citation-context-summarization

Citation Context based scientific summarization algorithm based on C-LexRank (http://www-personal.umich.edu/~vahed/c-lexrank.html)

Modified summarization algorithm:

Available at /summaryModAlgo

• The dataset is present in summaryModAlgo/contextFiles as {paperID}.txt
• The factoids for the dataset required for evaluation is present in summaryModAlgo/facts as {paperID}.ann
• The log for the runs is present in the file log.txt
• The scripts are present in summaryModAlgo/scripts

List of features

For 2 sentences in the citation context S1, S2

• 1 - Unigrams
  Cosine similarity using TD-IDF vectorization of unigrams after removing stopwords (stopwords are removed using NLTK corpuses)

• 2 - Bigrams
  Cosine similarity using TD-IDF vectorization of bigrams without removing stopwords

• 3 - Number of citations
  sim( S1, S2 ) = min (C(S1), C(S2)) / max (C(S1), C(S2))
  Where, C(Si) = Number of out-citations in the sentence Si

• 4 - POS Tag unigrams
  Cosine similarity using TD-IDF of unigrams of POS tags after removing stopwords.
  

• 5 - POS Tag Bigrams
  Cosine similarity using TD-IDF of bigrams of POS tags withought removing stopwords.
  

• 6 - Bibilographic Coupling
  sim (S1, S2) = |intersection(outCites(S1), outCites(S2)|/ min ( |outCites(S1)|, |outCites(S2)| )
  Where,
  inCites(Si) = Out-vertices of Si in the citation graph
  intersection(A, B) = set of common elements in A & B
  and |A| = cardinality of list A
  

• 7 - Co-citation matrix
  sim (S1, S2) = |intersection(inCites(S1), inCites(S2)|/ min ( |inCites(S1)|, |inCites(S2)| )
  Where,
  inCites(Si) = In-vertices of Si in the citation graph
  

• 8 - Title similarity
  sim (S1, S2) = Unigram-Similarity(Title(paper1), Title(paper2))
  

• 9 - Author similarty
  sim (S1, S2) = |intersection(authors(S1), authors(S2)|/ min ( |authors(S1)|, |authors(S2)| )
  Where,
  authors(Si) = list of authors of Si
  

• 10 - Time based similarity sim (S1, S2) = max( 0, (1 - abs(year1 - year2)/5))
  Where,
  abs (x) = absolute value of x
  year1 = publication year of S1
  year2 = publication year of S2
  

Running the algorithm

• Execute the algorithm over the whole dataset and run the pyramid evaluation scheme

python executeAndEvaluate.py <featureIndex> <comments>

• Execute the algorithm over a paper

python executeModAlgo.py <paperID> <featureIndex>

• Output
• The output summaries are added to summaryModAlgo/modOutput as {paperID}-C-LR.txt
• The pyramid scores can be obtained from the log file at summaryModAlgo/log.txt and summaryModAlgo/pScore.ods