Download presentation
Presentation is loading. Please wait.
Published byMabel Cobb Modified over 8 years ago
1
Morphosemantic Relations in and Across Wordnets: A Study Based on Turkish Orhan Bilgin, Özlem Çetinoglu, Kemal Oflazer Sabanci University Human Languages and Speech Technologies Laboratory Istanbul, Turkey {orhanb,ozlemc,oflazer@sabanciuniv.edu}
2
OBJECTIVES Using morphological processes in Language A, we can: extract explicit semantic relations in Language A and use these to enrich Wordnet A; automatically prepare machine-tractable synset glosses for Wordnet A and/or B; and most importantly discover implicit semantic relations in Wordnet B and use these to enrich Wordnet B.
3
METHODOLOGY 1) Determine derivational affixes in Language A Rule of Thumb: Prefer productive affixes with predictable semantics
4
METHODOLOGY -ci The chaotic “agentive” suffix -ci
5
A more well-behaved suffix: - li -li METHODOLOGY
6
2) Define the semantic effects of the affixes SUFFIXPOSEFFECT -laş n-v, a-v BECOME -lann-vACQUIRE -lHk a-n, n-n BE_IN_STATE -lHn-aWITH -sHzn-aWITHOUT -sAln-aPERTAINS_TO -(y)lAn-bWITH -Hşv-vRECIPROCAL -(H)lv-vCAUSES -(H)t, -DHr, -(H)r, -(A)r v-vIS_CAUSED_BY -Hşv-nACT_OF -CA a-b, n-b MANNER
7
2) Extract morphosemantically-related pairs boulder builder deer dresser father founder her killer laser maker mother never teacher CANDIDATES boulder builder deer dresser father founder her killer laser maker mother never teacher WINNERS MORPHOLOGICAL ANALYZER ROOT (v) + AGENT ? METHODOLOGY
8
2) Extract morphosemantically-related pairs build - builder dress - dresser found - founder kill - killer make - maker teach - teacher METHODOLOGY
9
-put on clothes -dress in a certain manner -dress with elaborate care -put a dressing on -convert into leather -apply a bandage to -give a neat appearance to -arrange hair attractively -put a finish on -kill and prepare for consumption -arrange in ranks -provide with clothes -cut back the growth of -furniture for keeping clothes -person who dresses in a particular way -a wardrobe assistant for an actor -a cabinet with shelves -low table with mirror or mirrors 3) Link pair members to ILI records dress (v)dresser (n) (1709-1784) METHODOLOGY ?
10
USES 1) Extract explicit semantic relations in the language taşlaşmak polimerleşmek iyonlaşmak kemikleşmek billurlaşmak kireçleşmek plastikleşmek izomerleşmek keratinleşmek BECOME taşlaştırmak polimerleştirmek iyonlaştırmak kemikleştirmek billurlaştırmak kireçleştirmek plastikleştirmek izomerleştirmek keratinleştirmek taş polimer iyon kemik billur kireç plastik izomer keratin IS_CAUSED_BY
11
USES 2) Share relations with other wordnets a) Pairs in importing language are morphologically related delidelilik madnessmad 13580347-n 02005975-a STATE_OF EXPORTING LANGUAGE IMPORTING LANGUAGE INTERLINGUAL INDEX
12
USES 2) Share relations with other wordnets a) Pairs in importing language are morphologically related delidelilik madnessmad 13580347-n 02005975-a STATE_OF EXPORTING LANGUAGE IMPORTING LANGUAGE INTERLINGUAL INDEX STATE_OF
13
USES 2) Share relations with other wordnets b) Pairs in importing language are morphologically unrelated yıkmakyıkılmak collapsetear down 01614562-v 01931110-v CAUSES EXPORTING LANGUAGE IMPORTING LANGUAGE INTERLINGUAL INDEX
14
USES yıkmakyıkılmak collapsetear down 01614562-v 01931110-v CAUSES EXPORTING LANGUAGE IMPORTING LANGUAGE INTERLINGUAL INDEX CAUSES 2) Share relations with other wordnets b) Pairs in importing language are morphologically unrelated
15
USES 3) Prepare simple synset glosses omurgaomurgalı vertebratespine 05268544-n 02422440-a WITH EXPORTING LANGUAGE IMPORTING LANGUAGE INTERLINGUAL INDEX
16
USES 3) Prepare simple synset glosses omurgaomurgalı vertebratespine 05268544-n 02422440-a WITH EXPORTING LANGUAGE IMPORTING LANGUAGE INTERLINGUAL INDEX vertebrate == with spine
17
USES 3) Prepare simple synset glosses Some examples based on Turkish Wordnet: ossify: become bonedress: cause to wear languish: become weakdissuade: cause to give up petrify: become stoneabrade: cause to wear away thin out: become sparseencourage: cause to take heart improve: become goodkill: cause to die saponify: become soapdisease: state of being sick caseate: become cheeseinfidel: without religion hush: become silentweak: without strength rejuvenate: become youngperfect: without defect calcify: become limesmooth: without roughness
18
RESULTS SUFFIX # OF PAIRS EFFECT -lik4,078BE_IN_STATE -li2,725WITH -siz1,001WITHOUT -iş991ACT_OF -lan758ACQUIRE -laş763BECOME -dir782CAUSES -ca710MANNER -sal115PERTAINS_TO TOTAL11,923 The current wordlist of Turkish contains a substantial number of words derived from a small set of suffixes.
19
RESULTS SUFFIX # IN WL # IN TWN # IN PWN NEW -dir782801877.5% -laş763831186.7% Detailed analysis of two suffixes: Although Turkish Wordnet is a small-sized resource (~10,000 synsets), it contains a significant number of synsets involving these two suffixes. In only a few cases does PWN indicate a CAUSES relation between the respective English synsets. In the case of the BECOME pairs, PWN provides the underspecified relation “ENG_DERIVATIVE”.
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.