2

I've been engaged in a conversation on another site pertaining to frequency analysis, particularly in relation to the 1966 work Buntús Gaeilge, Colmán Ó Huallacháin; Ireland. Department of Education.

This statement was made in the discussion:

Frequency analyses cannot produce statistically trustworthy results without having a heck of a lot of language data. Even now, corpuses can be spoiled by selective choosing of source material, and they contain billions of words. No matter anyone's best intentions, in 1966 it would have been impossible to build a large-scale, statistically relevant corpus, so they would have had to be very selective in what they thought was worth including, and that selective step destroys the objectivity of the study.

So this has prompted me to wonder about the answer to the following question(s)?

  • What is considered a large enough sample size when being used to guide the creation of pedagogical resources such as foreign language courses and why?

And a bonus question...

  • Is there any actual merit to the statement bellow?

No matter anyone's best intentions, in 1966 it would have been impossible to build a large-scale, statistically relevant corpus, so they would have had to be very selective in what they thought was worth including, and that selective step destroys the objectivity of the study.

  • Assume that you teach English to foreigners and the subject is conjunctions. Then you can take a look at researches based on frequently seen conjunctions in English (pg: 14) and you are able to know what to teach in lesson (Yes, I know it is a bit much.). Seeing that we can say enough sample size is enough when frequency list will not have major changes (In document, there's not a big chance of because to decrease or for to increase.). I was not sure it was the answer of your question so I posted it as comment. – Eray Erdin Aug 21 '14 at 19:15
2

What is considered a large enough sample size when being used to guide the creation of pedagogical resources such as foreign language courses and why?

How large your corpus should be depends on what exactly you want to use it for, and what alternatives are available. If you want to know the 2,000 most frequent words of a language, a 1 M. word corpus will be sufficient to conclude that these words should be taught earlier than the words with frequency ranks 10,000 to 12,000.

Of course, if a much larger corpus is available (say, 10 M. words), you will want to use this one. But in 1967, the Brown Corpus of American English, the first modern computer-based corpus, just became available, with a size of 1 M. words. Nowadays, much larger corpora are available for many languages and dialects. But for a smaller language, if a 1 M. word corpus is available, it will allow you to answer many questions (regarding frequent lexis etc.) that you couldn't hope to answer without a corpus of this size.

Is there any actual merit to the statement below:

No matter anyone's best intentions, in 1966 it would have been impossible to build a large-scale, statistically relevant corpus, so they would have had to be very selective in what they thought was worth including, and that selective step destroys the objectivity of the study.

As the Brown Corpus shows, this was in fact possible in the 1960s. Also, being selective about what you include is a good approach in empirical research. Rather than including just any language samples that were available, the Brown Corpus samples published material from a range of genres. The corpus make-up mirrors closely what was published in 1961 in the US, so that it can claim to be representative of 1961 written American English. This is basically the same approach as carefully choosing a group of 1,000 Americans (based on socio-economic and geographical criteria) to determine voting patterns, rather than asking 10,000 completely randomly selected Americans.

The "completely random" approach is also used nowadays, for example to compare web-based megacorpora. If you want to investigate low-frequency phenomena this is great, but a huge, automatically compiled corpus comes with its own set of problems - and depending on the research question, sometimes a web-based mega-corpus might be appropriate, and sometimes a carefully compiled smaller corpus.

| improve this answer | |
  • Thanks for the reply. Could you explain why 1 million is sufficiently large? – Robert Kaucher Aug 21 '14 at 19:47
  • Because with 1 M. words, the likelihood is sufficiently small that you include a word among the 2,000 most frequent words that is, in reality, very infrequent (say, not even among the 10,000 most frequent words). But if you have an (equally) well-balanced larger corpus, obviously you'd prefer the larger one. – robert Aug 21 '14 at 20:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.