Comparability of LSI and human judgment in text analysis tasks

The technique of latent semantic indexing (LSI) plays an important role in a growing number of text processing applications. In these applications, the most important aspect of LSI is that it can be used to emulate human judgment of textual content. In general, the economic savings that can be obtained thorough the use of LSI are directly dependent upon the fidelity with which LSI can be employed as a surrogate for human judgment. This paper presents an overview of 30 sets of studies in which the performance of LSI in text processing tasks can be compared directly to human performance on the same tasks. In half of the studies, performance of LSI was equal to or better than that of humans. The paper presents arguments that even this surprisingly high performance actually underestimates the potential performance of LSI. Two techniques are presented which can be used to obtain improved results in general LSI applications.