Masaki Eto

We propose a method for a comprehensive listing of comics and manga authors by combining a VIAF dataset, which internationally shares authority information created in each country, and a national bibliography. Specifically, it combines the dataset based on the descriptions of VIAF and the set of comics and manga authors extracted from comic titles in the Japanese national bibliography. The proposed method revealed that there are Japanese comics and manga authors who are not recognized as such in the Japanese national bibliography and that comics written by many Japanese comics authors are distributed only in Japan.

This study addressed the automatic classification of Japanese manga held by public libraries. Holdings of 4,681 public libraries and similar facilities were investigated, and 29,795 manga titles were identified. Hierarchical clustering was applied to 631 titles that were each held by more than one hundred libraries. Five clusters were identified in the upper hierarchy. Principal coordinate analysis and a manual examination of individual titles were performed to identify the common characteristics of the works in each cluster. The results suggest that the proposed method offers a novel approach to large-scale classification of manga titles.

This paper proposes a “rough co-citation”, which is a measure of relationship to expand co-citation networks so as to include new relevant documents. A rough co-citation relationship is a linkage between a pair of documents which are cited by two other documents in a similar citation context. The linkage strength of a rough co-citation relationship may be weaker than the original co-citation relationship, because a rough co-citation relationship is determined by citations in two separate documents. Rough co-citation linkages, however, may yield new relevant documents that are not identified by the original co-citation linkages. For example, the rough co-citation can identify relevant documents that are published after the citing document of the original co-citation becomes public. This study conducted IR experiments to evaluate the search performances of retrieval methods using the co-citation networks expanded by the rough co-citation relationships. Specifically, the random walk with restart, which is one of the latest graph search algorithms, is applied to the expanded and original co-citation networks. Scores of the normalized discounted cumulative gain (nDCG@K) are then compared. The results indicate that the search performance of the method using the expanded network outperforms a baseline method using the original network.

In the field of academic document search, citations are often used for measuring implicit relationships between documents. Recently, some studies have attempted to extend co-citation searching. However, these studies mainly focus on comparisons of traditional co-citation and extended co-citation search methods combination effects of word-based and extended co-citation search algorithms have not yet been sufficiently evaluated. This paper empirically evaluates the search performance of the combination search by using a test collection comprising about 152,000 documents and a metric 'precision at k.' The experimental results indicate that the combination search outperforms two baseline methods: a word-based search and a combination search of word-based and traditional co-citation search algorithms.

This paper proposes a graph-based retrieval technique on a weighted co-citation network, which allows users to find more relevant documents easily from the co-citation network. More specifically, the random walk with restart technique is applied to a weighted graph of documents, in which the degree of each edge weight is measured by the number of co-citation documents and the strength of the co-citation context both obtained by parsing the full text of the citing documents. To evaluate its effectiveness empirically, a special test collection was created from the Open Access Subset of PubMed Central, and the search performance of the proposed method was compared with traditional co-citation searching by "precision at k." The experimental results indicate that the proposed method tends to retrieve much more relevant documents without reducing precision.

This paper proposes a measure that uses a spread co-citation relationship for document retrieval. To clarify whether this proposed measure has potential for enhancing the search performance of co-citation searching, two retrieval methods are evaluated: one uses the relationship directly the other incorporates the co-citation context. Experiments with a special test collection comprising about 152,000 documents are conducted. Results indicate that this relationship tends to be able to detect relevant documents which are undetectable using a traditional co-citation relationship, and that using context has a positive effect to reduce the number of noise documents. © 2012 ACM.

The purpose of this research is to assign automatically appropriate keywords to the reference records on the basis of the sentences of questions and answers in the reference records. Former researches extracted appropriate words automatically as keywords from the reference records, but this method could assign only the words in the texts. This research proposes a method of assigning of keywords using titles and snippets of web pages. The method consists of three processes: (1) the process of extracting keywords as search terms from the reference records by the machine learning, (2) the process of acquiring titles and snippets from Google, and (3) the process of acquiring the keywords by estimation of weights. The result showed that the keywords were assigned successfully with precision ratio of 31.4% and recall ratio of 61.4%.

In this paper the author proposes a sophisticated document retrieval method using context based co-citation relationship. To evaluate the effectiveness of the proposed method, two experiments were conducted. The first experiment is to compare documents retrieved by word with the ones retrieved by co-citation relationship. The second one is to compare the proposed method with the traditional method by using binary co-citation relationship based on two typical metrics (MAP and nDCG), and by analyzing documents ranked top 10. The experiments showed that the proposed method will (1) retrieve relevant documents which cannot be retrieved by using word, (2) rank retrieved relevant documents more adequately, and (3) rank relevant documents highly, even though the number of the frequencies of their co-citation is few.

In recent years Kansei retrieval systems have been developed to be practical use. We have developed the retrieval system using Kansei parameters based on online book reviews of children's books. The relation between parameters assigned manually and the words in online book-reviews has yet to be analyzed enough, however. Moreover, how to set appropriate value of Kansei parameters based on the words in online book reviews has not been theoretically proved. To resolve such problem, we have conducted an experiment. In this experiment, Kansei parameters which are consistent with the value gained manually were assigned based on "the words, which represent human feelings, impressions or atmosphere of books" using the machine learning approach. We have used 1605 book reviews to assign the parameters automatically. The result showed that : 1)using the words in book reviews is useful to assign parameters automatically, 2)using about 30 frequent words is more effective than using all the words in book reviews, and 3) the recall ratio is not influenced by the parts of speech of the words extracted from book reviews.

The purpose of this research is to extract appropriate words automatically as keywords of the reference records from the sentences of questions and answers in the reference records. The automatic extraction method consists of two processes: (1) the process of extracting candidate words from given texts, and (2) the process of judging whether each extracted candidate is appropriate or not as a keyword. We focus on the process of (2) using the machine learning approach. In this method as a training data, we have used 42 features such as where each candidate word was, how many times it appeared in the records, and the postpositional particles surrounding the candidate. We then conducted an evaluation experiment using 9,375 words judged to be keywords manually in 507 reference records of Japanese history. The result showed that the keywords were assigned successfully with precision ratio of 56.9% and recall ratio of 48.3%

In this research, we conducted the experiment of automatically classifying the reference records in the Collaborative Reference Database of the National Diet Library into Nippon Decimal Classification (NDC). Three machine learning methods: decision tree, Naïve Bayes, and Support Vector Machine (SVM) methods - were tested on the words included in the sentences of questions and answers in 634 reference records. The results indicate that the using the SMV method and answer sentences were classified successfully compared to others. As for the recall and precision ratio, there was small difference among the methods, and between question and answer sentences. The significant difference, however, was observed in the recall ratio by the NDC category; the result of some category obtained more than 60% recall ratio, which suggested that classifying reference records into NDC by using machine learning methods was effective.

Co-citation, which is a typical similarity indicator among papers, has a premise that the degrees of similarity among cited papers are equal in one citing paper. But the similarity strength can be guessed from the distance between the places where cited papers are shown in a citing paper. In this paper, three distance measures ("physical distance", "co-occurrence of citing words", and "structural distance") are introduced for the purpose and these measures are evaluated.