About Word Association

[Andrealee1993]

Version number of KH Coder

3.02 macOS

Your Qeuestion

Your Operating System
Mac 10.15.7,

Your Question
A clear and concise description of your question.

Your Qeuestion:

Dear Professor Koichi,

I encountered another problem this afternoon. I have one excel file with three variables, respectively, DT2016, DT2020, and DT2024. When I was checking the word association via "List of variables & headings", I used this code "<>Time-->2024 country" to generate a network regarding the word "country" in 2024. I assume that all the words in this network are in DTC2024. Is my understanding right? However, when I was checking the collocates like in this screenshot, I found that concordances of this word are spreading in all the variables but not just only 2024.

Looking forward to your reply.

What language of text are you trying to analyze?
ex. English

[ko-ichi-h]

Well, I think that happens.

I assume that all the words in this network are in DTC2024. Is my understanding right?

All the words on the network came from [Word Association] screen.

And the words listed on the [Word Association] screen are those that appeared more frequently in 2024 than in the whole. These are not words that only appeared in 2024, but words that appeared more frequently in 2024 than in other parts. Therefore, it is possible that they may also appear in 2016 or 2020.

[Andrealee1993]

Oh I see. Thanks! "but words that appeared more frequently in 2024 than in other parts. " this means that these words are clustered by frequency, right?

[ko-ichi-h]

It's not too far off to think of it as Frequency, but strictly speaking, words are selected based on the Jaccard coefficient.

For example, if you could find the word "stop" on the [Word Association] screen and attach a screenshot of it, we'll be able to explain in more detail. It doesn't have to be "stop," any other word would be fine.

Section A.5.7 of the manual provides a explanation, but it doesn't go into detail about the Jaccard coefficient. When calculating the Jaccard coefficient, the number of documents or units containing the word is more important than the word's frequency itself.

[ko-ichi-h]

In the case of "Matthew", Jaccard coefficient is calculated like this:

a) Number of H5 units in part 01-07 that contain Matthew: 90
b) Number of H5 units in part 01-07 that do not contain Matthew: 338 - 90 = 248
c) Number of H5 units not in part 01-07 that contain Matthew: 285 - 90 = 195

Jaccard coefficient: a / (a + b + c) = 90 / (90 + 248 + 195) = 0.168856...

BTW, "Number of H5 units" means "number of cells in the excel file". H5 units means cells of Excel files.

It would be more accurate to say that the calculation is based on the probability of a word occuring within a given unit, rather than on word frequency. Because of this calculation, if you select a smaller unit such as "sentence" as the "Unit", you will probably get results similar to those of Frequency.

[Andrealee1993]

Thanks for your very patient guide! I will try to search for more information about it.

[Andrealee1993]

Dear prof.,
I happened to find the real meaning of your suggestion here:

It would be more accurate to say that the calculation is based on the probability of a word occuring within a given unit, rather than on word frequency.;

I am wondering whether it would be more accurate if it is caculated based on total tokens rather than sentences or units.

[ko-ichi-h]

While that's one option, I prefer the "probability of a word appearing in a sentence." There are two reasons for this.

[1] With this calculation, even if a word appears more than once in the same sentence, it's still considered to have appeared in only one sentence. It values ​​the word's appearance in many sentences more than its multiple occurrences in the same sentence. While repeated use in one sentence may be a some kind of coincidence, its appearance across many sentences is more likely to reflect some characteristic.

[2] For example, you can imagine that "out of 100 sentences, that word will appear in three of them." It is easier to imagine and interpret the meaning of the numbers.

(However, there may be a cultural gap here. I'm Japanese, and when expressing the length of Japanese text, we don't use word counts; we count characters. Since English seems to be more commonly measured in words, numbers calculated based on word counts may be easy enough for English-speaking people to imagine.)

For now, if you want to calculate based on token counts, you'll need to do it yourself using the word frequency list and the number of tokens displayed on the main window.

[Andrealee1993]

Thanks so much, Prof. Higuchi!
Using the Jaccard index formula you provided, I calculated the Jaccard co-efficient of my three variable values, respectively, 2016, 2020, and 2024. I got these figures shown in the screenshot.
My question is: From this table, the Jaccard co-efficient reveals an increasing trend in the association strength of the lemma we across the three campaign cycles from 0.1462 in 2016, 0.1562 in 2020 to 0.1757 in 2024. This indicates that we is more characteristic of variable value 2024, right? But KH Coder shows that we is characteristic of 2016, is that contradicted? Or my understanding of the figures are totally wrong?
I am desperately looking forward to your reply. Thanks again for your patience.

[ko-ichi-h]

So it means that there is no need to calculate the Jaccard index directly and the make a comparison, right? The first step is to look at the conditional probability then Jaccard right?

Yes, KH Coder first omits all words with low conditional probabilities.
Then it sorts the remaining words by Jacard coefficients.

[ko-ichi-h]

In addition to the tutorial, please read the manual.

If you don't understand something after reading the manual, please ask.

[Andrealee1993]

Thanks so much! Got that!