The code takes in an article and returns a summary of about 75 words. The code uses the following features

1. Similarity between a sentence and the first sentence of the text.
2. Importance to first line of every new paragraph
3. Similarity between sentence and first line of the corresponding paragraph.
4. PageRank algorithm.
5. Importance to sentence containing a quoted text.
6. A lexical database that accounts for synonyms.
7. Chronological sequence.
8. Frequently occuring unigrams and bigrams.

To compute the similarity between two sentences I have used tf-idf cosine similarity method. More info on:- https://janav.wordpress.com/2013/10/27/tf-idf-and-cosine-similarity/ PageRank refference:- http://www.aclweb.org/anthology/P04-3020 Lexical database from wordnet. Feature references http://anthology.aclweb.org/C/C08/C08-1124.pdf.

Summarization is of two types. Extractive and Abstractive. Extractive is what I have done and basically ranks individual sentences according to a specified criteria. Research in Abstractive is in its budding stages with no consolidatory instances of its implementation. It is hugely difficult and does not involve any conclusive solution. Please do read about it more. I checked into various products and pioneers like inShorts have their own manual team to curate and summarize articles while newstand does extractive.

I also implemented a Naive Bayess learning model to the above algorithm that trains the code to analyze results of an existing database and then compute the present summary.

5.2 Naïve Bayesian Classier (NBC) Naïve Bayesian Classier assumes features are independent. It learns prior probability and conditional probability of each feature, and predicts the class label by highest posterior probability. Given a feature vector (F1, F2, F3,…, Fn), the classifier need to decide the label c: argmax ( | , , ,... ) 1 2 3 n c c = P c F F F F By applying Bayesian rule, we have ( , , ,..., ) ( ) ( , , ,..., | ) ( | , , ,..., ) 1 2 3 1 2 3 1 2 3 n n n P F F F F P c P F F F F c P c F F F F = Since the denominator does not depend on c and the values of Fi are given, therefore the denominator is a constant and we are only interested in the numerator. As features are assumed independent,

n i i n n P c P F c P c F F F F P c P F F F F c 1 1 2 3 1 2 3 ( ) ( | ) ( | , , ,... ) ( ) ( , , ,... | ) where P(F | c) i is estimated with MLE from training data with Laplace Smoothing

Given: L is the set of labeled training examples U is the set of unlabeled training examples Loop: until the unlabeled data is exhausted Train the first classifier C1 (PSVM) on L Train the second classifier C2 (NBC) on L For each classifier Ci Ci labels examples from U Ci chooses p positive and n negative examples E from U. These examples have top classifying confidence. Ci removes examples E from U Ci adds examples E with the corresponding labels to L

https://www.youtube.com/watch?v=EGKeC2S44Rs

After implementing this I found the results to be pretty erratic and certainly worse than what was achieved before. (In technical terms, the probability of each sentence to not be included in the summary was either very high or very low). Owing to the results, I scraped the model completely. If you want I can redo it and show you the results.

I recommend all to test the code and run it for bugs and test it for different articles. A sample input is present in the input.txt file and the output in the output.txt file. In order to run it, please download an IDE(prefferably NetBeans 8.0.2) and watch the following video https://netbeans.org/kb/docs/ide/git_nb_ssh_screencast.html.

Thank you!.