Entity Localization Bug: Sentence. Sentences with colored citations have fragmented bounding boxes, in many papers.

[andrewhead]

Description: In papers that contain colored citations, sentence bounding boxes can be incorrect. In particular, they're fragmented, either missing part of the sentence that contains the citation, or only including long horizontal bars through the citations.

This problem has been observed for the following papers from among those reviewed from the list of 24 papers that appers in this issue #188:

• 1701.07481v3
• 1702.01287v1
• 1701.02810v2
• 1906.00414v2
• 1906.01502v1
• 1908.00300v1
• 1905.05475v2
• 1705.06566v2
• 1802.07740v2
• 1806.02371v1
• 1811.12359v4

I have also seen this problem for sentences that contain a URL that wraps more than one line:

• 1701.02810v2 (First sentence of Background)

Here are a couple of examples. To reproduce, open up the paper by going to https://scholarphi.semanticscholar.org/?file=https://arxiv.org/pdf/[PAPER_ID].pdf. Then open up the web console and enter the following style:

.sentence-annotation {
  background-color: rgba(0, 0, 255, 0.2);
}

Paper 1701.07481v3

How to fix (optional): I believe this issue is because citations are assigned a color by the hyperref package when it's enabled for papers. This color is preserved even when the surrounding sentences is assigned a different color. As a result, citations are not detected as belonging to the sentence because they do not have the expected color of that sentence.