ijmarshall / cochrane-nlp Goto Github PK

change to tokenize 'N=192' as 'N', '=', '192'

NLTK's word_tokenizer retains 'N=192' as a single word token; which means that our ML tools are retrieving 'N=192' as the population size.

Ideally, it would be tokenized as 'N', '=', '192', to be able to use the 'N=' as features, but also since the distant supervision requires the integer in order to match between the Cochrane and pubmed datasets.

However, the old hack way of doing it (re substitution pre-processing) will not work for any annotated data, since pretty much all instances will be tagged as e.g. 'N=<n>192</n>'

Currently quality.py uses the regex "Quote\:\s*[\'\"](.*?)[\'\"]") to get the first quotation from the Cochrane data.

This mostly works, but misses some data in the case where there are multiple quotes given to justify a domain

Quote: "… randomized into intervention (n = 1718) and control
(n = 1714) groups by an based on simple randomization. " "The
subjects were informed by letter to which group they were
randomized. The letter contained information concerning the
trial and the prescription for the intervention."

The code should identify and tag the text from both quotes in the full text of the study, rather than just the first where this happens.