STRATEGIES FOR FINDING RELEVANT DOCUMENTS AMONG NEIGHBOR LISTS

获取数据库访问权限的一种方法是让用户选择 一些关键术语或写一个简短的描述他的主题 兴趣直接给出或经过处理的结果关键词 然后，使用写入的请求中的 相关文件的数据库。众所周知， 这种搜索的特异性不如 从搜索者兴趣的更明确的描述中获得， 在许多情况下，他只能在找到一个 感兴趣的文章这就是所谓的相关反馈的基础 程序.相关反馈利用的相关性判断作出的， 已经检索到的材料进行重点搜索。在这项研究中，我们 关注一种特殊的相关反馈。 我们处理数据库，其中每个文档的顶部列表 已经计算了与给定文档相关的文档。 我们的目的是发现最有效地利用这些预先计算的 邻居列表以便于搜索，该搜索可以开始于基于 在前一段中描述的几个关键术语上。当 找到第一个相关文档时，它带有一个预先计算的 邻居问题是是否要回头看下一个相关的 在初始列表中查找文档或在预先计算的 与相关文档相关联的邻居。第二次相关 文件已经确定，然后又有一个问题，如何 最佳地使用它的邻居列表等等。我们已经尝试了一些 不同的策略，并发现有明确的改善， 并量化了所获得的性能 CISI、CRAN、CACM和MED测试文档集。 一篇论文 有关这些结果的描述正在印刷中。

One method for gaining access to a database is for a searcher to choose some key terms or write a short description of the topic of his interest. The resultant key terms which are given directly or processed out of the written request are then used to make a vector for searching the database for related documents. It is well known that the specificity of such a search is not as good as that which would be obtained from a more definitive description of the searchers interest, a description which he in many cases can only give after he has found an article of interest. This is the basis for so-called relevance feedback procedures. Relevance feedback makes use of relevance judgments made on already retrieved material to focus a search. In this study we are concerned with a special kind of relevance feedback. We deal with databases in which for each document the list of the top documents in relation to the given document have already been computed. Our purpose is to discover the most efficient use of these precomputed neighbor lists to facilitate a search which may begin by a search based on several key terms as described in the previous paragraph. When the first relevant document is found it comes with a precomputed list of neighbors. The question is whether to look back for the next relevant document on the initial list or to look on the list of precomputed neighbors associated with a relevant document. After the second relevant document has been identified then one again has the question of how to use its neighbor list optimally and so on. We have tried a number of different strategies and have found that there is definite improvement in using the neighbor lists and have quantified the performance obtained on the CISI, CRAN, CACM, and MED test sets of documents. A paper describing these results is in press.