Automatic Analysis of Epistemic Stance-Taking in Academic English Writing: A Systemic Functional Approach
Loading...
Date
2024-01-10
Authors
Eguchi, Masaki
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oregon
Abstract
Existing linguistic textual measures that investigate features of academic writing often focus on lexis, syntax, and cohesion, despite writing skills being considered more complex and multifaceted (e.g., Sparks et al., 2014). For this reason, writing assessment researchers seek ways to measure and assess various textual features beyond the traditional ones, including discourse moves and steps (Cotos, 2014), source use (Burstein et al., 2018; Kyle, 2020), and essay argument structures (Fiacco et al., 2022). The present dissertation attempts to extend this research by proposing an automated analysis of rhetorical discourse features of epistemic stance-taking strategies.
Drawing on a theoretical framework of the engagement system from Appraisal Analysis (Martin & White, 2005), which originates from the Sydney School of the systemic functional discourse analysis tradition, the dissertation develops and evaluates a series of end-to-end machine learning models to conduct automated engagement resource analysis. The experiment in Study 1 indicated that the developed system can perform as well as (or even outperform) trained annotators’ intercoder agreement. Study 2 uses the natural language processing (NLP) systems to conduct the first large-scale analysis of engagement resources in university written assignments across genres and disciplines. The findings suggested that the registers of university writings are far more complex and nuanced than simple characterization by genres or disciplines.
Study 3 investigates whether the developed measures of rhetorical features of engagement can provide additional information above and beyond the traditional linguistic measures at the levels of lexis, syntax, and cohesion, for modeling professional ratings of essay qualities in a standardized second language proficiency assessment. The results indicate that the features of engagement (particularly the diversity of rhetorical strategies) can complement the existing measures in predicting essay quality.
The three studies together indicate that the proposed machine-learning approach is beneficial to scale up the analysis of rhetorical discourse features in academic writing for research and educational purposes. The dissertation concludes with a call for increased collaboration among discourse analysts, second language researchers, assessment researchers, and computational linguists to define essential textual features for writing assessments across contexts and automate the analysis of such constructs (Lu, 2021, Burstein et al., 2016).
Description
Keywords
academic writing, bayesian statistics, epistemic stance, evaluative language, language assessment, natural language processing