서브메뉴
검색
Prompt and Circumstance: Investigating the Relationship Between College Writing and Postsecondary Policy
Prompt and Circumstance: Investigating the Relationship Between College Writing and Postsecondary Policy
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152956
- ISBN
- 9798384043393
- DDC
- 401
- 서명/저자
- Prompt and Circumstance: Investigating the Relationship Between College Writing and Postsecondary Policy
- 발행사항
- [Sl] : University of Michigan, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 279 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
- 주기사항
- Advisor: Aull, Laura.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2024.
- 초록/해제
- 요약In US-based postsecondary education, first-year students commonly have their compositional ability consequentially assessed on the basis of standardized tests. As a result, students who score above certain thresholds on ACT, SAT, or AP exams often are placed into honors or remedial courses; receive credit remissions; and/or test out of general education classes such as first-year composition. While the thresholds and applicable tests vary from institution to institution, over 2000 Title IV schools implement policies based on such tests. However, there is little evidence that the linguistic patterns that correlate with success on timed, high-stakes tests carry forward to college-level writing tasks. Consequently, contemporary composition scholars call for research that centers examinations of student writing itself rather than assessments of writing quality such as standardized tests. This dissertation responds to that call by answering the questions, How do linguistic features observed in college-level writing relate to institutionally sanctioned measures of writing quality? And, what are the implications for policy levers based on those measures?To answer these questions, I leverage a longitudinal corpus (2009-2019) of approximately 47,000 student essays, matched with data on test scores. Together, these data allow me to investigate whether the test scores, implemented as boolean policy levers, meaningfully distinguish between students who write using measurably distinct linguistic patterns. To measure such distinctions, this study employs natural language processing by incorporating large language models designed for text classification tasks: BERT, RoBERTa, and XLNet. The methods employed in this study identify a quadratic weighted kappa of 0.43, which indicates that the model was able to classify student essays better than random assignment; however, the relationship between student writing and test scores maintain a minimal relationship. Ideally, educational policy that consequentially sorts students into different educational tracks at the most vulnerable point of their college career would bear more than a weak relationship to their college-level performance.To uncover which linguistic features are most correlated with higher scores, I employ OLS, multiple, and logistic regression. These models find significant differences between the essays of students with high and low test scores. Across most models, students with higher test scores have on average fewer clauses per sentence; more prepositions, adverbs, colons, and adjectives; and write with the same number of personal pronouns. While these findings are statistically significant, they only weakly describe the differences between high- and low-scoring, such that distinguishing between essays of students who are near common policy thresholds would be an error-prone task for any human or algorithm. Additionally, while the logistic regression based on the existing policy threshold at University of Michigan had the greatest explanatory power (Pseudo R2 0.09), linear regressions based on a normalized ACT-SAT score had more explanatory power (R2 0.161). While these metrics cannot be directly compared, the difference in their relative strength nonetheless reveals a disparity in goodness-of-fit that demonstrates how educational policy based on a boolean threshold from one test is functionally less discriminating than the metric that is based on multiple measures. Significance notwithstanding, the overall weak correlation between standardized test scores and college-level writing evidences the inability for a timed, high-stakes writing test to relate to writing in other circumstances, including college-level writing tasks. These results evidence the brittleness of these test scores as measures of writing quality and cast doubt as to their utility as policy levers.
- 일반주제명
- Linguistics
- 일반주제명
- Education policy
- 일반주제명
- Computer science
- 일반주제명
- Educational evaluation
- 키워드
- College writing
- 기타저자
- University of Michigan English & Education
- 기본자료저록
- Dissertations Abstracts International. 86-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164387
■00520250211152956
■006m o d
■007cr#unu||||||||
■020 ▼a9798384043393
■035 ▼a(MiAaPQ)AAI31631201
■035 ▼a(MiAaPQ)umichrackham005695
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a401
■1001 ▼aGodfrey, Jason Michael.
■24510▼aPrompt and Circumstance: Investigating the Relationship Between College Writing and Postsecondary Policy
■260 ▼a[Sl]▼bUniversity of Michigan▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a279 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: B.
■500 ▼aAdvisor: Aull, Laura.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2024.
■520 ▼aIn US-based postsecondary education, first-year students commonly have their compositional ability consequentially assessed on the basis of standardized tests. As a result, students who score above certain thresholds on ACT, SAT, or AP exams often are placed into honors or remedial courses; receive credit remissions; and/or test out of general education classes such as first-year composition. While the thresholds and applicable tests vary from institution to institution, over 2000 Title IV schools implement policies based on such tests. However, there is little evidence that the linguistic patterns that correlate with success on timed, high-stakes tests carry forward to college-level writing tasks. Consequently, contemporary composition scholars call for research that centers examinations of student writing itself rather than assessments of writing quality such as standardized tests. This dissertation responds to that call by answering the questions, How do linguistic features observed in college-level writing relate to institutionally sanctioned measures of writing quality? And, what are the implications for policy levers based on those measures?To answer these questions, I leverage a longitudinal corpus (2009-2019) of approximately 47,000 student essays, matched with data on test scores. Together, these data allow me to investigate whether the test scores, implemented as boolean policy levers, meaningfully distinguish between students who write using measurably distinct linguistic patterns. To measure such distinctions, this study employs natural language processing by incorporating large language models designed for text classification tasks: BERT, RoBERTa, and XLNet. The methods employed in this study identify a quadratic weighted kappa of 0.43, which indicates that the model was able to classify student essays better than random assignment; however, the relationship between student writing and test scores maintain a minimal relationship. Ideally, educational policy that consequentially sorts students into different educational tracks at the most vulnerable point of their college career would bear more than a weak relationship to their college-level performance.To uncover which linguistic features are most correlated with higher scores, I employ OLS, multiple, and logistic regression. These models find significant differences between the essays of students with high and low test scores. Across most models, students with higher test scores have on average fewer clauses per sentence; more prepositions, adverbs, colons, and adjectives; and write with the same number of personal pronouns. While these findings are statistically significant, they only weakly describe the differences between high- and low-scoring, such that distinguishing between essays of students who are near common policy thresholds would be an error-prone task for any human or algorithm. Additionally, while the logistic regression based on the existing policy threshold at University of Michigan had the greatest explanatory power (Pseudo R2 0.09), linear regressions based on a normalized ACT-SAT score had more explanatory power (R2 0.161). While these metrics cannot be directly compared, the difference in their relative strength nonetheless reveals a disparity in goodness-of-fit that demonstrates how educational policy based on a boolean threshold from one test is functionally less discriminating than the metric that is based on multiple measures. Significance notwithstanding, the overall weak correlation between standardized test scores and college-level writing evidences the inability for a timed, high-stakes writing test to relate to writing in other circumstances, including college-level writing tasks. These results evidence the brittleness of these test scores as measures of writing quality and cast doubt as to their utility as policy levers.
■590 ▼aSchool code: 0127.
■650 4▼aLinguistics
■650 4▼aEducation policy
■650 4▼aComputer science
■650 4▼aEducational evaluation
■653 ▼aCollege writing
■653 ▼aStandardized tests
■653 ▼aAdvanced Placement
■653 ▼aScholastic Aptitude Test
■653 ▼aNatural language processing
■690 ▼a0458
■690 ▼a0290
■690 ▼a0984
■690 ▼a0443
■71020▼aUniversity of Michigan▼bEnglish & Education.
■7730 ▼tDissertations Abstracts International▼g86-03B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164387▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


