Catálogo de publicaciones - libros

In order to help readers grasp key information from web pages within limited time, a novel intelligent user interface for webpage navigation: Keyphrase Navigation Map (KNM) is proposed in this paper, which presents a map to the readers, assisting them to navigate their webpage. First, the key phrases are generated from the webpage and ranked by the keyphrase extraction engine, then they are clustered by keyphrase relevancy, and finally they are shown on the thumbnail according to their relative coordinates in the full page. The usability test proves that the KNM which presents an overview of the corresponding web page to readers can really help them to find information they are interested in effectively.

Query expansion is an important technique for achieving higher retrieval performance. However, a good framework for evaluating this technique in isolation does not exist because the effect of query expansion depends on the quality of the initial query. Feature quantities that characterize the quality of the initial query are defined in this study. Correlations between these quantities and retrieval performance were analyzed using the NTCIR-4 web test collection.

Recently, advanced multimedia applications, such as geographic information system, and content-based image/video retrieval system, require the efficient processing of -nearest neighbor queries with semantic predicates as well as visual predicates. In this paper, we propose an integrated index structure, so-called SPY-TEC+, that provides an efficient method for indexing visual and semantic feature information at the same time using the SPY-TEC and the signature file. We also propose an efficient incremental nearest neighbor algorithm for processing the -nearest neighbor queries with visual and semantic predicates on the SPY-TEC+.

Transliterating foreign entities into Chinese is usually done through direct-style approach. The direct-style approach transliterates each syllable rendered from foreign terms into Chinese directly. Not every syllable can be rendered. An approach utilizing contextual information for term transliteration is proposed in this paper to attack this problem. Traditionally, evaluating transliteration performances always uses character and word error rates. However, many transliteration variants of the same term always found. Validating the generated results using Web corpora is more suitable and is proposed in this paper. Using the proposed evaluation method, experiments on term transliteration were conducted. From the experimental results show that taking contextual information is helpful to term transliteration and validating the generated results using Web corpora can provide more concrete evidences to transliteration evaluation.

Orthographic varieties are common in the Japanese language and represent a serious problem for Japanese information retrieval (IR), as IR systems run the risk of missing documents that contain variant forms of the search term. We propose two different strategies for handling orthographic varieties: pronunciation or yomi-based indexing and “Fuzzy Querying”, comparing katakana terms based on edit distance. Both strategies were integrated into our multiple index and fusion system [1] and tested using two different test collections, newspaper articles (Mainichi Shimbun ’98) and scientific abstracts (NTCIR-1), to compare their performance across text genres.

It is common that representative words in a document are identified and discriminated by their statistical distribution of their frequency statistics. We assume that evaluating the confidence measure of terms through content-based document analysis leads to a better performance than the parametric assumptions of the standard frequency-based method. In this paper, we propose a new approach of term weighting method that replaces the frequency-based probabilistic methods. Experiments on Naïve Bayesian classifiers showed that our approach achieved an improvement compared to the frequency-based method on each point of the evaluation.

This paper proposes a fast and effective question-answer system for encyclopedia domain using a new answer indexing method. We define about 160 answer types. The indexer generates AIU(Answer Index Unit) structures between answer candidates and content words within LF(Logical Form) and sentence boundary. We select essential terms among question terms using syntactic information for ranking the answer candidates. Experiments show our new method is good for the encyclopedia question-answering system.

The medical thesaurus, MeSH, has been used to index medical documents. A Korean MeSH also has been developed, but it does not include many of the synonymous translations for the English terms. The coverage of synonymous translation is important to index medical documents correctly. In this paper, we propose an approximate phrase match method to extract synonymous translations from Korean medical documents, where parentheses are used to include English terms, or English keywords are used in the keyword field. The approximate phrase match is to handle the unregistered terms in a bilingual dictionary. The empirical evaluation showed that the proposed methods are very effective to compile translation phrase pairs.

Most proteins interact with other proteins, cells, tissues or diseases. They have biological functions and can be classified according to their functions. With the functions and the functional relations of proteins, we can explain many biological phenomena and obtain answers in solving biological problems. Therefore, it is important to determine the functions of proteins. In this paper we present a protein function classification method for the function prediction of proteins. With human proteins assigned to GO molecular function terms, we measure the similarity of proteins to function classes using the functional distribution.

Most of existing open domain question answering systems predefine the conceptual category to which answers can belong. So, they cannot generate appropriate answers in every case or must use a strategy that handles exceptions when the concept requested in the question is not prepared in the system. In this paper, we suggest a flexible strategy that can generate the candidate answers which correspond to any nominal target concepts. The proposed question answering system is equipped with general patterns that can extract hyponyms of the nominal target concept with their confidence scores. Therefore, it can create a set of candidate answers from the dynamically generated ontology when a user requests any nominal concept.