• Strategy for the summarization of scientific articles that concentrates on the rhetorical status of statements in an article: Material for summaries is selected in suchway that summaries can highlight the new contribution of the source article and situate it with respect to earlier work.

  • An algorithm that, on the basis of the annotated training material, selects content from unseen articles and classifies it into a fixed set of seven rhetorical categories. The output of this extraction and classification system can be viewed as a single-document summary.

Data Format

<PAPER>
<TITLE> Similarity-Based Estimation of Word Cooccurrence Probabilities </TITLE>
<AUTHORLIST>
  <AUTHOR>Ido Dagan</AUTHOR>
  <AUTHOR>Fernando Pereira</AUTHOR>
  <AUTHOR>Lillian Lee</AUTHOR>
</AUTHORLIST>
<ABSTRACT>
  <A-S ID='A-0' AZ='BKG'> In many applications of natural language processing it is necessary to determine the likelihood of a given word combination . </A-S>
</ABSTRACT>
<BODY>
  <DIV DEPTH='1'>
    <HEADER ID='H-0'> Introduction </HEADER>
    <P>
      <S ID='S-0' AZ='BKG'> Data sparseness is an inherent problem in statistical methods for natural language processing . </S>
    </P>
  </DIV>
</BODY>
<REFERENCELIST>
<REFERENCE>
<SURNAME>Sugawara</SURNAME>, K., M. <SURNAME>Nishimura</SURNAME>, K. <SURNAME>Toshioka</SURNAME>, M. <SURNAME>Okochi</SURNAME>, and T. <SURNAME>Kaneko</SURNAME>.
<DATE>1985</DATE>.
Isolated word recognition using hidden Markov models.
In Proceedings of ICASSP, pages 1-4, Tampa, Florida. IEEE.
</REFERENCE>
</REFERENCELIST>
</PAPER>
  • annotated data is present in xml format.
  • iterate through each tag and extract relevant feature information.

FeatureGenerator.py

  from FeatureGenerator import FeatureGenerator
  
  # locate corpus data 
  xmlcorpora = "../data/corpora/AZ_distribution/"

  # generate gold-standard features
  featureGen = FeatureGenerator(xmlcorpora)

  # sample feature vectors
  # featureGen.features[<xmlfile_name>][<line_ID>]
  
  # example
  print featureGen.features["9405001.az-scixml"]["A-0"]
  • generates features from parsing XML paper format
  • each sentence has 14 features along with a tag, stored in self.features in FeatureGenerator object
  • Implemented features capture
    • absolute location
    • Tf-Idf scoring
    • history
    • citations & references
    • title & length
    • explicit structure

NaiveBayes.py

  from FeatureGenerator import FeatureGenerator
  from NaiveBayes import NaiveBayes
  
  # locate corpus data
  xmlcorpora = "../data/corpora/AZ_distribution/"

  # generate gold-standard feature vectors
  featureGen = FeatureGenerator(xmlcorpora)

  # initialise and train NB model
  classifer = NaiveBayes(features=featureGen.features, distribution="Bernoulli", train_split=0.8)
  classifer.train()

  # Analysis
  confusionMatrix, precisionRecallF1, accuracy = classifer.test()
  classifer.plotConfusionMatrix(confusionMatrix, range(8))
  print("=== Accuracy: %f ===" %(accuracy))

  # generate summary
  classifer.getSummary('9405001.az-scixml')
  • gets the feature vectors of each sentence from FeatureGenerator object
  • trains with different classifer as choosen distribution="Bernoulli", available options include
    • Compliment
    • Bernoulli
    • Multinomial
    • Gaussian
    • Deep
  • Implemented classifiers
    • Gaussian Naive Bayes (acc: 25%)
    • Multinomial Naive Bayes (acc: 72%)
    • Bernoulli Naive Bayes (acc: 80%)
    • Compliment Naive Bayes (acc: 73%)
    • Deep CNN Network (acc: 71%)

clean.py

  python clean.py inputFile | awk '!uniq[substr($0, 0, 10000)]++' > outFile
  • Basic cleaning of the inputFile and writing to outFile, each line in outFile being a sentence from the inputFile.
  • May have errors, ignore them or manually correct if you want.

script.py

  python script.py inputFile outFile annotationFile
  • Send the inputFile (output of clean.py).
  • Each line from the inputFile pops up and you will be asked to tag it.
  • The details of the key for each tag will be output under each sentence.
  • Enter the corresponding tag and that line will be appended to outFile and your tag to the annotationFile.
  • Used RECHECK(R) tag incase you are not sure about it. Do it manually later, etc.
  • Used IGNORE(I) tag incase that sentence is trash and is not cleaned well from clean.py file.
  • The line and the tag wont be written incase of IGNORE tag.

Plots

  • set of true positive/negative & false positive/negative classification plots using confusion matrix
  • classifier prediction distribution plots using histograms ### Confusion Matrix
  • GaussianNB Confusion Matrix
  • ComplimentNB Confusion Matrix
  • MultinomialNB Confusion Matrix
  • BernoulliNB Confusion Matrix ### Histograms
  • GaussianNB Histogram
  • ComplimentNB Histogram
  • MultinomialNB Histogram
  • BernoulliNB Histogram

Dependencies

  # matplotlib
  pip install --user matplotlib
  
  # numpy
  pip install --user numpy
  
  # nltk
  pip install --user nltk
  
  # scikit-learn
  pip install --user scikit-learn
  
  # tensorflow
  pip install --user tensorflow
  
  # gensim
  pip install --user gensim
  
  # keras
  pip install --user keras