It is generally accepted that information retrieval based on full texts of documents will result in higher recall and lower precision compared with retrieval using paragraphs, abstracts, or controlled vocabularies. The present study tested this assumption by examining the effectiveness of full text retrieval compared with other approaches in terms of recall and precision. It then focused on how to improve the precision of full text retrieval with a minimum decrease in recall.;A subset of Harvard Business Review records for the time period of Jaunary 1979 - August 1983 was selected for analysis, and nine search questions were selected to test the effectiveness of full text retrieval. Twenty-nine weighting algorithms developed for automatic extractive indexing were examined as means of improving the precision of full text retrieval.;There was a significant difference in retrieval effectiveness between full text and other methods, at the .0001 level for recall and at the .0032 level for precision. That is, full text retrieval achieved significantly higher recall and lower precision ratios than paragraph, abstract, or controlled vocabulary searching. There was also a significant difference at the .0290 level between full text retrieval with weighting algorithms and full text retrieval without weighting algorithms for precision measured at a fixed level of recall. All 29 algorithms improved the precision of full text retrieval over that without algorithms. Twenty-two of 29 algorithms achieved higher precision than that of paragraph searching, although these results were not statistically significant.;However, there was no significant difference between algorithms. The relative performance of algorithms seemed to depend on the search strategy employed and level of recall achieved. In general an algorithm which gave high weights to common words and low weights to rare words in the Boolean intersection performed well at low levels of recall. The performance of algorithms did not vary much at high levels of recall.
展开▼
