본문

서브메뉴

Arabic Native Language Identification
Arabic Native Language Identification
Arabic Native Language Identification

Detailed Information

자료유형  
 학위논문 서양
최종처리일시  
20250211152807
ISBN  
9798384027027
DDC  
400
저자명  
Bassas, Yasmeen.
서명/저자  
Arabic Native Language Identification
발행사항  
[Sl] : Indiana University, 2024
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2024
형태사항  
177 p
주기사항  
Source: Dissertations Abstracts International, Volume: 86-02, Section: A.
주기사항  
Advisor: Kuebler, Sandra.
학위논문주기  
Thesis (Ph.D.)--Indiana University, 2024.
초록/해제  
요약Native Language Identification (NLI) is concerned with predicting the native language of an author's writing in a second language. NLI relies on the assumption that speakers of the same native language display certain linguistic patterns in their L2 texts which can be used as traces in NLI to predict their L1. In other words, NLI task makes use of this assumption by examining native language transfer effects or what is called the cross-linguistic influence (CLI) which shows that non-native speakers produce certain linguistic patterns in L2 texts as a result of their mother tongue. This influence is transferred to the author's second language writings and can be shown in the syntactic and orthographical errors as well as the choice of words the writer uses in L2 texts. These linguistic patterns can be used as features in NLI classification task. NLI research can be utilized in various fields such as to develop materials for second language learning, in forensic linguistics and in author profiling.Various types of linguistic features have been explored for English NLI using either learners' corpora or TOEFL data. Some of these features are function words, character ngrams, part of speech (POS) tags, syntactic structure, error-based features, style-based features and so on. With the emergence of learner corpora in other languages such as Chinese and Arabic, NLI research has been extended to these languages. The release of the first Arabic learner corpus by Alfaifi and Atwell (2013) provides a basis for work on Arabic NLI research. However, when it comes to investigating features, limited types of features have been investigated for Arabic NLI; all of them are syntactic based.As a supervised multi-class classification task, I base my system on a traditional machine learning technique of linear Support Vector Machine classifier (linear SVM). The contributions of this thesis focus on two realms of Arabic NLI. First, I explore the impact of using different tokenization schemes and POS tag sets on Arabic NLI performance. For the tokenization subtask, I use 5 different tokenization schemes provided by two Arabic tokenizers (Farasa vs Madamira) on the Arabic NLI predictions to determine which tokenization scheme is more suitable for Arabic NLI. For the POS tag sets subtask, I use two Arabic POS tag sets: PATB tag set (Maamouri et al., 2004) which makes use of a fine-grained scheme (21 POS tags), and CATIB tag set (Habash and Roth, 2009) which uses a very coarse-grained POS tag set (6 POS tags). Investigating these types of POS tag sets answers the question of which size of POS tag sets produces better results for Arabic NLI task. Findings of the tokenization subtask show that Madamira tokenizer (55.26%) outperforms Farasa tokenizer (52.37%) when feature selection is used. The best Madamira scheme for Arabic NLI is the one that provides the minimal tokenization, which only tokenizes question and conjunction proclitics (55.26%) when feature selection is employed. When investigating the different POS tag sets, results show that the best POS tag set for Arabic NLI predictions is CATIB (19.74%) as compared to PATB (18.68%). Second, I investigate different feature types for Arabic NLI predictions. I use a wide range of lexicalized features, unlexicalized features and a combination of lexicalized and unlexicalized features to know which ones are more accurate and informative when used for a morphologically rich language like Arabic. Findings show that the best lexicalized single feature is character ngrams that range from 2-6 (61.84%). The best unlexicalized single feature set is prefix-suffix when D2 1-4 is used (53.95%). The best combination is when I combine character ngrams ranging from 2 to 7 with function words (60.00%). Unlexicalized features still show solid results (53.95%) and thus can be useful when there is topic shift.Since my data can be imbalanced for topic and since I employ lexicalized features in my thesis, I consider investigating topic bias issue in my data. Results show that topic bias issue does not exist in my data set. I also introduce a comparison between my system and three previous Arabic NLI systems, one of which was the state-of-the-art system. Results show that my SVM-based system outperforms the other systems; it outperforms the state of the art system by more than 10 percent points.
일반주제명  
Language
일반주제명  
Middle Eastern studies
키워드  
Arabic learner
키워드  
Native Language Identification
키워드  
Cross-linguistic influence
키워드  
Arabic tokenizers
기타저자  
Indiana University Linguistics
기본자료저록  
Dissertations Abstracts International. 86-02A.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008250123s2024        us                              c    eng  d
■001000017163900
■00520250211152807
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798384027027
■035    ▼a(MiAaPQ)AAI31557219
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a400
■1001  ▼aBassas,  Yasmeen.
■24510▼aArabic  Native  Language  Identification
■260    ▼a[Sl]▼bIndiana  University▼c2024
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2024
■300    ▼a177  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  86-02,  Section:  A.
■500    ▼aAdvisor:  Kuebler,  Sandra.
■5021  ▼aThesis  (Ph.D.)--Indiana  University,  2024.
■520    ▼aNative  Language  Identification  (NLI)  is  concerned  with  predicting  the  native  language  of  an  author's  writing  in  a  second  language.  NLI  relies  on  the  assumption  that  speakers  of  the  same  native  language  display  certain  linguistic  patterns  in  their  L2  texts  which  can  be  used  as  traces  in  NLI  to  predict  their  L1.  In  other  words,  NLI  task  makes  use  of  this  assumption  by  examining  native  language  transfer  effects  or  what  is  called  the  cross-linguistic  influence  (CLI)  which  shows  that  non-native  speakers  produce  certain  linguistic  patterns  in  L2  texts  as  a  result  of  their  mother  tongue.  This  influence  is  transferred  to  the  author's  second  language  writings  and  can  be  shown  in  the  syntactic  and  orthographical  errors  as  well  as  the  choice  of  words  the  writer  uses  in  L2  texts.  These  linguistic  patterns  can  be  used  as  features  in  NLI  classification  task.  NLI  research  can  be  utilized  in  various  fields  such  as  to  develop  materials  for  second  language  learning,  in  forensic  linguistics  and  in  author  profiling.Various  types  of  linguistic  features  have  been  explored  for  English  NLI  using  either  learners'  corpora  or  TOEFL  data.  Some  of  these  features  are  function  words,  character  ngrams,  part  of  speech  (POS)  tags,  syntactic  structure,  error-based  features,  style-based  features  and  so  on.  With  the  emergence  of  learner  corpora  in  other  languages  such  as  Chinese  and  Arabic,  NLI  research  has  been  extended  to  these  languages.  The  release  of  the  first  Arabic  learner  corpus  by  Alfaifi  and  Atwell  (2013)  provides  a  basis  for  work  on  Arabic  NLI  research.  However,  when  it  comes  to  investigating  features,  limited  types  of  features  have  been  investigated  for  Arabic  NLI;  all  of  them  are  syntactic  based.As  a  supervised  multi-class  classification  task,  I  base  my  system  on  a  traditional  machine  learning  technique  of  linear  Support  Vector  Machine  classifier  (linear  SVM).  The  contributions  of  this  thesis  focus  on  two  realms  of  Arabic  NLI.  First,  I  explore  the  impact  of  using  different  tokenization  schemes  and  POS  tag  sets  on  Arabic  NLI  performance.  For  the  tokenization  subtask,  I  use  5  different  tokenization  schemes  provided  by  two  Arabic  tokenizers  (Farasa  vs  Madamira)  on  the  Arabic  NLI  predictions  to  determine  which  tokenization  scheme  is  more  suitable  for  Arabic  NLI.  For  the  POS  tag  sets  subtask,  I  use  two  Arabic  POS  tag  sets:  PATB  tag  set  (Maamouri  et  al.,  2004)  which  makes  use  of  a  fine-grained  scheme  (21  POS  tags),  and  CATIB  tag  set  (Habash  and  Roth,  2009)  which  uses  a  very  coarse-grained  POS  tag  set  (6  POS  tags).  Investigating  these  types  of  POS  tag  sets  answers  the  question  of  which  size  of  POS  tag  sets  produces  better  results  for  Arabic  NLI  task.  Findings  of  the  tokenization  subtask  show  that  Madamira  tokenizer  (55.26%)  outperforms  Farasa  tokenizer  (52.37%)  when  feature  selection  is  used.  The  best  Madamira  scheme  for  Arabic  NLI  is  the  one  that  provides  the  minimal  tokenization,  which  only  tokenizes  question  and  conjunction  proclitics  (55.26%)  when  feature  selection  is  employed.  When  investigating  the  different  POS  tag  sets,  results  show  that  the  best  POS  tag  set  for  Arabic  NLI  predictions  is  CATIB  (19.74%)  as  compared  to  PATB  (18.68%).  Second,  I  investigate  different  feature  types  for  Arabic  NLI  predictions.  I  use  a  wide  range  of  lexicalized  features,  unlexicalized  features  and  a  combination  of  lexicalized  and  unlexicalized  features  to  know  which  ones  are  more  accurate  and  informative  when  used  for  a  morphologically  rich  language  like  Arabic.  Findings  show  that  the  best  lexicalized  single  feature  is  character  ngrams  that  range  from  2-6  (61.84%).  The  best  unlexicalized  single  feature  set  is  prefix-suffix  when  D2  1-4  is  used  (53.95%).  The  best  combination  is  when  I  combine  character  ngrams  ranging  from  2  to  7  with  function  words  (60.00%).  Unlexicalized  features  still  show  solid  results  (53.95%)  and  thus  can  be  useful  when  there  is  topic  shift.Since  my  data  can  be  imbalanced  for  topic  and  since  I  employ  lexicalized  features  in  my  thesis,  I  consider  investigating  topic  bias  issue  in  my  data.  Results  show  that  topic  bias  issue  does  not  exist  in  my  data  set.  I  also  introduce  a  comparison  between  my  system  and  three  previous  Arabic  NLI  systems,  one  of  which  was  the  state-of-the-art  system.  Results  show  that  my  SVM-based  system  outperforms  the  other  systems;  it  outperforms  the  state  of  the  art  system  by  more  than  10  percent  points.
■590    ▼aSchool  code:  0093.
■650  4▼aLanguage
■650  4▼aMiddle  Eastern  studies
■653    ▼aArabic  learner
■653    ▼aNative  Language  Identification
■653    ▼aCross-linguistic  influence
■653    ▼aArabic  tokenizers
■690    ▼a0555
■690    ▼a0679
■71020▼aIndiana  University▼bLinguistics.
■7730  ▼tDissertations  Abstracts  International▼g86-02A.
■790    ▼a0093
■791    ▼aPh.D.
■792    ▼a2024
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163900▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

Preview

Export

ChatGPT Discussion

AI Recommended Related Books


    New Books MORE
    Statistics for the past 3 years. Go to brief

    detalle info

    • Reserva
    • No existe
    • Mi carpeta
    • Primera solicitud
    • Non-Book Loan Application
    • Nighttime Book Loan Application
    Material
    número de libro número de llamada Ubicación estado Prestar info
    TF11460 전자도서 대출가능 My Folder 부재도서신고 비도서대출신청 야간 도서대출신청

    * Las reservas están disponibles en el libro de préstamos. Para hacer reservaciones, haga clic en el botón de reserva

    Books borrowed together with this book

    Related Popular Books

    Available after logging in.