How few is too few? Determining the minimum acceptable number of LSA dimensions to visualise text cohesion with Lex

Building comprehensive language models using latent semantic analysis (LSA) requires substantial processing power. At the ideal parameters suggested in the literature (for an overview, see Bradford, 2008) it can take up to several hours, or even days, to complete. For linguistic researchers, this extensive processing time is inconvenient but tolerated— but when LSA is deployed in commercial software targeted at non-specialists, these processing times become untenable. One way to reduce processing time is to reduce the number of dimensions used to build the model. While the existing research has found that the model’s reliability starts to degrade as dimensions are reduced, the point at which reliability becomes unacceptably poor varies greatly depending on the application. Therefore, in this paper, we set out to determine the lowest number of LSA dimensions that can still produce an acceptably reliable language model for our particular application: Lex, a visual cohesion analysis tool. We found that, across all three texts that we analysed, the cohesion-relevant visual motifs created by Lex start to become apparent and consistent at 50 retained dimensions.