English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools annual

Global: Zendy Free Plan

Global: Zendy Tools monthly

Global: Zendy Plus monthly

s are unavailable to a certain extent or even completely. One of our analysis studies shows that crawled abstracts have 52% of the words on average, compared to their original. It is obvious that this loss of information should adversely affect the accuracy of similarity scores, which is confirmed by our results in Fig. 2. “Complete” in the graph represents the case where the missing terms of the abstracts have manually been filled in, while “Incomplete” is the asis case. The graph demonstrates that the similarity accuracy may be improved by up to 50%, if the loss can somehow be made up. It is noted that this experiment was conducted in the same way as in the previous section. When a paper is cited by another paper, we may expect the same or synonymous terms would appear in their titles and abstracts. Based on this expectation, we propose a new measure, called TL-Rank, that expands the term set of paper by adding the terms from the titles and abstracts of all the papers in citation relationship with that paper. For example, suppose that paper B cites paper C and is being cited by paper A. As the term set for paper B in similarity computation, we use not only the terms from B but also the terms out of A and C. Therefore, TL-Rank can be viewed as a similarity measure that makes use of both text and link information. The similarity computation is performed using the term set augmented by incorporating terms from related papers. This expansion results in a dramatic improvement over “Incomplete” case, as shown by TL-Rank in Fig. 2. It is noted that, for this experiment, we set the ratio of weights assigned to the terms from the original paper, cited papers, and citing papers to 1 : 1 : 1. Also, the weight ratio for titles and abstracts is set to 0.3 : 0.7 as in Fig. 1. In addition to the quantitative results above, Table 2 assesses the three cases from a qualitative comparison perspective. Given one of the most prominent papers [16] in the area of data clustering as a query paper, we let each scheme select 10 papers that they believe are the most similar to the query paper. Boldface words in the table indicate keywords in the area, while the underlined are considered keywords for other areas. We can see that all the selections by “Complete” and TR-Rank cases include the word “clustering” (which is a keyword in the area of data clustering research.) In contrast, the result of “Incomplete” case is contaminated by keywords for other areas such as sequenTable 2 Qualitative evaluation of similarity schemes. tial pattern mining. Notice that there are cross-over research efforts by which some clustering algorithms use sequential pattern mining, and vice versa. This can cause confusion of similarity measures, when title and abstract texts are not sufficient enough for the computation. 4.2 Recursive Expansion As discussed above, TL-Rank tries to compensate for any loss in the original paper by importing terms from its citing and cited papers. However, the utility of it will be limited considerably, if the imports face the same problem of incomplete text. This problem can be dealt with effectively, if we apply the term expansion in a recursive fashion. In other words, the term set expansion is repeatedly performed over several hops in a citation relationship network. The idea is similar, in spirit, to the expansions of Bibliographic coupling, Co-citation, and Amsler to their respective recursive versions of rvs-SimRank, SimRank, and P-Rank. If k denotes the number of recursive expansion steps, then TLRank above corresponds to k = 1. Results being presented here is the case of k being set to 2. We have found that going further does not necessarily lead to a better result because of noisy information introduced by the expansion process. Different weights assigned to the term set from cited papers, a query paper, and citing papers affect resultant similarity scores. Varying the weight combination, we have found that the ratio of 2 : 1 : 1 yields best result. Figure 2 compares the results of “Incomplete”, “Complete”, and TL-Rank cases. Results from the recursive expansion scheme are referred to as Recursive TL-Rank in the graph, while the case of k = 1 is indicated simply as TLRank. The accuracy of Recursive TL-Rank increases substantially up to 3.3 times compared to that of the incomplete abstract case. Moreover, our Recursive TL-Rank shows the results comparable with the complete abstract case, demon-

TL-Rank: A Blend of Text and Link Information for Measuring Similarity in Scientific Literature Databases