vtechworks.lib.vt.edu

vtechworks.lib.vt.edu

CS 4984/5984 Big Data Text Summarization Taught By: Edward A. Fox Virginia Tech Blacksburg, VA 24061 Abstractive Text Summarization of the Parkland Shooting Collection Ryan Kingery, Jiacheng Ye, Li Huang, Chao Xu, Sudha Yellapantula Submitted on December 10, 2018

Funding Source: NSF: IIS-1619028 Introduction Goal: Use deep learning techniques to generate a decent abstractive summary of the Parkland shooting web collection. Complications:

Dataset had to be converted from WARC file to text files. Dataset was extremely noisy. Denoised data was too homogeneous to easily segment. Deep learning models only trained on individual articles, not collections. Training necessary deep learning models from scratch was prohibitively slow. Most of team had no real prior experience with ML, NLP, DL.

Collection Summarization Pipeline Doc2Vec Embeddings + K-Means Clustering https://cs.stanford.edu/~quocle/ paragraph_vector.pdf Sequence-to-Sequence Model https://arxiv.org/pdf/

Pointer Generator Network https://arxiv.org/pdf/1704.04368.pdf Reinforced Extractor-Abstractor Network https://arxiv.org/pdf/1805.11080.pdf Example Representative Article

Presidential Proclamation Honoring the Victims of the Tragedy in Parkland, Florida: Our Nation grieves with those who have lost loved ones in the shooting at the Marjory Stoneman Douglas High School in Parkland, Florida. As a mark of solemn respect for the victims of the terrible act of violence perpetrated on February 14, 2018, by the authority vested in me as President of the United States by the Constitution and the laws of the United States of America, I hereby order that the flag of the United States shall be flown at half-staff at the White House and upon all public buildings and grounds, at all military posts and naval stations, and on all naval vessels of the Federal Government in the District of Columbia and throughout the United States and its Territories and possessions until sunset, February 19, 2018. I also direct that the flag shall be flown at half-staff for the same length of time at all United States embassies, legations, consular offices, and other facilities abroad, including all military facilities and naval vessels and stations. IN WITNESS WHEREOF, I have hereunto set my hand this fifteenth day of February, in the year of our Lord two thousand eighteen, and of the Independence of the

United States of America the two hundred and forty-second. Generated Summaries for Representative Article The victims of the united states shall be flown at half-staff at the white house. The united states of america, i hereby order that the flag of the united states shall be flown at half-staff at the white house. The flag of the united states shall be flown at half-staff for the same length of time.

Seq2Seq Presidential proclamation honoring the victims of the tragedy in parkland, florida. As a mark of solemn respect for the victims of the terrible act of violence perpetrated on february 14, 2018, by the authority vested in me as president of the united states of america, i hereby order that the flag of the united states shall be flown at halfstaff at the white house and upon all public buildings and grounds, at all military posts and naval stations. In witness whereof, i have hereunto set my hand this fifteenth day of february, in the year of our lord two thousand eighteen years.

PGN Use the transcript to help students with reading comprehension and vocabulary. The flag shall be flown at half-staff for the same length of time at all united states embassies, consular offices, and other facilities and naval vessels. He set my hand this fifteenth day of february, in the year of our lord two thousand eighteen. Presidential proclamation honoring the victims of the tragedy in parkland, florida our nation grieves with those who have lost loved ones. REAN

Subjective Rank-Based Summary Evaluations REAN with clusters was the best summary based on rankings alone. Seq2Seq summaries were universally seen as bad. Objective Summary Evaluations PGN with random articles was the best summary based on ROUGE-1, and gave the only ROUGE score that was higher than 0.1

Lessons Learned Abstractive summarization is still a long ways off before becoming anywhere near human-level summarization. Deep learning helps, but only superficially. Not all text datasets are easy to cluster semantically. Our dataset was nigh on impossible to do any coherent clustering or topic modeling on. Models trained on larger articles tend to give larger summeries. This suggests it wouldve perhaps been better to train on Wikipedia instead of CNN News. Using other peoples pretrained models was easier than training our own, but still wasnt easy due to build and formatting issues.

The REAN seems to capture more noise than the PGN, making PGN summaries seem more readable. Both are well above the Seq2Seq approach.

Recently Viewed Presentations

  • Learning from our students through learning assessments

    Learning from our students through learning assessments

    Articulate the value of a literature review. Differentiate between empirical research and review articles. Identify keywords, synonyms, and search terms. Search a variety of databases and resources. Locate related resources through forward and backward citation tracking. Knowledge-Based. Skills-Based
  • Toward a Science of Consciousness - University of Florida

    Toward a Science of Consciousness - University of Florida

    Explanatory gap: Why should the new physicalism entail experience? A multiverse of possibilities All that physics specifies is a measure on a set of possibilities or possible worlds. Slogan: Multiverse as a quantum computer (Lloyd 2006).
  • Critically Appraising a Journal Article

    Critically Appraising a Journal Article

    One of the most effective means by which students and professionals keep up with current biomedical literature. ... It is believed that Sir William Osler established the first formal journal club at McGill University in Montreal in 1875, though Osler...
  • PowerPoint Presentation Last modified by: Andrew J. Robison IV Created Date: 1/1/1601 12:00:00 AM Document presentation format: On-screen Show Other titles:
  • CIE generic PowerPoint template

    CIE generic PowerPoint template

    Meet the college-ready qualification score of "C" on 1 additional Cambridge IGCSE exam from English Literature, the sciences**, or history exam areas. Fine Art or Career and Technical Education (CTE) G. Meet the college-ready qualification score on IGCSE Art &...
  • Lesson Plans You Can Use Now! - National Consortium for ...

    Lesson Plans You Can Use Now! - National Consortium for ...

    Lesson Plan #7 Health ScreeningCo-curricular connection: Health, Anatomy and Physiology 1.13 Analyze basic structures and functions of Cardiovascular (components of blood, structures and functions of blood components, structures and functions of the cardiovascular system, conduction system of the heart, cardiac...
  • Система Для Внутрішньокісткових Інфузій Fast1®

    Система Для Внутрішньокісткових Інфузій Fast1®

    Пояснення системи fast1®. Використовує лише силу оператора - це не пістолет, не має пружини, не залежить від акумулятора, не пневматична.
  • What-ere you do in word or deed, do

    What-ere you do in word or deed, do

    What-e're you do in word or deed, do all in the name of the Lord. Do naught in name of man or creed, do all in the name of the Lord. Silence of the Scriptures