Files
Abstract
This dissertation delves into the phenomenon of pro-drop, where pronouns can be grammatically omitted in certain languages. The research focuses on three key questions: (1) How does discourse coherence support pro-drop languages, and how can we quantify discourse coherence? (2) What is the order of importance among linguistic factors relevant to pro-drop? (3) If an AI model is trained to predict pro-drop based on language dependency structures, what factors are crucial cross-linguistically? These questions are addressed in three experiments.
Experiment 1 quantifies verb usage continuity to compare character salience in pro-drop and non-pro-drop cases across CN, BP, and ES. Results indicate that pro-drop cases necessitate higher character salience in verb usage continuity, highlighting the role of discourse coherence. Differences exist among languages, with radical pro-drop languages demanding more coherence than partial pro-drop and consistent pro-drop languages.
Experiment 2 employs Binomial Logistic Regression and Random Forests to model pro-drop based on syntactic, semantic, morphological, and logical features. Findings reveal the significance of character consistency between main-embedded and current-previous clauses in all languages (Chinese, Brazilian Portuguese, and Spanish). ES and BP exhibit richer verbal morphology, potentially contributing to a higher model fit than Chinese. Sentential discourse relation features, such as coordinate structures, encourage pro-drop in CN and ES.
Experiment 3 employs Graph Attention Networks (GATs) to classify story characters using dependency structure graphs from CN, BP, and ES discourse material. Results align with key elements in pronoun resolution, highlighting the importance of subjects, main verbs, and objects. Language-specific importance emerges, with the marker (e.g. '就'(jiu4, means "just")) and auxiliary (e.g. '了'(le0, means "already"), '着'(zhe0, means "going on")) crucial in CN, while determiner and case play vital roles in BP and ES.
Experiment 1 quantifies verb usage continuity to compare character salience in pro-drop and non-pro-drop cases across CN, BP, and ES. Results indicate that pro-drop cases necessitate higher character salience in verb usage continuity, highlighting the role of discourse coherence. Differences exist among languages, with radical pro-drop languages demanding more coherence than partial pro-drop and consistent pro-drop languages.
Experiment 2 employs Binomial Logistic Regression and Random Forests to model pro-drop based on syntactic, semantic, morphological, and logical features. Findings reveal the significance of character consistency between main-embedded and current-previous clauses in all languages (Chinese, Brazilian Portuguese, and Spanish). ES and BP exhibit richer verbal morphology, potentially contributing to a higher model fit than Chinese. Sentential discourse relation features, such as coordinate structures, encourage pro-drop in CN and ES.
Experiment 3 employs Graph Attention Networks (GATs) to classify story characters using dependency structure graphs from CN, BP, and ES discourse material. Results align with key elements in pronoun resolution, highlighting the importance of subjects, main verbs, and objects. Language-specific importance emerges, with the marker (e.g. '就'(jiu4, means "just")) and auxiliary (e.g. '了'(le0, means "already"), '着'(zhe0, means "going on")) crucial in CN, while determiner and case play vital roles in BP and ES.