"Could you recommend me any paper on (over)interpretation of processing data" seems to infer that you are rather sceptical of this method. "Under-interpretation" does, however, seem to do it more justice. A key point here is the issue of 'online processing' - unlike written texts (that leave time to edit) language in its primary form can now be analysed to an extremely high degree. Based on those insights, neural networks employed by machine learning can very mimic these processes to an eerily life-like degree.

Thank you. I think that our perspectives are just quite different. I'm not skeptical of language processing methods, I think quite the opposite. My point is the differentiation between representation as implicite/explicite language knowledge represented in our memory and processing it, e.g. using this knowledge during production and comprehension. Let's take cognate facilitation effect. Many studies have shown that cognates are recognized faster during visual/spoken word recognition and this results is often taken as evidence of a shared lexical representation. Isn't is just about processing?

I would not disagree :-) Maybe ask to find "any critical evaluation of the cognate facilitation effect".