While we may initially think of keywords as being the keywords we are often asked to provide when uploading an essay or a video; the term has a different meaning within corpus linguistics. When we talk about keywords within our field, we refer to words that are quantitatively characteristic of a text when compared to others, meaning words that appear more frequently in one text or type of text than other texts. So, we take a comparative perspective on our target texts and see them in relation to a reference corpus. When selecting our reference corpus, we must consider the types of texts we are comparing, as well as when they were written and other possible differences between the two datasets. So, how do we measure the importance of a keyword?

Keyness is the primary metric for keywords, but it can mean different things. We sometimes see research where the T-test, chi-squared, or log-likelihood are used to represent keyness by testing the significance of the difference in the words' relative frequencies. These three tests are used to see whether we can reject the null hypothesis or not by providing a probability value (often referred to as a P-value) used to indicate statistical significance. In the context of keyness, these tests are intended to indicate whether or not the high frequency of a keyword is likely to be caused by chance or not.

A different set of approaches focuses on the relative frequencies' effect size. What this means is that we look at how much the relative frequencies of our keyword differ in our target corpus and our reference corpus and then assign keyness according to the difference we see. Today, it is not uncommon for corpus linguists to use a combination of these methods, for instance, measuring keyness through the P-value of the effect sizes being compared. While statistics are a great way of producing a list of keywords for our target corpus, we must, at some point, move on to the actual content of the list in order to understand what is going on.

The cultural and social importance of a keyword is often a criterion for us selecting it as being important for further exploration, and indeed, the term keyword was used long before statistical keywords were discussed in linguistics (Williams 1976). The way that keywords were used by Williams could also be seen as an example of early topic modelling, which we will explore later in  3.3. Of course, these two aspects tie into the specific research questions for our study, but roughly speaking, we tend to focus on the words that appear to carry some important meaning. Here, our understanding of language patterns will play a crucial role in our ability to actually pick out the keywords that are key to our interests. If we want to explore grammatical features, we might want to focus on function words related to those grammatical patterns, while an interest in the content might lead us to focus on nouns, or place names, or whatever else might be relevant to us.

We can think of these words we have selected as keywords. This brings us back to the beginning of keyword analysis, as before corpus linguistics, researchers qualitatively hand-picked keywords when creating representations of texts. What researchers were interested in then was the importance of a word when moulding a discourse, for instance, the use of "democracy" within texts partaking in the discourse on education. The idea was to find words that indicated the social, cultural, or political theme.

When we discuss corpus linguistic approaches to keyword analysis today, we can choose to either let the retrieval be automatic through the approach selected above and then pick our words from the finished list, or we can perform a reading of our target material and then focus on specific words from the beginning. In the literature on the topic, this is sometimes discussed as an explorative or a targeted approach. As with everything else, which we elect to apply will depend on what we want to accomplish. If we are performing research on specific words, selecting those already from the start makes sense, while an explorative effort into unknown texts would benefit more from the broader approach.

---

References:

• Williams, Raymond. 1976. Keywords: A Vocabulary of Culture and Society. Croom Helm.