Skip to content

karthikbmk/is-karthikbmk

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

96 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Text Summarization using Deep Learning and Ridge Regression

summ

Problem

Develop Models to Summarize long documents.

Research questions

1.What are the relevant features for summarization?
2.How to tune the Hyper-parameters for a given model?
3.Which model works best for summarization?
4.How to stich together multiple sentences for summarization?

Related work

Here's how other people have tried to solve this problem:

  1. Query-Oriented Multi-Document Summarization via Unsupervised Deep Learning http://www.aaai.org/ocs/index.php/AAAI/AAAI12/paper/view/5058

  2. Ranking with Recursive Neural Networks and Its Application to Multi-document Summarization:
    http://www.aaai.org/ocs/index.php/AAAI/AAAI15/paper/download/9414/9520

  3. Reader-Aware Multi-Document Summarization via Sparse Coding:
    www.aaai.org/ocs/index.php/IJCAI/IJCAI15/paper/download/11398/10841

Data

DUC 2001 dataset will be used to build the initial model.Here is an example data record:

LA021090-0005 174161

February 10, 1990, Saturday, Home Edition

Metro; Part B; Page 6; Column 4; Op-Ed Desk

873 words

GUN NUTS HAVE A REAL POINT;

CONSTITUTION: THE CLIMATE MAY NOW BE RUNNING IN FAVOR OF MORE RESTRICTIONS, INCLUDING A BAN ON HANDGUNS. BUT THERE'S STILL THE SECOND AMENDMENT.

By MICHAEL KINSLEY, Michael Kinsley writes the TRB column for The New Republic.

Around the world, national theologies are crumbling: communism, apartheid and, here in America, the worship of guns -- to foreigners, the single craziest thing about us.

Do you sense an outbreak of sanity about gun control? I do. There was retired Chief Justice Warren Burger preaching sacrilege on the cover of Parade magazine a couple of weeks ago. A Time Magazine/Cable News Network poll reports that 87% of gun owners themselves favor a seven-day waiting period for handgun purchases; three-quarters favor registration of semi-automatic weapons and handguns, and half favor registration of rifles and shotguns.

Unfortunately, there is the Second Amendment to the Constitution: "A well regulated militia, being necessary to the security of a free State, the right of the people to keep and bear arms, shall not be infringed."

Most right thinkers take comfort in that funny stuff about the militia. Since the amendment's stated purpose is arming state militias, they reason, it creates no individual right to own a gun. That reasoning is good enough for the ACLU. But would civil-libertarians be so stinting about an amendment they felt more fond of? Say, the First?

The purpose of the First Amendment's free-speech guarantee was pretty clearly to protect political discourse. But liberals reject the notion that free speech is therefore limited to political topics, even broadly defined. True, that purpose is not inscribed in the amendment itself. But why leap to the conclusion that a broadly worded constitutional freedom ("the right of the people to keep and bear arms") is narrowly limited by its stated purpose, unless you're trying to explain it away? A colleague says that if liberals interpreted the Second Amendment the way they interpret the rest of the Bill of Rights, there would be law professors arguing that gun ownership is mandatory.

The most thorough parsing of the Second Amendment is a 1983 article in the Michigan Law Review by Don Kates, a gun enthusiast. Kates expends most energy demonstrating that at the time of the Bill of Rights, all able-bodied men were considered to be part of the "militia" and were expected to defend the state if necessary. I'm not sure this is as clinching an argument as Kates seems to think. The fact that once upon a time everyone was a member of the militia doesn't prove that everyone still has a right to a gun even after the composition of the militia has changed.

But Kates has other bullets in his belt. The phrase "right of the people" appears four other times in the Bill of Rights (including the First Amendment). In all these other cases, everyone agrees that it creates a right for individual citizens, not just some collective right of states as a whole. Kates also marshals impressive historical evidence that the Second Amendment, like other Bill of Rights protections, was intended to incorporate English common law rights of the time, which pretty clearly included the right to keep a gun in your home for reasons having nothing to do with the militia.

If there is a good reply to Kates's fusillade, the controllers haven't made it. Of course the existence of an individual right to own guns doesn't mean that it is absolute. What are the limits? In the Supreme Court's one 20th-Century treatment of the Second Amendment, it held somewhat ambiguously in 1939 that sawed-off shotguns aren't necessarily protected by the Constitution without proof that they are the kind of weapon a militia might have used.

Working from that decision and the common law, Kates says the amendment's protection should be limited to weapons "in common use among law-abiding people," useful for law enforcement or personal defense, and lineally descended from weapons known to the Framers. (No nuclear bombs.) He adds that they must be light enough for an ordinary person to carry ("bear"), and even that they can't be especially "dangerous or unusual." He says that the amendment places no limit on mandatory registration or laws against concealed weapons in public.

This list seems quite reasonable and moderate, though where it all comes from is not clear. In suggesting, for example, that it would be fine to ban automatic rifles but not semi-automatics, Kates is slicing the constitutional salami pretty thin. But in what I suspect was the main purpose of his exercise -- establishing that a flat ban on handguns would be hard to justify under the Constitution -- Kates builds a distressingly good case.

The downside of having a Bill of Rights is that the protection of individual rights usually entails social costs. This is as true of the Second Amendment as it is of the First, Fourth, Fifth and Sixth. The downside of having those rights inscribed in a Constitution, protected from the whims of majority rule, is that they can't be re-defined as life changes. It would be remarkable indeed if none of the Bill of Rights became less sensible and more burdensome with time.

Talking and writing are as central to American democracy as they ever were; shooting just isn't. Gun nuts are unconvincing (at least to me) in their attempts to argue that the individual right to bear arms is still as vital to freedom as it was in 1792. But the right is still there.

Opinion

Methods

  1. Extract Features. This will be X_train
  2. Compute ROGUE score for every sentence across all the documents. This is will be y_train
  3. This is now a Regression problem. Use Ridge as the initial Baseline Model
  4. Predict using the Model
  5. Select Sentences for Summarization based on a Salience Function
  6. Evaluate and Tune the Model
  7. Repeat steps 4,5,6 with Deep Models
  8. Identify the Best Model
  9. Generate Summaries using the Best Model

Validation Errors

Ridge

Ridge

Ridge

Ridge

Ridge

Ridge

Ridge

Test Accuracy:

Test

Sample Summaries :

Predicted Summary for doc 1 :

Second Amendment states (1791): "A well regulated militia, being necessary to the security of a free State, the right of the people to keep and bear Arms, shall not be infringed."

Real Summary for doc 1:

The Second Amendment to the Constitution reads: "A well regulated militia, being necessary to the security of a free state, the right of the people to keep and bear arms shall not be infringed". The Second Amendment was originated by some of the states as protection against the federal government. The word "people" is used collectively and refers to a political entity such as a state. The Supreme Court on several occasions in the past has ruled that this amendment did not create a federally- protected right of individuals to bear arms but rather, limited the power of the federal government.

Predicted Summary for doc 2 :

Johnson testified this week at a Canadian inquiry that his seven-year involvement with illegal performance-enhancing drugs included injections before the 1987 World Championships in Rome, where he set the existing world record of 9.83 seconds.

Real Summary for doc 2 :

US track and field officials said that Johnson should be stripped of his world record in the 100 meter dash. Johnson testified at a Canadian inquiry that he had been involved with illegal performance-enhancing drugs for seven years and had injections before he set the world record in the 1987 World Championship in Rome. His coach and physician had also testified at the inquiry. Carl Lewis said that while Johnson had not told the truth earlier, he was doing so now and speaking out against drugs and should be allowed back in the sport after he finishes his suspension.

Predicted Summary for doc 3 :

Exxon Corp. on Wednesday increased its estimate of the total 1989 costs of cleaning up the massive Alaskan oil spill to $2 billion and said it would take another $500-million charge in the fourth quarter to cover costs from what is now the most expensive environmental disaster in history.

Real Summary for doc 3 :

Exxon increased its estimate of the 1989 costs of cleaning up the most expensive environmental disaster in history to $2B and it could take another $500M. Exxon noted that these costs were the major reason for their lower net income for the year, but analysts said that Exxon's size would insulate it from serious long-term financial damage. Exxon's stock has lagged behind that of major oil firms and Exxon could suffer from continuing negative publicity from the Valdez spill and the subsequent heating oil spill into a waterway between New York and New Jersey.

Predicted Summary for doc 4 :

Reforestation would give wooded areas in Yellowstone a "five-year jump" on natural regrowth of the woods and would "get it green again," says John Davis, Willamette Industries Inc.'s general manager for Western timber logging operations in Lewiston, Idaho.

Real Summary for doc 4 :

Following the worst fire on record in Yellowstone National Park, biologist, Donald Despain examines the forest looking for signs of new life. By next spring, many seeds released from rock-hard pine cones will be fertilized by nutrients leaching into the soil from the now-packed ash, will sprout. Grasses and several types of wildflowers are already sprouting less than seven weeks after the fire. Next year, the ground cover replacing the park's burnt stands of aging lodgepole pines should provide a 30-fold increase in plant species. The new growth will attract a larger variety of birds and other animal life to the area.

Predicted Summary for doc 5 :

The director of the county Social Services Agency wants all Orange County residents, including illegal aliens, to be counted in the 1990 Census, and he will ask the Board of Supervisors to support his position.

Real Summary for doc 5 :

The Social Services director in Orange County wants all residents in the county to be counted in the 1990 Census. He will ask the Board of Supervisors to pass a resolution supporting his position. If the approximate 200,000 illegal aliens were not counted, the county would loose an estimated $56 million a year in federal revenue and lose representatives in Congress. On Thursday, the U.S. Senate and House agreed that illegal aliens should be counted. Before Thursday, the Senate had voted to bar the Census Bureau from counting illegal aliens. The House had twice rejected efforts to exclude aliens.

Conclusions / Future Work

Here's the main conclusions and a list of directions for improvement.

  1. Generate the summary instead of extracting it
  2. Train on all the DUC Datasets - DUC 2002, 2003....2006
  3. Regularize the Deep Models using Dropout.
  4. Learn the Features from Text instead of Hand-Crafting it
  5. Work on Query Relevant Summarization

About

Text summarization with deep regression models

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 98.8%
  • Other 1.2%