Is there any research work that shows that we should explicitly mark the word boundaries for 1D CNNs?

I'm doing character embedding for NLP tasks using one-dimensional convolutional neural networks (see Chiu and Nichols (2016) for the motivation). I haven't found any empirical evidence of whether or not marking the word boundaries makes a difference. As an example, a 1-D CNN with kernel size 2 would take "the" as input and use {"th", "he"} in its filters. But if I explicitly marked the boundaries it would give me {"t", "th", "he", "h"}.

Is there a go-to paper or project that definitively answers this question?