Performance on single-sentence documents

[ahalterman]

On pg. 1959 of the paper, it says

Particularly, the frequency-based approach is as follows. We form a histogram of each location
detected by NER tools. The location having the highest count is selected as the focus location.

Does this mean that on a corpus of single-sentence documents (i.e. one where there's no need to find the "focus sentence"), Profile will just return the highest-frequency place name from each sentence as the focus location?

Relatedly, if each place name in a sentence is mentioned once, how does it decide which to return?

[Maryam-Imani]

On pg. 1959 of the paper, it says

Particularly, the frequency-based approach is as follows. We form a histogram of each location
detected by NER tools. The location having the highest count is selected as the focus location.

Does this mean that on a corpus of single-sentence documents (i.e. one where there's no need to find the "focus sentence"), Profile will just return the highest-frequency place name from each sentence as the focus location?

Relatedly, if each place name in a sentence is mentioned once, how does it decide which to return?

It means that for each document, we extract all sentences which have a location in them. Then, based on the trained model, we assign a label focus/non-focus to them. Then, if there is more than one focus sentence, we consider the location names in those sentences and select the location with the higher frequency. If there are two locations with the same frequency, we select one of them randomly.