際際滷

際際滷Share a Scribd company logo
Automatic Extraction of Arabic
       Multiword Expressions
*Mohammed Attia, Antonio Toral, Lamia Tounsi, Pavel Pecina and
                    Josef van Genabith
     School of Computing, Dublin City University, Ireland
Outline
¢   Introduction
¢   Data Resources
¢   Methodology
    ¢   Crosslingual Correspondence Asymmetries
    ¢   Translation-Based Approach
    ¢   Corpus-Based Approach
¢   Discussion of experiments and results
¢   Conclusion
Introduction
¢   Criteria of MWEs
    ¢   Ubiquity
    ¢   Diversity
    ¢   Low polysemy
    ¢   Statistically significant co-occurrence
¢   Focus
    ¢   Arabic
    ¢   Nominal MWEs
¢   Purpose is building an MWE lexicon for Arabic
Data Resources
?   Multilingual, bilingual and monolingual settings
?   Availability of rich resources that have not been
    exploited in similar tasks before.
¢   Arabic Wikipedia (March 2010)
    ¢   117,491 titles, of them 89,623 multiword titles
    ¢   Arabic is ranked 27th according to size (article count) and
        17th according to usage
    ¢   Information helpful for linguistic processing
Data Resources
¢   Princeton WordNet 3.0
    ¢   An electronic lexical database for English
    ¢   Arabic WordNet contains only 11,269 synsets (including
        2,348 MWEs)
Data Resources
¢   Arabic Gigaword
    ¢   Unannotated corpus distributed by the Linguistic Data
        Consortium (LDC).
    ¢   Articles from news agencies and newspapers from different
        Arab regions, such as Al-Ahram in Egypt, An Nahar in
        Lebanon and Assabah in Tunisia.
    ¢   Largest publicly available corpus of Arabic to date.
    ¢   Contains 848 million words.
Methodology
3 different techniques for 3 different data sources


Motivation for using different techniques
   ¢   The extraction of MWEs is a problem more complex than
       can be dealt with by one simple solution.
   ¢   The choice of technique depends on the nature of the task
       and the type of the resources used.
Pipeline
Technique 1: Crosslingual Asymmetries

¢   Data: Titles of Wikipedia Articles in Arabic and corresponding
    titles in 21 languages.
¢   Definition: We rely on many-to-one correspondence relations
¢   The non-compositionality of MWEs makes it unlikely to have
    a mirrored representation in the other languages.
¢   Compositionalily varies:
    ¢   highly compositional, "?" ,"????? ???????military base",
    ¢   with a degree of idiomaticity, such as, "?" ,"????? ???????amusement
        park", lit. "city of amusements".
    ¢   extremely opaque , "?" ,"??? ??????grasshopper", lit. "the horse of the
        Prophet".
Technique 1: Crosslingual Asymmetries

¢   Steps
    (1) Candidate Selection. All Arabic Wikipedia multiword titles
       are taken as candidates.
    (2) Filtering. We exclude titles of disambiguation and
       administrative pages.
    (3) Validation. We check if there is a single-word translation in
       any of 21 selected languages.
Technique 1: Crosslingual Asymmetries

¢   Evaluation:
    ¢   1100 multiword titles are randomly selected from Arabic
        Wikipedia and manually tagged as: MWEs, non-MWEs, or
        NEs.
    ¢   Baseline: all multi-word titles are considered as MWEs
¢   Results
Example
Language Ranking

How likely will each language give many-to-one correspondence?
Technique 2: Translation-Based

¢   Data: Princeton WordNet
    ¢   Assumption: MWEs in one language are likely to be
        translated as MWE in another language.
    ¢   Ontological advantage
¢   Steps
    ¢   Extracting the list of nominal MWEs from PWN 3.0.
    ¢   Translating the list into Arabic using Google Translate.
    ¢   Validating the results using pure frequency counts from three
        search engines: Al-Jazeera, BBC Arabic and AWK.
Technique 2: Translation-Based

¢   Evaluation (automatic)
    ¢   Gold Standard: PWN-MWEs found in English Wikipedia and have
        correspondence in Arabic: 6322 expressions.
    ¢   We test the Google translation without any filtering, and consider this as
        the baseline.
    ¢   Then we filter the output based on the number of combined hits from the
        search engines.


¢   Results
Technique 2: Translation-Based

¢   Evaluation (Manual)
    ¢   On 200 MWE candidates
    ¢   Precision
         C   Baseline (before validation): 45.5%
         C   After validation: 83%
Technique 2: Translation-Based

¢   Notes on Google Translate
    ¢   Word Order
         C   shark repellent      =>        ?????? ?????
         C   accordion door       =>        ?????????? ??????
    ¢   Transferring source word to target
         C   acroclinium roseum =>             acroclinium roseum
         C   actitis hypoleucos        =>       actitis hypoleucos
Technique 3: Corpus-Based

¢   Data: Arabic Gigaword corpus
¢   Association Measures used:
    ¢   Pointwise Mutual Information (PMI)
    ¢   Pearson¨s chi-square
¢   Steps
    (1) Compute the frequency of all the unigrams, bigrams, and trigrams
    (2) Computing the association measures for all bigrams and trigrams (threshold to 50)
    (3) Ranking bigrams and trigrams
    (4) Conducting lemmatization of Arabic words using MADA.
    (5) Filtering the list using the MADA POS-tagger. The patterns included for bigrams are: NN NA, and for
      trigrams: NNN NNA NAA
Technique 3: Corpus-Based

¢   Why is lemmatization important?
    ¢   Al>mm AlmtHdp
        (the-nations united) ^the United Nations ̄
        Al>mm@>um~ap_1@N@1#AlmtHdp@mut~aHid_1@AJ@2#


    ¢   ll>mm AlmtHdp
        (to-the-nations united) ^to the United Nations ̄
        ll>mm@>um~ap_1@N@3#AlmtHdp@mut~aHid_1@AJ@3#


    ¢   wAl>mm AlmtHdp
        (and-the-nations united) ^and the United Nations ̄
        wAl>mm@>um~ap_1@c-N@3#AlmtHdp@mut~aHid_1@AJ@3#


    ¢   bAl>mm AlmtHdp
        (by-the-nations united) ^by the United Nations ̄
        bAl>mm@>um~ap_1@N@3#AlmtHdp@mut~aHid_1@AJ@3#
Technique 3: Corpus-Based

¢   Evaluation: 3600 expressions are randomly selected
    and classified into MWE or non-MWE by a human
    annotator.
¢   Results
Discussion results
¢   Combination of yields
Discussion of results
¢   Similarities and dissimilarities of output
The set of collocations detected by the association
measures may differ from the those which capture the
interest of lexicographers and Wikipedians
    ¢   ?????? ??????     ^Menachem Mazuz ̄
    ¢   ??????? ??????    ^fresh fruits ̄
    ¢   ??????? ???????   ^Ladies and gentlemen ̄
Conclusion
¢   Applicability to other languages
¢   the heterogeneity of the data sources helps to enrich
    the MWE lexicon.
¢   A lexical resource of:
    ¢   33,000 MWEs
    ¢   39,000 NEs
Thank you!

More Related Content

Arabic mwe presentation 07

  • 1. Automatic Extraction of Arabic Multiword Expressions *Mohammed Attia, Antonio Toral, Lamia Tounsi, Pavel Pecina and Josef van Genabith School of Computing, Dublin City University, Ireland
  • 2. Outline ¢ Introduction ¢ Data Resources ¢ Methodology ¢ Crosslingual Correspondence Asymmetries ¢ Translation-Based Approach ¢ Corpus-Based Approach ¢ Discussion of experiments and results ¢ Conclusion
  • 3. Introduction ¢ Criteria of MWEs ¢ Ubiquity ¢ Diversity ¢ Low polysemy ¢ Statistically significant co-occurrence ¢ Focus ¢ Arabic ¢ Nominal MWEs ¢ Purpose is building an MWE lexicon for Arabic
  • 4. Data Resources ? Multilingual, bilingual and monolingual settings ? Availability of rich resources that have not been exploited in similar tasks before. ¢ Arabic Wikipedia (March 2010) ¢ 117,491 titles, of them 89,623 multiword titles ¢ Arabic is ranked 27th according to size (article count) and 17th according to usage ¢ Information helpful for linguistic processing
  • 5. Data Resources ¢ Princeton WordNet 3.0 ¢ An electronic lexical database for English ¢ Arabic WordNet contains only 11,269 synsets (including 2,348 MWEs)
  • 6. Data Resources ¢ Arabic Gigaword ¢ Unannotated corpus distributed by the Linguistic Data Consortium (LDC). ¢ Articles from news agencies and newspapers from different Arab regions, such as Al-Ahram in Egypt, An Nahar in Lebanon and Assabah in Tunisia. ¢ Largest publicly available corpus of Arabic to date. ¢ Contains 848 million words.
  • 7. Methodology 3 different techniques for 3 different data sources Motivation for using different techniques ¢ The extraction of MWEs is a problem more complex than can be dealt with by one simple solution. ¢ The choice of technique depends on the nature of the task and the type of the resources used.
  • 9. Technique 1: Crosslingual Asymmetries ¢ Data: Titles of Wikipedia Articles in Arabic and corresponding titles in 21 languages. ¢ Definition: We rely on many-to-one correspondence relations ¢ The non-compositionality of MWEs makes it unlikely to have a mirrored representation in the other languages. ¢ Compositionalily varies: ¢ highly compositional, "?" ,"????? ???????military base", ¢ with a degree of idiomaticity, such as, "?" ,"????? ???????amusement park", lit. "city of amusements". ¢ extremely opaque , "?" ,"??? ??????grasshopper", lit. "the horse of the Prophet".
  • 10. Technique 1: Crosslingual Asymmetries ¢ Steps (1) Candidate Selection. All Arabic Wikipedia multiword titles are taken as candidates. (2) Filtering. We exclude titles of disambiguation and administrative pages. (3) Validation. We check if there is a single-word translation in any of 21 selected languages.
  • 11. Technique 1: Crosslingual Asymmetries ¢ Evaluation: ¢ 1100 multiword titles are randomly selected from Arabic Wikipedia and manually tagged as: MWEs, non-MWEs, or NEs. ¢ Baseline: all multi-word titles are considered as MWEs ¢ Results
  • 13. Language Ranking How likely will each language give many-to-one correspondence?
  • 14. Technique 2: Translation-Based ¢ Data: Princeton WordNet ¢ Assumption: MWEs in one language are likely to be translated as MWE in another language. ¢ Ontological advantage ¢ Steps ¢ Extracting the list of nominal MWEs from PWN 3.0. ¢ Translating the list into Arabic using Google Translate. ¢ Validating the results using pure frequency counts from three search engines: Al-Jazeera, BBC Arabic and AWK.
  • 15. Technique 2: Translation-Based ¢ Evaluation (automatic) ¢ Gold Standard: PWN-MWEs found in English Wikipedia and have correspondence in Arabic: 6322 expressions. ¢ We test the Google translation without any filtering, and consider this as the baseline. ¢ Then we filter the output based on the number of combined hits from the search engines. ¢ Results
  • 16. Technique 2: Translation-Based ¢ Evaluation (Manual) ¢ On 200 MWE candidates ¢ Precision C Baseline (before validation): 45.5% C After validation: 83%
  • 17. Technique 2: Translation-Based ¢ Notes on Google Translate ¢ Word Order C shark repellent => ?????? ????? C accordion door => ?????????? ?????? ¢ Transferring source word to target C acroclinium roseum => acroclinium roseum C actitis hypoleucos => actitis hypoleucos
  • 18. Technique 3: Corpus-Based ¢ Data: Arabic Gigaword corpus ¢ Association Measures used: ¢ Pointwise Mutual Information (PMI) ¢ Pearson¨s chi-square ¢ Steps (1) Compute the frequency of all the unigrams, bigrams, and trigrams (2) Computing the association measures for all bigrams and trigrams (threshold to 50) (3) Ranking bigrams and trigrams (4) Conducting lemmatization of Arabic words using MADA. (5) Filtering the list using the MADA POS-tagger. The patterns included for bigrams are: NN NA, and for trigrams: NNN NNA NAA
  • 19. Technique 3: Corpus-Based ¢ Why is lemmatization important? ¢ Al>mm AlmtHdp (the-nations united) ^the United Nations ̄ Al>mm@>um~ap_1@N@1#AlmtHdp@mut~aHid_1@AJ@2# ¢ ll>mm AlmtHdp (to-the-nations united) ^to the United Nations ̄ ll>mm@>um~ap_1@N@3#AlmtHdp@mut~aHid_1@AJ@3# ¢ wAl>mm AlmtHdp (and-the-nations united) ^and the United Nations ̄ wAl>mm@>um~ap_1@c-N@3#AlmtHdp@mut~aHid_1@AJ@3# ¢ bAl>mm AlmtHdp (by-the-nations united) ^by the United Nations ̄ bAl>mm@>um~ap_1@N@3#AlmtHdp@mut~aHid_1@AJ@3#
  • 20. Technique 3: Corpus-Based ¢ Evaluation: 3600 expressions are randomly selected and classified into MWE or non-MWE by a human annotator. ¢ Results
  • 21. Discussion results ¢ Combination of yields
  • 22. Discussion of results ¢ Similarities and dissimilarities of output The set of collocations detected by the association measures may differ from the those which capture the interest of lexicographers and Wikipedians ¢ ?????? ?????? ^Menachem Mazuz ̄ ¢ ??????? ?????? ^fresh fruits ̄ ¢ ??????? ??????? ^Ladies and gentlemen ̄
  • 23. Conclusion ¢ Applicability to other languages ¢ the heterogeneity of the data sources helps to enrich the MWE lexicon. ¢ A lexical resource of: ¢ 33,000 MWEs ¢ 39,000 NEs