Masaki Eto

Since machine-readable documents have become widespread, some recent studies have proposed retrieval methods using a combination of citation linkage and its context. In the case of co-citation linkage, there have been attempts to discern 'strong' co-citations from 'weak' ones by examining the positions of citations in a document. However, this promising concept has not yet been sufficiently evaluated, and it remains unclear whether search performance is significantly improved. Therefore, this paper explores the effects of using co-citation context more deeply and more widely by comparing the search performance of six retrieval methods, which differ as to whether co-citation context and normalization using cited frequency are used. For empirically evaluating the effects, a special test collection was created from CiteSeer Metadata, and the search performances of the six retrieval methods were compared by two IR metrics (AP and nDCG). The main conclusions of this paper are: (1) co-citation context has a positive effect on co-citation searching; (2) the normalization technique using cited frequency is useful for context-based co-citation searching; (3) approaches of using co-citation context tend to affect the characteristics of search performance.

The co-citation measure is widely used to retrieve similar documents. The method is based on the premise that all degrees of similarity between a pair of co-cited papers have an equal weight in a single citing paper. In addition, the conventional measure is binary in that only whether two papers are "co-cited" or "not co-cited" is considered in the calculation process. In order to estimate similarities between co-cited papers more precisely, the author proposes a new co-citation measure based on structures of citing papers, i.e., we focus on structural distances between the positions where two co-cited papers appear in a citing paper. By the proposed measure, each co-citation is classified into four types: "different paragraph", "same paragraph", "same sentence" and "enumeration (i.e., a set of references to papers is included in a single sentence of the citing paper)". To evaluate the effectiveness of the proposed measure, the five typical similarities between co-cited papers that are found by the above four types of co-citation measure were respectively calculated and compared. In the experiment, the degree of calculated similarities gradually became higher with shorter structural distance; the highset one was "enumeration" and the lowest was "different paragraph". The proposed co-citation measure was thus shown to be able to estimate similarities between co-cited papers more precisely.

Purpose: One typical document retrieval method is to use co-citation. The method is based on the premise that the degree of similarity among co-cited papers is equal in a particular paper. The degree is calculated with binary values: "co-cited" or "not co-cited". To improve upon this method, the author proposes a multivalued co-citation measure based on semantic distance between co-cited papers. Methods: To determine the distance between citations, the author measured two machine. parseable relationships (location and citing words) between places where papers are cited. In order to evaluate the proposed method, we identified two categories of co-citation: a group with strong relationships indicating "enumerated co-citation" (papers cited within one statement) and a group with weak relationships showing "non enumerated co-citation". Similarities within each group were calculated and compared using the CiteSeer dataset and 6 major similarity indicators. Results: All of the similarity indicators showed that the degree of "enumerated co-citation" is higher than "non enumerated co-citation". Consequently, it became clear that the proposed co-citation measure can be used to distinguish the strength of co-citation more precisely and that it can be applied to large-scale document collections.