Other comments that I wasn't sure how to suggest fixes for:
When you define "ConceptNet vector ensemble", it implies that it doesn't include PPDB, but everywhere that you subsequently refer to it, it seems that it does include PPDB. So it would probably be good to clarify this somehow.
The text in the dataflow figure is really tiny.
Although you list the paper's contributions in the introduction, I felt that after reading the paper I still didn't have a good concise idea of "this is what this paper contributes", it was more like "they tweaked a bunch of things and stuff got better". That might just be the nature of the paper, but I wonder if it would be possible to put a bit near the end with a succinct explanation of how this is substantively different from previous work. I liked the phrase "wide learning" from previous versions of the paper.
Other comments that I wasn't sure how to suggest fixes for: