You basically have two choices: random and purposive (purposeful) sampling. With random sampling, the usual approach is to select some substantial but manageable proportion of the total (e.g., 25%) at random, so that you can generalize to the full set of documents.

With purposive sampling, you select documents according to some set of pre-specified inclusion criteria. Thus, rather than examining everything, you determine which documents are most likely to meet your purpose.

This paper will help you a lot. It's a study on media values based on the kind of content analysis you want to do. Very detailed methodology. 

 Dear Professor Morgan,

Thank you very much sir for your help.

Cameen

Dear Professor Jorge,

Thank you very much sir for the link. 

Cameen

In the literature of statistics, some authors have reported methods in which the finite population (in this question its units are documents) is divided in two strata: (1) Stratum of secure inclusion in the sample, in which all (the main) units are included in the sample of this stratum (a census of main units), and (2) Stratum of random selection, in which a simple random sample without replacement of units are drawn of this second stratum. Using in this case the usual unbiased estimator for stratified sampling there are conditions to assure more precision than with simple random sampling without replacement of all the population and the sample mean as estimator.

References:

Glasser, G. J. (1962). On the complete coverage of large units in a statistical study. Review of the International Statistical Institute 30, 28-32.

Ruiz Espejo, M. (1985). Equiprecisional allocation and optimum stratification. Statistics 16, 559-562.

 Dear Mariano Ruiz Espejo

Thank you very much sir for providing another perspective for me.  I'll go check out the two sources too.

Cameen

This is is a good introduction to Content Analysis in general (including sampling) from Bucy Herbert, and Holbert, Lance (eds) (2011): The Sourcebook for Political Communication Research. Routledge.  

 Dear Professor Jorge,

Thank you very much sir for the file. It is of very much help.

Cameen

When some auxiliary variable can be observed in all document population. It could be e.g book values of documents. In this situation the sampling with probabilities proportional to book values are proposed in textbooks about statistics in auditing. You can find many sampling designs in book by Y. Tille: Sampling Algorithms. Spronger 2006. 

I have some brief but scattered comments.  First of all, you should consult the accounting literature, specifically work on auditing.  Secondly, a skillful use of stratified probability sampling provides the benefits of purposive sampling without the drawbacks (lack of rigorous statistical propertires).  The design described by Mariano Ruiz Espejo is a particular case of this.  Lastly, Tille (2006), mentioned by Janusz Wywial, nicely illustrates the wide variety of probability sampling designs available.

Thank you so much for the shared information. My direct question would be - how do one go through the sampling method of the relevant categories of documents and the predicting the sample size?

Several propositions about auditing documents can be found in the book by Wywiał J. L. (2016). Cotributions to Testing Statistical Hypotheses in Auditing. Wydawnictwo Naukowe PWN Warszawa.

http://ksiegarnia.pwn.pl/Contributions-to-Testing-Statistical-Hypotheses-in-Auditing,404979797,p.html