Developing Novel Machine Learning Techniques to Improve Comparative Judgements for e-Learning and e-Assessment

在这个项目中，我们可以研究的目标有很多：1)了解机器学习（ML）如何作为教育工作者的支持而不是紧身衣。关键问题：不同的措施适用于不同的教师吗？他们如何解释数据，并用它来指导他们的教学？2)了解我们如何收集大量关于定性评估的可量化数据，从而允许机器学习（ML）方法对数据进行处理。3)为了在这个复杂的设计过程中与学生和老师互动，适合使用哪些设计方法。关键问题是：我们如何表达从数据中收集到的不同见解的可能性？我们如何抓住ML中更模糊的元素（如隐私）的想法，并确保数据收集的持续同意？4)确定适当的机器学习方法，并开发一个完整的机器学习框架，用于分析定性评估。5)积极地查询所有提交的一小部分，以了解哪些属性构成了优秀的质量，采取与教育工作者互动的方法来改进我们提出的框架的性能。6)验证并使开发的工具在现实世界中可用。我们的目标是纠正由自动化评分工具造成的不平衡，促进具体的评估方法。我们将开发一种决策支持工具，用于教育工作者评估定性作业，通过增加教育工作者对学生群体作业的了解，并可能减少他们需要花在批改作业上的时间，使此类评估更具吸引力。将制定一个比较判断框架bbb10，使教育工作者能够了解不同群组之间的实践变化如何影响评估。如果比较判断框架能够在队列水平上显示可靠的表现，我们将探索如何将其应用于个人定性作业，以支持教育者对他们的评估。第一年-识别评估：定性评估是主观的，学习好的和坏的做法是自主的挑战。为了解决这个问题，学生将调查教学从业者，以确定最重要的定性评估类型，以及这些评估中的关键学习成果。我们还将与教育工作者和学生合作，了解他们对工作自动化评估的态度，他们对工具的担忧，以及他们可以确保遵循价值敏感设计方法b[2]的评估有效性的方式。这将缩小工具的范围，并用于创建概念验证。一年级和二年级：学习知识和发现知识差距：CDSM有丰富的提交和各自成绩的数据。在这个阶段，我们的目标是从这些数据中学习，以确定在前一步确定的范围内的关键特征。在这里，我们将首先专注于从语义角度寻找数据集中有意义的结构[3]，可能利用监督方法（使用现有等级）和某种形式的上下文嵌入[4]。这将有助于深入了解是什么使提交内容解决特定的学习结果。假设我们使用概率模型，我们可以推断出我们的模型对哪些提交具有低置信度（或预测中的高不确定性）：这显示了我们知识中的空白，只有人类从业者才能帮助填补。第三年-交互式改进模型：开发一个交互式视觉系统，以显示对队列数据和个人提交的见解，并允许教育从业者回答关于提交的两个关键问题：我们的模型预测正确吗？他们会如何评价一个给定的任务？这些提交将通过主动学习策略进行仔细挑选，我们选择我们不确定的提交或可能改进我们的模型以提供更好的预测的提交。

There is a wide selection of aims we could examine under this project:1) To understand how machine learning (ML) can function as a support for an educator and not a strait jacket. Key questions to ask:Are different measures relevant to different teachers? How do they interpret the data and use it to guide their teaching?2) To understand how we can gather large volumes of quantifiable data about qualitative assessments allowing machine learning (ML) approaches to work on the data.3) What design methods are suitable to use in order to engage with students and teachers in this complex design process. Key questions are:How do we express possibilities for different insights that we can gather from the data?How do we capture ideas about more nebulous elements of ML like privacy and ensure ongoing consent for data collection?4) To identify appropriate ML methods and develop a complete ML framework for analysing qualitative assessments.5) To actively query a small subset of all submissions to learn what attributes constitute excellent quality, taking an interactive approach with the educators to improve our proposedframework's performance.6) To validate and make the developed tool useable in the real-world.We aim to redress the imbalance caused by automated marking tools promoting specific approaches to assessment. We will develop a decision support tool for educators assessing qualitative work that will make such assessments more attractive by increasing the educators understanding of the student cohorts' work and, potentially, reducing the amount of time they need to spend marking it. A comparative judgement framework [1] will be developed to allow educators to understand how changes in practice between different cohorts' impact on assessments. If the comparative judgement framework can show reliable performance at the cohort level, we will explore how it can be applied to individual qualitative assignments to support educators' assessments of them. Year One - Identifying assessments: Qualitative assessments are subjective and learning about good and bad practices autonomously challenging. To address this, the student will survey teaching practitioners to identify the most important types of qualitative assessments, and key learning outcomes in those assessments. We will also work with educators and students to understand their attitudes to automated assessment of their work, their concerns about the tool and the ways in which they can be reassured of the validity of such assessments following Value Sensitive Design approaches [2]. This will narrow down thescope of the tool and be used to create a proof-of-concept.Year One and Two - Learning knowns and identifying gaps in knowledge: CDSM has awealth of data on submissions and their respective grades. At this stage, we aim to learn from this data to identify key features within the scope identified in the previous step. Here, we will first focus on finding meaningful structures in the dataset from a semantic perspective [3], potentially utilising a supervised approach (using existing grades) and some form of contextual embeddings [4]. These will help derive insight into what makes a submission address specificlearning outcomes. Given we use a probabilistic model, we can deduce which submissions our model has low confidence (or high uncertainty in predictions) about: this shows the gaps in our knowledge, which only a human practitioner can help fill.Year Three - Interactively improving models: develop an interactive visual system to display insights into cohort data and individual submissions and allow educations practitioners to respond to two key questions about a submission:Were our model predictions correct? How would they rate a given assignment?These submissions will be carefully picked via an active learning strategy where we select submission that we are unsure about or that are likely to improve our model to provide better predictions.