Dealing with multi-level categorical variables in regression

#1
Hi everyone,
I have been working in R and using lmer() to run mixed-effect linear and logistic regression models to study language processing.

I have recently learned a lot about how to code and include multi-level categorical variables in regression models (e.g., dummy coding, effects coding, Helmert, polynomial). One thing that keeps on coming up is that only orthogonal coding schemes like Helmert and polynomials have no collinearity among the coding variables included in the models; however, in my research, we usually care about differences among group means, which is what dummy coding tends to test. But, when I look at the correlation matrix after running these models, I see fairly high correlations between levels of these coding variables.

For example, say I have a categorical factor called Part of Speech with three levels (nouns, verbs, adjectives). I want to include these as dummy coded variables with a baseline of "nouns". So I'd have two coding variables (NounsVsVerbs, NounsVsAdjs). I find that these coding variables are often correlated (0.5 or greater). How can I determine if I need to worry about this level of collinearity between these predictors, especially when there are other predictors in the models? How can I know if I can I trust the size and direction of these effects when I report, for example, that nouns are faster than verbs when I get a significant beta for the NounsVsVerbs coding variable predictor?

Does anyone know where to find a good reference that addresses this issue?

Thanks in advance for your help!

~Maureen