101035 中文信息处理 chinese nlp lecture 16. 应用 —— 自动摘要 automatic summarization...
TRANSCRIPT
101035 中文信息处理
Chinese NLP
Lecture 16
2
应用——自动摘要Automatic Summarization
• 背景简介( Background)
• 摘要类型( Types of Summary)
• 主流技术(Mainstream Techniques)• 摘要评测( Summarization Evaluation)• 研究前沿( Research Frontiers)
3
背景简介Background
• A summary is a short text produced from one text or texts, which conveys the most important information for generic use or for a particular task/user.
• Text summarization is the process of producing a summary for the given text(s).
• Automatic (text) summarization is the process of producing the summary by computer programs.
4
• Examples
5
• Why Summarization
• Information overload (Big text)
• Tens of Billions of Webpages
• PB-level data
• Related fields
• Information extraction
• Information retrieval
• Question answering
6
• Summarization Objectives
• Content
• Conveying the most important ideas in the original documents
• Length
• Considerably shorter than original or as required
• Presentation
• Readable, in natural coherent sentences instead of diagrams or a collection of key words
7
摘要类型Types of Summary
Form Input Purpose
Extract: composed of representative textual segments (usually sentences)
Single document
Generic
Abstract: rewritten version based on the salient information in the text
Multiple documents Query-focused
8
• Extract vs Abstract • Four score and seven years ago our fathers brought forth on this
continent, a new nation, conceived in Liberty, and dedicated to the proposition that all men are created equal. Now we are engaged in a great civil war, testing whether that nation, or any nation so conceived and so dedicated, can long endure. We are met on a great battle-field of that war. We have come to dedicate a portion of that field, as a final resting place for those who here gave their lives that that nation might live. It is altogether fitting and proper that we should do this. But, in a larger sense, we can not dedicate -- we can not consecrate -- we can not hallow -- this ground. The brave men, living and dead, who struggled here, have consecrated it, far above our poor power to add or detract. The world will little note, nor long remember what we say here, but it can never forget what they did here. It is for us the living, rather, to be dedicated here to the unfinished work which they who fought here have thus far so nobly advanced. It is rather for us to be here dedicated to the great task remaining before us -- that from these honored dead we take increased devotion to that cause for which they gave the last full measure of devotion -- that we here highly resolve that these dead shall not have died in vain -- that this nation, under God, shall have a new birth of freedom -- and that government of the people, by the people, for the people, shall not perish from the earth.
9
• Extract vs Abstract Extract
Four score and seven years ago our fathers brought forth on this continent, a new nation, conceived in Liberty, and dedicated to the proposition that all men are created equal. Now we are engaged in a great civil war, testing whether that nation, or any nation so conceived and so dedicated, can long endure. We are met on a great battle-field of that war. We have come to dedicate a portion of that field, as a final resting place for those who here gave their lives that that nation might live. It is altogether fitting and proper that we should do this. But, in a larger sense, we can not dedicate -- we can not consecrate -- we can not hallow -- this ground. The brave men, living and dead, who struggled here, have consecrated it, far above our poor power to add or detract. The world will little note, nor long remember what we say here, but it can never forget what they did here. It is for us the living, rather, to be dedicated here to the unfinished work which they who fought here have thus far so nobly advanced. It is rather for us to be here dedicated to the great task remaining before us -- that from these honored dead we take increased devotion to that cause for which they gave the last full measure of devotion -- that we here highly resolve that these dead shall not have died in vain -- that this nation, under God, shall have a new birth of freedom -- and that government of the people, by the people, for the people, shall not perish from the earth.
AbstractThis speech by Abraham Lincoln commemorates soldiers who laid down their lives in the Battle of Gettysburg. It reminds the troops that it is the future of freedom in America that they are fighting for.
10
• Summarization Process
Selecting the most important
content
Ordering and structuring the
extracted content
Generating a summary from the selected sentences
11
主流技术Mainstream Techniques
• Sentence Ranking / Selection • Feature-based (Unsupervised)
• Graph-based (Unsupervised)
• Supervised
• Redundancy Control
• Maximum marginal relevance
• Sentence ordering
• Textual ordering (single-document)
• Chronological ordering (multi-document)
• Majority ordering
12
• Feature-Based Method• Process
• Defining a set of features for the sentences
• Using the features to estimate the importance of the sentences by a scoring function
• Ranking the sentences by the importance scores
• Useful features
• TF
• TF.IDF
• Position
• Cue words
• …
13
• Frequency Features• What is more important in the given texts?
• Important terms (words) often appear frequently.
• Sentences with frequent terms (words) should be important.
• Stopwords should be discounted.
• Sentence length should be considered.
• Algorithm
• Using TF.IDF to calculate the importance of words
• Using the sum of word importance as the sentence importance
• Optionally, using measures other than TF.IDF
• Ranking sentences
14
In-Class Exercise
• Although TF.IDF can be replaced by other measures of importance, it is still the most widely used. The reason is ________________.
A) it is relatively easy to calculate the tf and idf statistics
B) no deep NLP is needed
C) it naturally disfavors stopwords
D) all of the above
15
• A Simple Algorithm
• Counting the frequency of words
score(w) = freq(w)
• Calculating the sum of the word scores, normalized by sentence length
score(s) = ∑w in sscore(w) /|s|
• Ranking the sentences by score(s)
16
• Position• Important information tends to be at the beginning of
text
• Good heuristic for news documents
• Use
• Using the position of sentences in a document to calculate its importance
E.g., score(s) = 1 / pos(s); score(s) = 1 – pos(s) / N
• An extreme case: lead-based summarization
• Selecting the first K sentences from the documents
• Effective for single-document summarization
• Very common in practical applications
17
• Redundancy Control
• Sometimes sentences may contain redundant information. After ranking the sentences, consider sentence redundancy.
• Maximum Marginal Relevance (MMR)
• When a sentence is selected, the scores of the remaining sentences are discounted by the similarity to the selected sentence.
C = doc collectionQ = user queryR = IR(C,Q,)S = already retrieved documents
18
• Sentence Ordering
• The selected sentences may not be well arranged.
• Single document summarization
• Following the original textual order
• Multi-document summarization
• Chronological ordering
19
摘要评测Summarization Evaluation
• Intrinsic Evaluation(内部评测)• Automatic
• Semi-automatic
• Manual
• Extrinsic Evaluation(外部评测)
20
• Intrinsic Evaluation• Summary quality is evaluated directly.
• Automatic evaluation
• ROUGE
• BE
• Semi-automatic evaluation
• Manual evaluation
21
• Automatic Methods• ROUGE (Recall-Oriented Understudy of Gisting
Evaluation)
• Inspired by BLEU, from machine translation
• ROUGE-N (Ngram matching, N = 1, 2)
• ROUGE-L
• ROUGE-SU4
22
In-Class Exercise
• According to the reference summary (left) and the system summary (right). What are the ROUGE-1 and ROUGE-2 using character-based ngrams (ignoring non-Chinese characters)?
港珠澳大桥拱北隧道管幕工程近日成功。
近日,港珠澳大桥拱北隧道成功完成。
23
• Automatic Methods• BE (Basic Elements)
• Similar to ROUGE, but use basic syntactic units (BE) instead of ngrams
• Pyramid
• Manually determine SCUs (Semantic Content Unit)
• Calculate SCU coverage
• Manual
• Grammaticality, non-redundancy, referential clarity, focus, structure and coherence
24
• Extrinsic Evaluation
• Summary quality is evaluated via its performance in another task.
• Goal-oriented
• Example
• Compare summaries and the original documents in solving GMAT reading comprehension questions (Morris et al., 1992)
25
研究前沿Research Frontiers
• User-customized Summarization
• Quality-oriented Summarization
• Inexpensive Abstractive Summarization
• Big and Noisy Text Summarization
• Special Genre (Non-news) Summarization
26
• Demos• AutoSummarize in MS Word (<2010)
• Mining Essential
• https://essential-mining.com/es/index.jsp?ui.lang=en
• Columbia NewsBlaster
• http://newsblaster.cs.columbia.edu/index.html
• A Chinese Product Demo
• http://lietu.com/demo/Summary.jsp
27
• 背景简介• 摘要类型• Extract vs Abstract
• Summarization Process
• 主流技术• Sentence Ranking
• Redundancy Control
• Sentence Ordering
Wrap-Up
• 摘要评测• Intrinsic Evaluation
• Extrinsic Evaluation
• 研究前沿