Comprehensive computational lexicons areessential to practical natural languageprocessing (NLP). To compile such computationallexicons by automatically acquiring lexicalinformation, however, we previously requiresufficiently large corpora. This study aims atpredicting the ideal size of suchautomatic-lexical-acquisition oriented corpora,focusing on six specific factors: (1) specificversus general purpose prediction, (2)variation among corpora, (3) base forms versus inflected forms, (4) open class items,(5) homographs, and (6) unknown words.Another important and related issue withregard to predictability has something to dowith data sparseness. Research using theTOTAL Corpus reveals serious datasparseness in this corpus. This, again, pointstowards the importance and necessity ofreducing data sparseness to a satisfactorylevel for the automatic lexical acquisition andreliable corpus predictions. The functions ofpredicting the number of tokens and lemmas in acorpus are based on the piecewisecurve-fitting algorithm. Unfortunately, thepredicted size of a corpus for automaticlexical acquisition is too astronomicalto compile it by using presently existingcompiling strategies. Therefore, we suggest apractical and efficient alternative method. Weare confident that this study will shed newlight on issues such as corpus predictability,compiling strategies and linguisticcomprehensiveness.