Log-likelihood and odds ratio: Keyness statistics for different purposes of keyword analysis

Abstract

Keyword analysis is used in a range of sub-disciplines of applied linguistics from genre analyses to critically-oriented studies for different purposes ranging from producing a general characterization of a genre to identifying text-specific ideological issues. This study compares the use of log-likelihood (LL), a probability statistic, and odds ratio (OR), an effect size statistic, for keyword identification and argues that the two methods produce different keywords applicable to research focusing on different purposes. Through two case studies, keyword analyses of advance fee scams against the British National Corpus and research articles in applied linguistics against research articles from other academic disciplines, we show that both the LL and OR keywords concern the aboutness of the corpus, but differ in their specificity and pervasiveness through the corpus. LL highlights words which are relatively common in general use serving genre purposes, whereas OR highlights more specialized words serving critically-oriented purposes. Methodological and practical contributions to keyword analysis are discussed.

Developing and Validating an Academic Vocabulary List in Russian: A Computational Approach

“…The resulting word list does indeed contain many relevant academic words, but it also includes lemmas such as slave , protein , and verse , which are highly discipline specific. This finding provides additional evidence for the claim that likelihood and odds ratio statistics are competing keyness statistics and their relative performance is highly case specific (Pojanapunya & Todd 2018).…”

Mining Facebook Data of People with Rare Diseases: A Content-Based and Temporal Analysis

“…To make this comparison, LL was chosen to compare among different texts because common words are highlighted. The reason of this choice was because LL puts more emphasis in common words for a given topic than other measures [ 30 ], which are most likely to be used by non-expert users. LL was computed by constructing the contingency table of Table 1 .…”

The Language of Destructive Cults

“…In these analyses, we used the probability statistic called log-likelihood (LL) to measure the keyness of linguistic features which helps us to see the difference between the target corpus and the benchmark corpus. LL is an appropriate keyness statistic application since it has been widely used in previous research and it allows us to characterize the specific register under investigation (Dunning, 1993;Pojanapunya & Watson Todd, 2018). To reveal the LL values, this significance test statistic compares the frequencies of linguistic features in the target corpus to the frequencies of linguistic features in the benchmark corpus while taking the overall size of corpus in each data set into account (Pojanapunya & Watson Todd, 2018).…”