Accept vs. Reject ICLR Reviews LORIDP

iclr_positive_negative_loridp.py

reviews_df = pd.read_csv('https://github.com/JasonKessler/ICLR18ReviewVis/raw/master/iclr2018_reviews.csv.bz2')
reviews_df['parse'] = reviews_df['review'].apply(spacy.load('en', parser=False))

# Create Corpus based on accept/reject/workshop decision
full_corpus = st.CorpusFromParsedDocuments(
          reviews_df, category_col='decision', parsed_col='parse').build()

# A two-category corpus to use for plotting, with unigrams which only occur in bigrams removed.
# Terms used in <5 documents are removed as well.
corpus = st.CompactTerms(full_corpus.remove_categories(['Workshop']), 
                         st.OncePerDocFrequencyRanker, 5).compact()

# Use counts of unigrams and bigrams from the Workshop corpus as the Dirichlet prior
priors = (st.PriorFactory(full_corpus, term_ranker=st.OncePerDocFrequencyRanker)
          .use_categories(['Workshop'].align_to_target(corpus).get_priors()))
term_scorer = LogOddsRatioInformativeDirichletPrior(
          priors, reviews_df.parse.apply(len).mean(), 'word') # use the original approach to scaling prior
 
html = st.produce_frequency_explorer(corpus, 
  category='Accept', not_categories=['Reject'],
  term_ranker = st.OncePerDocFrequencyRanker,
  term_scorer = term_scorer,
  grey_threshold = 1.96,
  metadata = corpus.get_df()['metadata'])