Automatic Sentence Compression in the MUSA project

1
Automatic Sentence Compression in the MUSA project

• Walter Daelemans Anja Höthker
• walter.daelemans_at_ua.ac.be
• http//cnts.uia.ac.be
• CNTS, University of Antwerp, Belgium
• Languages The Media 2004, Berlin

2
MUSA

• MUltilingual Subtitling of multimediA content
• EU IST 5th framework, Sep. 2002 - Feb. 2005
• Goals
• Conversion of audio streams into TV subtitles
  (monolingual)
• Translation of subtitles into French or Greek

3
(No Transcript)
4
Partners

• ILSP, Athens coordination, integration
• ESAT, KU Leuven Automatic Speech Recognition
• CNTS, U. Antwerp Sentence compression
• Systran, Paris Machine Translation
• BBC, London Main User, Data provider, Evaluation
• Lumiere, Athens Main User, Multilingual Data
  Provider, Evaluation

5
Goals for Sentence Compression

• Automatically and dynamically generate subtitles
  based on constraints (words and characters)
• Reduce the time needed for producing subtitles by
  expert subtitler
• Provide an architecture that can easily be ported
  to other languages

6
Example

• SPEECH
• The task force is in place and ready to attack
  without mercy.
• Constraints
• Delete 3 words and 14 characters
• Compression Module output
• The task force is in place and ready to fight
  without mercy .
• SUBTITLE
• The task force is ready ...
• ...to fight without mercy.

7
Approach

• Remove disfluencies compress sentence by
  removing repetitions introduced by hesitation
• I, I know that this war, this war will last for
  years
• Paraphrasing replace part of the input sentence
  by shorter paraphrase
• an increasing number of ? more and more
• Rule-Based Approach compress sentences based on
  handcrafted deletion rules that combine
• Shallow-parsing information (identifying
  constituents used by deletion rules)
• Relevance measures (determine in which order to
  delete constituents)

8
Shallow Parsing POS Tagging

• The/Det woman/NN will/MD give/VB Mary/NNP a/Det
  book/NN

9
Shallow Parsing Chunking

• The/Det woman/NNNP will/MD give/VBVP
  Mary/NNPNP a/Det book/NNNP

10
Shallow parsing Sense Tagging

• The/Det woman/NNNP-PERSON will/MD give/VBVP
  Mary/NNPNP-PERSON a/Det book/NNNP-MATERIAL-OBJ
  ECT

11
Shallow Parsing Relation Finding
person
material-object
person
12
MBSP (Perl)
Text In
Tokenizer (Perl)
MBT server POS Tagger
TiMBL server Known words
TiMBL server Relation Finder
TiMBL server Unknown words
MBT server Concept Tagger
Timbl server Phrase Chunker
TiMBL 5.0 MBT 2.0 http//ilk.uvt.nl/
TiMBL server Known words
TiMBL server Unknown words
13
Rule-Based Approach (syntax)

• Deletion rules mark phrases for deletion based on
  shallow parser output
• Rules for adverbs, adjectives, PNPs, subordinate
  sentences, interjections, ...
• Phrases are deleted iteratively until target
  compression rate is met

14

• Example Rule ADJECTIVES
• if(POS(word) JJ CHUNK(word) ! ADJP-END
  word-1 ! most least more less)
• delete(word)
• if (word-1CC word-2JJ)
• delete(word-1)
• elseif (word1CC word2JJ)
• delete(word1)
• Adam 's only serious childhood illness
  had been measles
• The virus triggered an 1 extremely 1 2 rare 3
  and 2 fatal 3 condition

15
Relevance Measures (semantics)

• Deletion rules suggest more deletions than
  necessary for reaching target compression
• System rates the different possibilities and
  starts with deleting the least important phrases
• Relevance measures in MUSA are based on (a
  weighted combination of)
• Word frequencies (in BNC corpus)
• Rule Probabilities (as encountered in parallel
  BBC corpus of transcripts with associated
  subtitles)
• Word Durations (compare estimates with actual
  durations)

16
Example

• This is a basic summarizer for English used for
  demonstration purposes.
• (NP This) is (NP a basic11 summarizer) (PNP for
  English)10 used (PNP for demonstration
  purposes)12.
• This is a basic summarizer used for demonstration
  purposes
• This is a summarizer used for demonstration
  purposes
• This is a summarizer used

17
Evaluation Data (Lumiere)

• MMR - every parents choice
• 243 segments
• 39.5 of the segments need compression
• Average target compression rate 4.58 words, 1.98
  chars
• The Tranquiliser Trap
• 287 segments
• 50.52 of the segments need compression
• Average target compression rate 3.21 words, 2.0
  chars

18
Human Evaluation
19
Conclusions

• We presented the Sentence Compression Module of
  the MUSA system
• Eclectic system combining statistical techniques
  for relevance detection with handcrafted deletion
  rules based on shallow parser output
• Evaluation suggests usefulness (with transcripts
  as input)
• Future Work
• Porting to other languages
• Machine Learning of paraphrases

20
Demos

• Sentence Compression http//cnts.uia.ac.be/cgi-bin
  /anja/musa
• MUSA demo http//sifnos.ilsp.gr/musa/demos