正在加载页面
REF-2026-013 · PRIMARY SOURCES VERIFIED
导师论文与项目研读路线
不是文献堆积，而是一条从“理解导师”到“讲清方法”再到“守住评价”的阅读路径。
完成首轮 0 篇
正在研读 0 篇
进度只保存在当前浏览器，不写入权威总控台。
CURRENT SUPERVISOR CO-AUTHORED · 导师本人署名
These papers are co-authored by the current prospective supervisor. A 15-minute targeted scan of each is enough for the first meeting; you do not need to read every paper in full.
这3篇均有当前潜在导师本人署名（共同作者）。第一次面谈前，每篇定向看15分钟即可，不需要把所有论文逐字读完。
Huang, Sun, Tani Bertuol, Zhang, Jiang & Jha (2025)
MINIMUM READ · 最少看这些
ONE LINE FOR THE MEETING · 面谈一句话
I was especially interested in your work on LLM-generated misinformation in cyber threat intelligence, because it also connects domain evidence, human judgement and model detection.
我尤其关注您关于大模型生成网络威胁情报错误信息的研究，因为它同样把领域证据、人工判断与模型检测联系起来。
Sun, Ding, Jiang, Xu, Mo, Tai & Zhang (2023)
MINIMUM READ · 最少看这些
ONE LINE FOR THE MEETING · 面谈一句话
Your cyber-threat intelligence survey helped me think of this project as a full pipeline—data, representation, validation and human use—not only a classifier.
您的网络威胁情报综述让我把这个项目理解为数据、表示、验证和人工使用组成的完整管线，而不只是一个分类器。
Yang, Jiang, Pal, Yu, Chen & Yu (2020)
MINIMUM READ · 最少看这些
ONE LINE FOR THE MEETING · 面谈一句话
Your work on COVID-19 myths showed me that harmful information can be studied through both content and propagation. I see propagation as a later extension of my project.
您关于新冠谣言的研究让我看到，有害信息可以同时从内容和传播角度研究；我把传播视为本项目后续可扩展的方向。
她还发表过什么？展开相关论文地图
只列与本项目接近的代表性署名论文，不等于全部成果，也不表示全部由她主导。
最接近你的消息理解、证据使用与可信检测主线。
展示领域证据如何被采集、对齐并转成可复核的情报。
可以作为消息排序实验稳定后的后续扩展，不是当前MPhil核心。
READING ORDER · 阅读顺序
VERIFIED PAPERS · 真实论文
READ FIRST
优先精读
01 · Understand the supervisor's lens / 先理解导师的研究视角
Huang, Sun, Tani Bertuol, Zhang, Jiang & Jha (2025) · Future Generation Computer Systems, 173, 107877
当前导师本人署名（共同作者）
CORE · 核心内容
The paper studies whether LLM-generated cyber-threat intelligence can be distinguished from authentic reports, combining generated samples, human/expert assessment and detector benchmarking.
论文研究能否区分大模型生成的网络威胁情报与真实报告，并结合生成样本、人工/专家判断和检测模型基准测试。
WHY IT MATTERS · 与你课题的关系
It reveals the supervisor cluster's interest in trustworthy AI, domain-specific misinformation and human-versus-model verification—the closest conceptual bridge to your project.
它直接体现了导师研究集群对可信AI、领域型错误信息以及人机核验的兴趣，是与你项目最接近的概念桥梁。
READ FOR · 带着问题读
How are authentic and synthetic CTI samples constructed and validated?
真实与合成CTI样本如何构造和核验？
Which claims depend on this domain and should not be transferred automatically?
哪些结论依赖CTI领域，不能直接搬到交易推广？
DO NOT OVERCLAIM · 不要过度推断
It does not validate a Chinese trading-promotion detector or your proposed weak-supervision comparison.
它不能证明中文交易推广检测有效，也不能替代你计划中的弱监督对比实验。
RL-01 · Publisher record · 2025
READ NEXT
随后研读
01 · Understand the supervisor's lens / 先理解导师的研究视角
Sun, Ding, Jiang, Xu, Mo, Tai & Zhang (2023) · IEEE Communications Surveys & Tutorials, 25(3), 1748–1774
当前导师本人署名（共同作者）
CORE · 核心内容
A survey of how data mining, machine learning, NLP and graph methods turn heterogeneous threat information into actionable cyber-threat intelligence.
这是一篇综述，梳理数据挖掘、机器学习、NLP和图方法如何把异构威胁信息转成可行动的网络威胁情报。
WHY IT MATTERS · 与你课题的关系
It helps you speak the supervisor's pipeline language: collection, representation, mining, validation and operational use—not merely model accuracy.
它帮助你用导师熟悉的“管线语言”交流：采集、表示、挖掘、验证与实际使用，而不只谈模型准确率。
READ FOR · 带着问题读
How does the survey separate data sources, methods and downstream decisions?
综述如何拆分数据来源、方法和下游决策？
Which parts resemble a human-review queue rather than automatic enforcement?
哪些环节更像人工复核队列，而不是自动执法？
DO NOT OVERCLAIM · 不要过度推断
A broad cybersecurity survey provides vocabulary and architecture, not direct evidence for your target population.
这篇广义网络安全综述提供术语和架构，但不是你目标人群的直接证据。
RL-02 · DOI / publisher record · 2023
EXTEND LATER
进阶扩展
01 · Understand the supervisor's lens / 先理解导师的研究视角
Yang, Jiang, Pal, Yu, Chen & Yu (2020) · IEEE Open Journal of the Computer Society, 1, 209–219
当前导师本人署名（共同作者）
CORE · 核心内容
The paper analyses categories, diffusion, engagement and affect around COVID-19 myths on Twitter, linking message content with propagation and human response.
论文分析Twitter上新冠谣言的类别、传播、互动与情绪，把消息内容、扩散方式和人的反应联系起来。
WHY IT MATTERS · 与你课题的关系
It shows how the supervisor cluster studies harmful information as a socio-technical phenomenon, which supports—but does not replace—your stance-aware message analysis.
它展示导师研究集群如何把有害信息视作社会—技术现象，可启发你的立场感知消息分析，但不能取代它。
READ FOR · 带着问题读
Which content and behaviour variables are kept separate?
哪些内容变量与行为变量被分开记录？
What changes when the unit is a tweet rather than a private-channel message?
研究单位从公开推文变成私域频道消息后，哪些假设会改变？
DO NOT OVERCLAIM · 不要过度推断
Public-health myths on Twitter differ materially from Chinese-language trading promotion in platform, intent and governance.
Twitter公共卫生谣言与中文交易推广在平台、意图和治理上有实质差异。
RL-03 · Institutional full text · 2020
READ FIRST
优先精读
02 · Master the core method / 掌握项目核心方法
Ratner et al. (2017) · Proceedings of the VLDB Endowment, 11(3), 269–282
CORE · 核心内容
Snorkel encodes noisy heuristics as labeling functions, estimates their accuracies and dependencies, combines their votes into probabilistic labels, and trains a downstream model.
Snorkel把有噪声的启发式规则写成标注函数，估计其准确性与依赖关系，把投票合成概率标签，再训练下游模型。
WHY IT MATTERS · 与你课题的关系
This is the conceptual foundation for your case-informed rules → weak labels → shared classifier pipeline.
这是你“案例驱动规则 → 弱标签 → 相同分类器”管线的概念基础。
READ FOR · 带着问题读
What exactly is a labeling function, and when can it abstain?
标注函数究竟是什么，什么时候可以弃权？
Why train an end model instead of using rule votes directly?
为什么不直接用规则投票，而要再训练最终模型？
DO NOT OVERCLAIM · 不要过度推断
Snorkel shows a general framework; it does not imply weak supervision will beat human labels in your matched-budget setting.
Snorkel证明的是一般框架；它不代表弱监督一定会在你的同预算实验中胜过人工标签。
RL-04 · Conference full text · 2017
READ FIRST
优先精读
02 · Master the core method / 掌握项目核心方法
Zhang et al. (2021) · NeurIPS 2021 Datasets and Benchmarks Track
CORE · 核心内容
WRENCH standardises weak-supervision evaluation across 22 real-world datasets and many weak-label sources and methods, exposing how results change with data and protocol choices.
WRENCH在22个真实数据集及多种弱标签来源和方法上统一评测，揭示结果会怎样随数据与协议选择而变化。
WHY IT MATTERS · 与你课题的关系
It tells you what must be controlled and reported before claiming that one weak-supervision pipeline is better than another or than a supervised baseline.
它告诉你，在声称某种弱监督优于另一种方法或监督基线前，必须控制并报告哪些因素。
READ FOR · 带着问题读
Which weak-label properties make comparisons unstable?
弱标签的哪些属性会让对比不稳定？
Which reporting fields should enter your frozen protocol?
哪些报告字段应写进你的冻结协议？
DO NOT OVERCLAIM · 不要过度推断
A benchmark cannot choose your ontology, data permissions or harmful-false-alert constraint for you.
基准不能替你决定本体、数据权限或有害误报约束。
RL-05 · Conference record and full text · 2021
READ FIRST
优先精读
02 · Master the core method / 掌握项目核心方法
Hardalov et al. (2022) · Findings of NAACL 2022, 1259–1277
CORE · 核心内容
The survey maps stance detection as both a component of misinformation systems and a standalone task, covering formulations, datasets, methods and open challenges.
该综述把立场检测既视为错误信息系统的组成部分，也视为独立任务，并梳理任务定义、数据集、方法和开放问题。
WHY IT MATTERS · 与你课题的关系
It gives you the literature language for separating a risky claim from the speaker's attitude toward that claim—the central protection against victim and warning false alerts.
它提供文献语言来区分“风险命题”与“说话者对命题的态度”，这是保护受害者叙述和警示信息不被误伤的核心。
READ FOR · 带着问题读
Is stance target-conditioned, message-level, or conversation-level?
立场是针对目标、单条消息，还是整段对话定义？
How should your four-way stance ontology differ from standard labels?
你的四类立场本体应如何区别于标准标签？
DO NOT OVERCLAIM · 不要过度推断
The survey supports the importance of stance, not the novelty or validity of your exact ontology.
综述支持立场的重要性，但不能证明你具体本体的新颖性或有效性。
RL-06 · ACL Anthology · 2022
READ FIRST
优先精读
03 · Defend the evaluation / 把评价设计讲明白
Søgaard et al. (2021) · EACL 2021, 1823–1832
CORE · 核心内容
The paper argues that random train/test splits often give optimistic estimates when deployed data differ from the sampled data, and studies more realistic alternatives.
论文指出，当部署数据与抽样数据不同时，随机训练/测试切分往往给出过于乐观的结果，并研究更贴近现实的替代方案。
WHY IT MATTERS · 与你课题的关系
This is the cleanest theoretical defence for not scattering forwarded templates randomly across training and test rows.
它是你反对把转发模板随机打散到训练集和测试集中的最直接理论依据。
READ FOR · 带着问题读
What deployment population is each split supposed to represent?
每种切分究竟代表哪个部署总体？
Why is a new channel closer to the intended use than a new random row?
为什么新频道比随机新行更接近你的预期使用场景？
DO NOT OVERCLAIM · 不要过度推断
It motivates non-random evaluation but does not uniquely prescribe your A–D versus E channel protocol.
它支持非随机评价，但不会唯一规定你必须采用A–D对E的频道协议。
RL-07 · ACL Anthology · 2021
READ NEXT
随后研读
03 · Defend the evaluation / 把评价设计讲明白
Koh et al. (2021) · ICML 2021, PMLR 139, 5637–5664
CORE · 核心内容
WILDS benchmarks models under real shifts across domains such as hospitals, cameras, geography and time, showing that in-distribution scores can hide deployment failures.
WILDS在医院、相机、地区和时间等真实分布偏移下评测模型，说明同分布高分可能掩盖部署失败。
WHY IT MATTERS · 与你课题的关系
It helps position held-out-channel and later-time tests as distribution-shift evaluation rather than ordinary validation.
它帮助你把“留出频道+未来时间复测”定位为分布偏移评价，而不是普通验证。
READ FOR · 带着问题读
Which metadata define a meaningful domain shift?
哪些元数据能定义有意义的领域偏移？
Which performance gap should be reported alongside average precision?
除平均精确率外，还应报告哪种性能落差？
DO NOT OVERCLAIM · 不要过度推断
WILDS supplies evaluation principles and examples, not a ready-made Chinese trading dataset.
WILDS提供评价原则和案例，不提供可直接使用的中文交易数据集。
RL-08 · PMLR conference record · 2021
EXTEND LATER
进阶扩展
03 · Defend the evaluation / 把评价设计讲明白
Geifman & El-Yaniv (2019) · ICML 2019, PMLR 97, 2151–2159
CORE · 核心内容
SelectiveNet learns to make predictions on a chosen coverage of cases and abstain on the rest, formalising the trade-off between coverage and prediction risk.
SelectiveNet让模型只对一定覆盖范围的样本作答，并对其余样本弃权，从而形式化覆盖率与预测风险之间的权衡。
WHY IT MATTERS · 与你课题的关系
It provides later-stage language for uncertainty, abstention and human review if the MPhil protocol expands beyond ranking alone.
如果MPhil后续从排序扩展到不确定性管理，它为弃权与人工复核提供进阶理论语言。
READ FOR · 带着问题读
What is the risk–coverage curve?
什么是风险—覆盖率曲线？
Could abstention protect difficult victim and warning cases?
弃权机制能否保护困难的受害者和警示样本？
DO NOT OVERCLAIM · 不要过度推断
This is an optional extension; it should not enlarge the first MPhil experiment before the core comparison works.
这是可选扩展；在核心对比跑通前，不应因此扩大第一阶段MPhil实验。
RL-09 · PMLR conference record · 2019