoming the limitations of our current technology.

We present an approach to summarisation based on the use of a question answering system to select the most relevant sentences. We used AnswerFinder, a question answering system that is being developed at Macquarie University. The sentences returned by AnswerFinder are further re-ranked and collated to produce the final summary. This system will serve as a baseline upon which we intend to develop methods more specific to the task of questiondriven summarisation. 1