서브메뉴
검색
Artificial Intelligence and Fake Reefs: What Privative Inferences and LLMs Tell Us About Adjective-Noun Composition
Artificial Intelligence and Fake Reefs: What Privative Inferences and LLMs Tell Us About Adjective-Noun Composition
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103546
- ISBN
- 9798280715639
- DDC
- 401
- 저자명
- Ross, Hayley.
- 서명/저자
- Artificial Intelligence and Fake Reefs: What Privative Inferences and LLMs Tell Us About Adjective-Noun Composition
- 발행사항
- [Sl] : Harvard University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 240 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Davidson, Kathryn.
- 학위논문주기
- Thesis (Ph.D.)--Harvard University, 2025.
- 초록/해제
- 요약The fact that people understand completely novel phrases is often taken as an argument that linguistic meaning is composed from the meaning of its parts. Thus, a central concern for the study of meaning is how that meaning is composed, especially for open-class content words like adjectives and nouns. This dissertation studies meaning composition and its interaction with context through the lens of adjective-noun modification and the privative inferences that sometimes result (e.g., a fake gun is (usually) not a gun, and a stone lion is not a (living) lion). This dissertation shows that privativity is not limited to a particular class of adjectives, which leads to a new, non-intersective semantics for adjective-noun composition which handles potential contradictions as part of composition. Further, we find that humans and modern large language models (LLMs) can generalize to the inferences of adjective-noun combinations that they have not seen before. Working with LLMs foregrounds the possibility that these inferences could be drawn by other means than meaning composition, such as memorization or analogy. In fact, success on this task is not explained by analogical generalization, as a computational analogy model and a human experiment involving analogy do not yield the expected inferences for all of the dataset. More broadly, the necessary adaptation in experiment design as well as reflection on our standards of evidence feeds into the broader, currently emerging discussion about how to study compositionality in humans and language models alike.
- 일반주제명
- Linguistics
- 일반주제명
- Language
- 키워드
- Adjectives
- 키워드
- Composition
- 키워드
- Generalization
- 키워드
- Semantics
- 기타저자
- Harvard University Linguistics
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357684
■00520260202103546
■006m o d
■007cr#unu||||||||
■020 ▼a9798280715639
■035 ▼a(MiAaPQ)AAI32041315
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a401
■1001 ▼aRoss, Hayley.▼0(orcid)0000-0003-2288-3847
■24510▼aArtificial Intelligence and Fake Reefs: What Privative Inferences and LLMs Tell Us About Adjective-Noun Composition
■260 ▼a[Sl]▼bHarvard University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a240 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Davidson, Kathryn.
■5021 ▼aThesis (Ph.D.)--Harvard University, 2025.
■520 ▼aThe fact that people understand completely novel phrases is often taken as an argument that linguistic meaning is composed from the meaning of its parts. Thus, a central concern for the study of meaning is how that meaning is composed, especially for open-class content words like adjectives and nouns. This dissertation studies meaning composition and its interaction with context through the lens of adjective-noun modification and the privative inferences that sometimes result (e.g., a fake gun is (usually) not a gun, and a stone lion is not a (living) lion). This dissertation shows that privativity is not limited to a particular class of adjectives, which leads to a new, non-intersective semantics for adjective-noun composition which handles potential contradictions as part of composition. Further, we find that humans and modern large language models (LLMs) can generalize to the inferences of adjective-noun combinations that they have not seen before. Working with LLMs foregrounds the possibility that these inferences could be drawn by other means than meaning composition, such as memorization or analogy. In fact, success on this task is not explained by analogical generalization, as a computational analogy model and a human experiment involving analogy do not yield the expected inferences for all of the dataset. More broadly, the necessary adaptation in experiment design as well as reflection on our standards of evidence feeds into the broader, currently emerging discussion about how to study compositionality in humans and language models alike.
■590 ▼aSchool code: 0084.
■650 4▼aLinguistics
■650 4▼aLanguage
■653 ▼aAdjectives
■653 ▼aComposition
■653 ▼aGeneralization
■653 ▼aLarge language models
■653 ▼aSemantics
■690 ▼a0290
■690 ▼a0679
■690 ▼a0800
■71020▼aHarvard University▼bLinguistics.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a0084
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357684▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
