What question did this study set out to answer?

The research aims to develop an automated framework for labeling and analyzing syntactic variations in English across different dialects.

March 28, 2026International Journal of Information and Communication Technology0 citationsOpen Access

Deep learning-based automatic labelling of English syntactic variation and cross-dialect comparison

YJYanli Jia XYXinhua YuanBeijing Foreign Studies University

Key Points

The research aims to develop an automated framework for labeling and analyzing syntactic variations in English across different dialects.
Developed an end-to-end deep learning framework for automated labeling of syntactic variants.
Integrated parser-generated annotations with a human-reviewed subset for accuracy.
Employed a dual-head neural model using CRF-based sequence tagging and span classification.
Utilized techniques like gradient reversal and supervised contrastive learning for domain adaptation.
Tested the model on diverse English dialect corpora including American, British, Australian, and Indian English.
Achieved a 6.9-point increase in out-of-dialect macro-F1 score compared to a strong RoBERTa baseline.
Reduced domain divergence in encoder space by over 55%.
Successfully identified stable contrasts for grammatical variants such as that-complementiser drop and dative alternation.

Abstract

With the growing demand for precise cross-dialect syntactic analysis, this work proposes an end-to-end framework that automatically labels English syntactic variants and quantifies their distribution across dialects.The approach integrates parser-generated silver annotations, a human-audited gold subset, and a dual-head neural model combining CRF-based sequence tagging and span classification.Domain adaptation with gradient reversal, moment matching, and supervised contrastive learning enhances robustness to dialectal shift, while probability calibration ensures accurate rate estimation.Evaluations on multi-source corpora covering American, British, Australian, and Indian English show that the proposed model improves out-of-dialect macro-F1 by 6.9 points over a strong RoBERTa baseline, reduces domain divergence in encoder space by over 55%, and recovers stable, interpretable contrasts for variants such as that-complementiser drop, particle movement, and dative alternation.

AIに質問

Bookmark

View Full Paper