How to optimize uptake of uploaded corpora?

Question

When we upload corpora to MS Hub, segments are extracted, aligned, and filtered.  
We call "uptake" the proportion of segments that are used in the end, over the initial number of segments uploaded.
In our experience, only about 70-75 % of segments are in fact used by the system:  an uptake of 70-75%. So it seems that 25-30% of our valuable domain-specific corpora are not being leveraged at all.
Example:  A single, bilingual TMX that shows 142,791 segments in WorldServer, appears with 134,557 "extracted" in Hub, with 104,100 "aligned", and 94,730 "used".  That's an uptake of only 66% of the original TM; i.e.,
48,000+ segments were ignored!
When we extract and filter segments from the same TM outside of Hub, then review the alignments manually and upload them as .align files, we apparently get no significant improvement:  18% of the segments that we _know_ are aligned correctly are still
eliminated by Hub.
Questions:
How can we improve the uptake of corpus segments -- i.e., increase the proportion that the system actually uses?
What criteria does the system use to eliminate segments?
How can we override the system's alignment step?
Thanks!

Answer 1

Thanks for your reply!
Re #3:  I guess I wasn't clear in my question.  I trained the same system once with a TM, again with extracted and aligned segments from the same TM.  So, I _did_ supply perfectly aligned .align files and the system STILL eliminated 18% of
segments (the difference between extracted and aligned) from the .align files . This is why I am puzzled.
Update:  I extracted aligned segments from TMX files with a separate script; I'm not reporting numbers for the extracted sentences that Hub provides.
Re #2:  Thanks for the hints. I do see one- and two-word segments eliminated.  But most of the segments that were eliminated do not match the criteria that you mentioned. In these cases, there are no tags, the alignment is correct, the segment
is not a duplicate (I only checked for segments that were no longer present at all), and the length is fine -- but they are not in the set of aligned sentences.
For example:
The image you uploaded no longer exists.
L’immagine caricata non esiste più.
Any further information or examples that you can provide will be appreciated!