A Conception-Based Approach to Automatic Subject Term Assignment for Scientific Journal Articles

0.69 0.68 0.68 full text 0.71 0.69 0.70 Introduction 0.73 0.72 0.72 title of cited works 0.76 0.73 0.74 keyword 0.82 0.80 0.81 From the results of the eight semantic sources and the characteristics of semantic sources, two comparisons can be noted. One is the comparison between the attributes ‘abstract’ and ‘keyword’. While both ‘abstract’ and ‘keyword’ are provided by the authors for representing a concise version of the full text, the difference in effectiveness of ‘keyword’ and ‘abstract’ showed a substantial difference. Another comparison can be noted between ‘introduction’ and ‘conclusion’. In general, with almost the same length of data, both were extracted from the full text of the article. However, the results of these two semantic sources again presented a considerable difference in effectiveness. While the effectiveness of ‘introduction’ shows better effectiveness than ‘full text’, ‘conclusion’ is the least effective among the eight semantic sources. In order to indicate how the effectiveness of semantic sources differs from the effectiveness of ‘full text’, comparisons were made using t-tests between ‘full text’ and the remaining semantic sources. The result of TC using ‘full text’ of documents was selected as the baseline because the majority of current TC research uses full text-based classification. In order to see a significant difference between the baseline and each semantic source, seven pairs of t-tests were applied. Table 3 indicates that while there is no significant differences with the baseline in terms of precision and F-measure, there is a significant difference between ‘keyword’ and the baseline in recall. In addition, ‘title of cited works’ presents a nearly significant difference with the baseline. This result indicates that all of the individual semantic sources performed as well as or better than the full text sources in effectiveness in assigning subject terms compared with the full text. Table 3. T-test between each semantic source and baseline semantic source precision recall F-measure T p t p t p conclusion -1.230 0.144 -2.228 0.056 -1.727 0.092 title -0.041 0.485 -0.408 0.355 -1.739 0.090 source title -0.529 0.317 -0.305 0.390 -0.461 0.338 abstract -0.418 0.352 -0.147 0.446 -0.340 0.3780.418 0.352 -0.147 0.446 -0.340 0.378 introduction -0.204 0.426 -0.942 0.208 -0.444 0.344 title of cited works -0.404 0.357 -2.113 0.063 -0.685 0.272 keyword -0.971 0.202 -2.828 0.033* -1.524 0.113