Argumentative Writing Support System for EFL Learners

本研究旨在开发一种支持工具，用于分析和改进英语学习者的议论文写作。该工具提取并可视化了一篇文章的议论文结构，然后通过重新排列句子来建议一篇文章的改进版本。该项目分为几个步骤：(A)数据集构建。2020年，我构建了第一个语料库(ICNALE-AS2R)，语料库中有434篇英语学习者的随笔，这些语料库已经标注了议论文结构和句子的重新排序。我还为这一步开发了一个内部注释软件(开源)tiara。(B)结构解析模型。在给定一篇输入文章的情况下，我开发了一个深度学习系统来预测文章的论证结构。我使用了双仿射注意力和伯特模型的组合。提出了多任务、多语料库的训练策略，提高了基本模型的句法分析性能。我们的想法是指导模型如何减少搜索空间，并通过选择性采样利用现有的非EFL数据集。(C)句子重新排序模型。以一篇文章及其相应的论辩结构为输入，我提出了一个基于连贯语言学理论的重新排序系统。该任务被描述为一个遍历问题，包含两个步骤。第一个是句子对之间的成对排序约束任务(使用Albert模型解决)。第二个步骤是使用自组织算法的遍历(输出生成)步骤。详细的实证评估表明，本文提出的系统具有很好的应用潜力。

My research aims to develop a support tool for analysis and improving argumentative essays written by English-as-foreign-language (EFL) learners. The tool extracts and visualizes the argumentative structure of a given essay, and then suggests an improved version of the essay by reordering sentences. This project was divided into several steps:(a) Dataset Construction. In 2020, I constructed the first corpus of 434 essays written by EFL learners (ICNALE-AS2R), which have been annotated with argumentative structure and sentence reordering. I also developed an in-house annotation software (open-source) TIARA for this step.(b) Structure Parsing Model. Given an input essay, I developed a deep learning system to predict the argumentative structure of the essay. I used the combination of Biaffine Attention and BERT models. I also proposed multi-task and multi-corpora training strategies which enhanced the parsing performance of the base model. The idea was to guide the model on how to reduce the search space and to utilise existing non-EFL datasets via selective sampling.(c) Sentence Reordering Model. Given an essay and its corresponding argumentative structure as input, I proposed a reordering system based on the linguistic theory of coherence. The task was formulated as a traversal problem, containing two steps. The first is a pairwise ordering constraint task between pairs of sentences (tackled using ALBERT model). The second is a traversal (output generation) step using an ad-hoc algorithm. Detailed empirical evaluation showed that my proposed system has a good potential.