New "combined score" for entity linking when place of pub is known

[thobson88]

Linker algorithm using place of publication & combined scores

The first two steps of the pipeline are unchanged:

• identify toponyms (Recogniser)
• rank candidates with deezymatch (Ranker)

The third step, disambiguation via the Linker, is new. We assume the place of publication wqid and latlon are known.

For each identified toponym:

• if the place of publication wqid is found in the list of candidates and is not the prediction:
  • get string match score for the place of publication (from the deezymatch scores computed by the Ranker)
  • if the string match score for the place of publication is greater than 0.01:
    • set the prediction to the place of publication
    • exit
• attach the place of publication information to the sentence
• for each candidate:
  • compute the REL+pub cross_cand_score
  • if the latlon coordinates for the candidate are not known/available:
    • set the combined_score equal to the cross_cand_score
  • if the wikidata popularity for the candidate is not known/available:
    • set the combined_score equal to the cross_cand_score
  • if both the candidates latlon coords and popularity are computable:
    • compute the popularity (i.e. the Wikidata "most popular" score)
    • compute the proximity
    • set the combined_score = cross_cand_score * max(popularity, proximity)
• rank candidates by the combined score and set the prediction to be the top one
• exit

See also: https://github.com/Living-with-machines/data-culture-newspapers/issues/17