1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours:...

62
1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116 Office Hours: Thurs 12-1, 8-9
  • date post

    19-Dec-2015
  • Category

    Documents

  • view

    214
  • download

    0

Transcript of 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours:...

Page 1: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

1

Natural Language Processing for the Web

Prof. Kathleen McKeown

722 CEPSR, 939-7118

Office Hours: Wed, 1-2; Tues 4-5

TA:

Yves Petinot

719 CEPSR, 939-7116

Office Hours: Thurs 12-1, 8-9

Page 2: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

2

Projects

Proposal due today Hand in via courseworks (by midnight)

Page 3: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

3

Today

Discussants – Each person should sign up for TWO papers

Automated Discovery and Analysis of Social Networks from Threaded Discussions (Kathy)

From Social Bookmarking to Social Summarization (Kathy) Joint Group and Topic Discovery from Relations and Text

(Kathy) Discovering Authorities in Question Answering Communities

Using Link Analysis (Kenny)

Discussants: Lauren, Weiwei

Page 4: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

4

Automated Discovery and Analysis of Social Networks Threaded discussion:

Online class discussion board “…examining social networks – including the roles and positions

of actors in a social network, their influence on others, and what exchanges support and sustain the network – is an important goal for understanding networked learning processes”

Social Network background Most studies use meta-data about links This work uses nl analysis of text within

postings

Page 5: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

5

Who talks to whom? (ties)

Chain network Create a link from poster to previous poster Create a link from poster to thread starter

plus previous poster Create a link from poster to all previous

posters in a thread, decreasing weight with distance

Page 6: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

6

Page 7: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

7

Topic of this paper: identifying names in a posting Possible names

Previous poster (to line, direct reference, indirect reference, subject of discussion)

someone else entirely (author), current poster (self-reference, in address

line, signature)

Page 8: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

8

Methods for identifying “Who” Use of name lists Class roster Use of titles (Prof.), addresses (dear) Exclusion of 3 word capitalized sequences Confidence level (page vs “Page”) Mis-spellings (manual review and edits)

Results: Local vs Ling-Pipe Precision: .88 vs. .60 Recall: .66 vs. .68

Page 9: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

9

Methods for identifying “ties”: links

Chance of a tie proportional to the number of times each mentions the other as addressee or subject Add 1 to a poster and all names found in posting

More than one name: link name to userid using collocational analysis (A vs. P) Association type P (poster) A (addressee)

Information Exchange: essential social interaction Measure via content of exchange (vs. network structure) Information weight of an exchange (vs. discourse act

such as announcement) Yahoo Term Extractor to find content nouns

Page 10: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

10

Example

Keep in mind that google and other search technology are still evolving and getting better. I certainly don't believethat they will be as effective as a library in 2-5 years, but if they improve significantly, it will continue to be difficult for the public to perceive the difference.

Page 11: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

11

Evaluation

Metric: QAP correlation

Page 12: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

12

From Social Bookmarking to Social Summarization

Exploit user-created content Del.icio.us web page tags, Flick’r, review sites

Approach expands on query-focused summarization Extract bookmark tags for a page p: (b1, b2, ..) Issue a search with tags as query Extract snippets associated with p in result: S(bi, p) Normalize snippets Score each snippet according to frequency Rank order as summary

Page 13: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

13

Some details

Limit results of search to top N Normalize by

Determining overlap (like cosine) Match if overlap score above threshold T Take shorter sentence of a match

Determine frequency of selected snippets in search result to rank

Page 14: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

14

Page 15: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

15

Evaluation

Baselines: OTS, MEAD Metric: Rouge Set-up 1: SS used full set of tags,

average length = 24% Problem? Relative improvement:

SS to OTS: 31-39% relative improvement SS to MEAD: 24-29% relative improvement

Page 16: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

16

Set-up 2: Vary summary length from 10% to 50%

Page 17: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

17

Set-up 3: Community based summarization Use tags generated by a specific

community: skier community (“skiing” as seed bookmark) vs. a travel community (“travel” as seed bookmark)

Evaluation: recall in summary of terms in seed set

Page 18: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

18

Page 19: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

19

Joint Group and Topic Discovery from Relations and Text

Example: legislative body and alliances Different alliances may form depending on the

resolution topic (taxation vs. foreign trade)

GT model Discovery of groups guided by emerging topics Discovery of topics guided by emerging groups Example: resolutions that would have been assigned

to one group based on topic may be assigned to different one given voting patterns; distinct word-based topics may be merged if entities vote similarly.

Page 20: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

20

GT Model

Simultaneously clusters entities into groups and words into topics

Data set: voting data from the US Senate and the UN General Assembly

Page 21: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

21

Sentence extraction

Sparck Jones:

`what you see is what you get’, some of what is on view in the source text is transferred to constitute the summary

Page 22: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

22

Background

Sentence extraction the main approach

Some more sophisticated features for extraction in recent years

Lexical chains, anaphoric reference, topic signatures

Machine learning models for learning an extraction summarizer (e.g., Kupiec)

Page 23: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

23

Today’s systems

How can we edit the selected text?

Page 24: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

24

Karen Sparck JonesAutomatic Summarizing: Factors and Directions

Page 25: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

25

Sparck Jones claims

Need more power than text extraction and more flexibility than fact extraction (p. 4)

In order to develop effective procedures it is necessary to identify and respond to the context factors, i.e. input, purpose and output factors, that bear on summarising and its evaluation. (p. 1)

It is important to recognize the role of context factors because the idea of a general-purpose summary is manifestly an ignis fatuus. (p. 5)

Similarly, the notion of a basic summary, i.e., one reflective of the source, makes hidden fact assumptions, for example that the subject knowledge of the output’s readers will be on a par with that of the readers for whom the source was intended. (p. 5)

I believe that the right direction to follow should start with intermediate source processing, as exemplified by sentence parsing to logical form, with local anaphor resolutions

Page 26: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

26

Questions (from Sparck Jones)

Does subject matter of the source influence summary style (e.g, chemical abstracts vs. sports reports)?

Should we take the reader into account and how?

Is the state of the art sufficiently mature to allow summarization from intermediate representations and still allow robust processing of domain independent material?

Page 27: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

27

Consider the papers we read in light of Sparck Jones’ remarks on the influence of context: Input

Source form, subject type, unit Purpose

Situation, audience, use Output

Material, format, style

Page 28: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

28

Cut and Paste in Professional Summarization

Humans also reuse the input text to produce summaries

But they “cut and paste” the input rather than simply extract automatic corpus analysis (Zipf Davis)

300 summaries, 1,642 sentences 81% sentences were constructed by cutting

and pasting

Page 29: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

29

Major Cut and Paste Operations (1) Sentence reduction

~~~~~~~~~~~~

Page 30: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

30

Major Cut and Paste Operations (1) Sentence reduction

~~~~~~~~~~~~

Page 31: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

31

Major Cut and Paste Operations (1) Sentence reduction

(2) Sentence Combination

~~~~~~~~~~~~

~~~~~~~~~~~~~~ ~~~~~~

Page 32: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

32

(3) Generalization

"a proposed new law that would require Web publishers to obtain parental consent before collecting personal information from children" -> "legislation to protect children's privacy on-line"

Page 33: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

33

Cut and Paste Based Single Document Summarization -- System Architecture

Extraction

Sentence reduction

Generation

Sentence combination

Input: single document

Extracted sentences

Output: summary

Zipf DavisCorpus

Decomposition

Lexicon

Parser

Co-reference

Page 34: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

34

Sentence Reduction Step 1: Use linguistic knowledge to decide what

phrases MUST NOT be removed Obligatory arguments of verbs are saved

Step 2: Determine what phrases are most important in the local context Phrases with words that link forward or backward

Step 3: Compute the probabilities of humans removing a certain type of phrase

Step 4: Combine the three factors to decide

Page 35: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

35

Sentence Fusion for Multi-document Summarization http://newsblaster.cs.columbia.edu

Page 36: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

36

Page 37: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

37

Fusion

Page 38: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

38

Sentence Fusion Computation: Content Selection

Common information identification Alignment of constituents in parsed

theme sentences: only some subtrees match

Bottom-up local multi-sequence alignment

Similarity depends on Word/paraphrase similarity Tree structure similarity

Page 39: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

39

Sim(T,T’) = max (nodecompare(T,T’), Sim(T, children(T’)), Sim(children(T),T’))

Nodecompare searches for best possible alignment of all childnodes

Nodesimilarity dependson similarity between words of atomic nodes

Page 40: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

40

Page 41: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

41

Sentence Fusion: Generation

Fusion lattice computation Choose a basis sentence Add subtrees from fusion not present in basis Add alternative verbalizations Remove subtrees from basis not present in

fusion Lattice linearization

Generate all possible sentences from the fusion lattice

Score sentences using statistical language model

Page 42: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

42

Page 43: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

43

Questions

Jing: Not a statistical approach, not learned. Is this OK? Does it buy us anything over the approaches using learning?

Barzilay: Also not statistical, OK? How to compare with Jing? Is redundancy a good criteria for content selection? What could go wrong?

Page 44: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

44

Sparck Jones claims

Need more power than text extraction and more flexibility than fact extraction (p. 4)

In order to develop effective procedures it is necessary to identify and respond to the context factors, i.e. input, purpose and output factors, that bear on summarising and its evaluation. (p. 1)

It is important to recognize the role of context factors because the idea of a general-purpose summary is manifestly an ignis fatuus. (p. 5)

Similarly, the notion of a basic summary, i.e., one reflective of the source, makes hidden fact assumptions, for example that the subject knowledge of the output’s readers will be on a par with that of the readers for whom the source was intended. (p. 5)

I believe that the right direction to follow should start with intermediate source processing, as exemplified by sentence parsing to logical form, with local anaphor resolutions

Page 45: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

45

Supervised and Unsupervised Learning for Sentence Compression

J. Turner and E. Charniak

Page 46: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

46

Knight and Marcu Model

Noisy Channel Model Zipf Davis corpus Given a long sentence, determine the short

sentence that maximizes P(s|l) Bayes rule:

P(l) is constant across all long, dropped

Language model combination of PCFG and bigram of S

Page 47: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

47

Two problems with K&M

Lack of training data – Why?

Probability model is ad hoc

Page 48: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

48

Turner and Charniak Approach – K&M modification Use syntactic language model Slightly change channel model:

Parameter to encourage compression

Page 49: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

49

Alternate models “Special rule” additions + K&M variation

NP(1) -> NP(2) CC NP (3) compressed to NP(2) Unsupervised version using PTB: no parallel corpus.

P(l|s) learned by comparing similar rules

NP -> DT JJ NN (3X) NP -> DT NN (4X) P(l|s) = 3/7

Semi-supervised: fall back on unsupervised when no data from supervised

Constraints: complement/adjunct distinction: never allow deletion of complement

Page 50: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

50

Results (evaluated using judges)

Page 51: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

51

Questions

How does this compare with Jing? Will same manual rules be captured?

Verb arguments not deleted? Context determines importance

What does statistics capture that is not captured by the manual approach?

How about revisions other than reduction?

Page 52: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

52

Compression Beyond Word Deletion Cohn and Lapata

Page 53: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

53

Goal and Approach

To delete, substitute, re-order

Collect a new corpus: why? 30 newspaper articles, 575 sentences Is this adequate?

Extract compressions

Collect paraphrases using MT

Page 54: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

54

Abstraction Example

High winds and snowfalls have, however, grounded at a lower level the powerful US Navy Sea Stallion helicopters used to transport the slabs.

Bad weather, however, has grounded the helicopters transporting the slabs.

Page 55: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

55

Extraction of compression rules Synchronous Tree Substitution

Grammar (S,S) -> (NPVBD NP, NP was VBN by NP)

Probabilistic (each grammar rule assigned a learned weight)

Prediction: generation finds the best scoring compression using the grammar rules

(Skip training section)

Page 56: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

56

Extension (contribution)

Paraphrasing with their corpus a problem

Learn paraphrase grammar rules Parallel bilingual corpus Learns over syntax tree fragments Translate from English to French and back

again -> an English paraphrase of original

These rules are added into extracted compression grammar

Page 57: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

57

Combined grammar

Incorporates an ngram language model as a feature Helps prevent ungrammatical output

Like K&M, Turner and Charniak, a parameter to penalize short output

Union of compression plus paraphrasing grammar plus a COPY grammar derived from the source side

Page 58: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

58

ResultsModels Grammaticality Importance Comp Rate

extract 3.1 2.43 82.5

abstract 3.38 2.85 79.2

gold 4.51 4.02 58.4

O: The scheme was intended for people of poor or moderate means.E: The scheme was intended for people of poor means.A: The scheme was intended for poor people.G: The scheme was intended for the poor.

O: He died last Thursday at his home from complications following a fall, said his wife author Margo Kurtz.E: He died last at his home from complications following a fall, said wife, author Margo Kurtz.A: His wife author Margo Kurtz died from complications after a decline.G: He died from complications following a fall.

Page 59: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

59

Quesitons

Is this comparable to K&M, Turner and Charniak?

Is it OK to take a risk? What are the weak points?

Page 60: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

60

Page 61: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

61

Page 62: 1 Natural Language Processing for the Web Prof. Kathleen McKeown 722 CEPSR, 939-7118 Office Hours: Wed, 1-2; Tues 4-5 TA: Yves Petinot 719 CEPSR, 939-7116.

62