Entity Localization Bug: Sentence. Runaway sentences in many papers

[andrewhead]

Description: In some papers, it appears that the end of a sentence is never detected, such that that sentence bounding boxes start somewhere in the middle of the paper, and then continue throughout the entire rest of the paper until the end.

I have inspected 24 papers from PR #188, and observed this problem in the following papers:

• 1702.01287v1: Starting at "Finally, we use..." in Sec 3)
• 1701.02810v2: Starting at "Finally we briefly..." in Sec 4.4)
• 1905.05475v2: Starting at "The idea is to reuse..." in Sec 3.3)
• 1706.08482v1: Starting at ("d in every rame..." in Sec 3). Note that htis paper has multiple runaway sentences.
• 1704.05838v1: Starting at ("When training with the..." in Sec. 3.5)
• 1705.06566v2: Starting at "The GAN architecture was..." in Sec 1.2.
• 1802.07740v2: Starting at "Finally, as a ToM needs to..." in Sec 2.1.
• 1901.10159v1: Starting at "Therefore a single update..." in Appendix E
• 1905.10887v2: Starting at A.2.

_How to fix: One fix is to fix the sentence segmenter; I would guess that the command for undoing the color at the end of a sentence is placed in a part of the TeX that is invalid, and hence doesn't successfully reverse the color. Another simpler fix is to do outlier detection, not uploading sentences which contain more than some threshold number of bounding boxes (e.g., more than 100).