Finding experts is useful in such cases: seeking for consultants, collaborators, and speakers. It also provides a source of information to supplement or complement academic sources including metadata [7], thus, receives increased attention in recent years. However, identification resolution is not considered significantly even though this research topic mainly deals with persons. Many studies concentrate only on string-based person names [1] [2] [5] [6]. Semantic Web can be one of competent solutions for managing identified experts through underlying URI scheme. Another consideration is to guarantee reliability on the results of the task. Deep analysis based on full text documents is needed in that topically-classified documents in high precision ensure finding the right persons for each topic. On the basis of these considerations, we propose an experts-finding method based on identity resolution and full text analysis, and further extract topic-centric information such as 'Topic Trends' and 'Institutions by Topic'. Chapter 2 indicates several previous studies. Chapter 3 explains how to acquire topic-centric information based on a Semantic Web Framework.
The sources for finding experts are various: documents, programs, e-mails, databases, citations, communities and so on. Finding expertise information from e-mails with four simple binary association methods was proposed by [1]. [5] investigated the expertise of users and experts by combining information retrieval techniques. However, such e-mails and communities are insufficient to extract the right experts for a specific topic because they give clues about only relationship and context. An experts-finding study based on full text documents related with persons and on a set of terms in them was introduced [2]. It extracts similar experts by measuring similarity between term vectors. However, it is not able to indicate which topics are related with experts, but only provides a bundle of persons as the results. ExpertFinder [6] recommends persons with a lot of documents for a given topic. A keyword phrase is used to retrieve relevant documents, but the results are unsatisfactory because reasonable candidates are not listed within the top three or four candidates in most cases. Its slow response time and incorrect relationship between persons and documents are also problems. Another interesting study, performed by [8], introduced three innovative points: document authority in terms of their PageRanks, co-occurrence model, and multiple levels of associations between experts and query terms. It finds variants in experts' names for identity recognition, but failed to identify different persons with the same name uniquely. OntoFrame is a Semantic Web-based service which provides academic research information for supporting R&D activities [3]. Its two main components are URI server and OntoReasoner (inference engine). The latter interacts with user interfaces through receiving SPARQL queries and returning XML results. We introduce SPARQL rather than inflexible SQL because it is easy to construct queries with only knowledge on ontology schema. OntoReasoner also expands knowledge in ways of forwardchaining inference. The URI server has several functions: ontology schema parsing and loading, DB schema creation, ontology instance loading, and RDF triple generation as shown in figure 1. When a new instance is inserted into the server, triple generator makes triples for the instance. The triples are then stored in RDF triple store, and further would be referred by OntoReasoner. OntoFrame distinguishes from other academic research information services such as CiteSeer (http://citeseer.ist.psu.edu/) and Google Scholar (http://scholar.google.com/) because it provides information acquired by inference beyond metadata. 'Persons by Topic', 'Topic Trends', and 'Social Network' are representative information served by OntoFrame.
The Open Archives Initiative (OAI, http://www.openarchives.org/) develops and promotes interoperability standards that aim to facilitate the efficient dissemination of content. CiteSeer (http://citeseer.ist.psu.edu/oai.html) also supports OAI, and thus allows downloading its own open access metadata which includes title, authors, publication year and so on. Identity resolution is an obligatory task for transforming string-based data to semantic data [4]. Various forms of institution names in the metadata are mapped to a set of normalized institution names1 , e.g. "U. Kassel" and "University of Kassel." We also identify different persons with the same name. There are a few metadata fields available for distinguishing authors such as affiliation, e-mail, and co-authors. It is possible to determine whether two authors with the same name are different or not using their affiliations and e-mails. However, affiliation and e-mail fields are not obligatory in many cases including CiteSeer metadata. Co-authorship information plays an important role in resolving identity problems because co-author field is usually filled up in metadata, and further many authors maintain co-authorship relation regardless of affiliation change. We consider two authors with the same name as the identical person when they share the identical co-author(s), otherwise they remain as different persons. 'sameAs' relation would compensate the short coverage of this method based on co-authorship. All of their information, including papers and topics, will be merged as one when we connect two authors with 'sameAs' relation later.
After identity resolution, we assign URI for each entity; for example, paper "A Bayesian Multiple Models Combination Method for Time Series Prediction" with 'http://www.kisti.re.kr/isrl/ResearchRefOntology#ART_00000000000000458673', topic "markov model" with 'http://www.kisti.re.kr/isrl/ResearchRefOntology#TOP_00000000000000046687' and person "V.
Petridis" with 'http://www.kisti.re.kr/isrl/ResearchRefOntology#PER_00000000000000128292'.
Extracting topics from papers is the most basic task to acquire topic-centric experts.
As full text documents as well as metadata of CiteSeer are available, we use the documents. Extracted topics are assigned to each paper. The followings explain the stages of the extraction as shown in figure 2; First, indexer extracts index terms from a given document. Second, the terms are matched with topic keywords in topic index DB 2 . Third, successfully matched terms are ranked by the following algorithms, and then we select top-n (currently, five) topics for the input document.
Many factors can be considered for finding experts: the number of papers, impact factor of sources, the degree of citations, hub persons in social network and so on. Currently, we take into account only the number of papers for several reasons. A great portion of source field in CiteSeer open access metadata has no information. Citation information also may be incomplete when compared with CiteSeer service page. We also do not consider social network because prosperous co-authorship with other persons does not always guarantee specialty on a topic. Acquiring topic-centric experts on OntoFrame requires querying to RDF triple store based on DBMS. 'Persons by Topic' is retrieved directly from the database through SPARQL query (shown as follows) and automatic SPARQL-to-SQL conversion. The query searches papers (?accomplishment) of which topic area is topicTerm, and then retrieves authors (?person) of the papers. Figure 3 shows backward chaining flow starting from topicTerm. 'createdByPerson' is one of derived properties induced by user-defined inference rules. It reduces the distance of backward path to find 'Persons by Topic' in ways that go through directly to 'Person' rather than without passing through 'CreatorInfo' (the dotted line in figure 3). After retrieving persons, OntoReasoner performs postprocessing for ranking them by descending order of the number of their own papers.
We gathered 114,337 papers (2000 ~ 2006) from CiteSeer open access metadata. They include 161,853 persons and 17,093 institutions. 160,568 topic keywords 3 were extracted from titles and abstracts. Average consuming time for extracting maximum 5 topics from a paper is about 1.6 seconds. Within three seconds are enough to generate an entity page including 'Persons by Topic' on OntoFrame4 .