In this talk… What do we mean by Summarization

Document Summarization

• Madhavi Ganapathiraju Graduate Student Language Technologies Institute Carnegie Mellon University

In this talk…

• What do we mean by Summarization

• Expectations

• Intuitive guesses on “how to”

• Current approaches

• One specific method in detail

Objective of Summarization

• Reduce length of document

• But preserve:

• Key Information
• Style of writing

• Expected Qualities:

• Cohesiveness
• Coherence
• Readability

How can we say a summary is good or bad?

• Be able to answer questions

• Compression ratio

• Preserve Chronology

• No Redundancy

How is it done manually

• Read document

• Identify important “phrases”

• Identify Chronology of events if any

• Synthesis new sentences

How to do it automatically: Edmundson’s method

• His work at IBM

• In 1969!!
• Forms major component even in today’s systems!!!!

Edmundson’s method

Scoring schemes derived

• Keyword-occurrence

• Title-keyword

• Location heuristic

• Indicative phrases

• “this report …”, “in conclusion…”

• Short-length cutoff

• Upper-case word feature

• Acronyms, proper names

Graph theoretic method

How to put key information together?

• Synthesis new sentences?

• Too difficult… to synthesize accurately
• Systems exist
• Undesirable
• Original style of writing lost
• Subtle information like tone of presentation lost

Summary = collection of sentences

• Take top most scoring sentences

• Arrange them by descending scores

• Preserve chronology if exists

Redundancy

• Edmundson’s procedure:

• No explicit mention of avoiding redundancy
• The only difference to modern methods

• Novel methods to avoid redundancy

• Maximum “marginal relevance” (MMR)

Similarity between sentences

• Semester begins tomorrow

• New semester is beginning on Monday

• S1 = [Semester(1) begin(1) tomorrow(1)]

• S2 = [New(1) semester(1) begin(1) Monday(1)]

• Similarity 

MMR features

• Clusters of sentences

• Candidature of a sentence to be in summary:

• Similarity to query.
• Coverage of the passage
• Content in the passage, eg., proper nouns, dates, etc.
• Time Sequence: more recent ones

• Undesirable features in sentences:

• Similarity to passages already included in the summary
• Belonging to the cluster/document that has already contributed a sentence to the summary

MMR algorithm

Future presentations on Summarization & Contact persons for research in this area:

• Future presentations on Summarization & Contact persons for research in this area:

• Nikesh Garera ([e-mail])

• Learning Methods

• Ravindra G. ([e-mail])

• Statistical Methods