😎 Awesome list of papers about knowledge-enhanced Question generation with notes.
:white_check_mark: : already reading carefully
:fire:: high citation in recent years
:hammer_and_wrench:: available code
Content
[TOC]
一、利用答案和语言特征
:white_check_mark: :fire: Neural question generation from text: A preliminary study, in EMNLP 2017. [pdf]
graph LR
en((encoder)) --bi-GRU--> fe((feature-Rich)) --> word-vecotr
fe --> lexcial-feature-embedding-vectors --> POS+NER
fe --> answer-position-embedding --> BIO-tagging
word-vecotr --> 双向的隐藏层
POS+NER --> 双向的隐藏层
BIO-tagging --> 双向的隐藏层
de((decoder)) --带注意力机制,使用加性注意力--> maxout-hidden+具体需要看reference论文
de --> GRU
de --> Copy-Mechanism,一样使用加性注意力 --> 计算出概率从source句子中直接copy单词
:white_check_mark: :fire: :hammer_and_wrench: Learning to Ask: Neural Question Generation for Reading Comprehension, in ACL 2017. [pdf] [official code (torch)]
graph LR
任务难点 --更加接近于人类--> 同义词替换+知识引入 --> 相关工作 --> 过去:rule-based
相关工作 --> 其他数据映射自然语言
Seq2Seq --> en((encoder)) --bidirectional--> soft计算注意力分数 --> lstm((LSTM)) --> only-sentence
lstm --> sentence+paragraph --> truncate截断,当然更好的方法是切片
Seq2Seq --> de((decoder)) --word-level-prediction--> LSTM((LSTM)) --> 隐藏层初始化 --basic-model --> 句子encoder的最后隐藏层
LSTM --oours--> 句子+段落的encoder输出
:white_check_mark: :fire: Improving Neural Question Generation using Answer Separation, in AAAI 2019. [pdf]
传统的有POS(词性标注)和NER(命名实体识别)。后续还有一些更加细微的处理
:fire: Learning to Generate Questions by Learning What not to Generate, in WWW 2019. [pdf]


Question-type Driven Question Generation, in EMNLP 2019. [pdf]




二、段落级别特征
:fire: :hammer_and_wrench: Paragraph-level Neural Question Generation with Maxout Pointer and Gated Self-attention Networks, in EMNLP 2018. [pdf] [torch]
主要贡献都在模型上面,基于seq2seq设计:
gate self-attention: 个人觉得是一套很常用的框架,可以学习一下,也非常简单
Maxout ==Pointer== & Decoding 全新的处理 copy 机制 (有空可以自行去看看代码!)
之前copy得分:$\operatorname{sc}^{\text {copy }}\left(y_{t}\right)=\left{\begin{array}{l}\sum_{k, \text { where } x_{k}=y_{t}} r_{t, k}, \quad y_{t} \in \chi \ -i n f, \text{otherwise}\end{array} \quad\right.$ , 问题在于若文章中某个单词重复出现多次,则对该单词copy也会多,影响语句通顺。
为此改进为Maxout Pointer: $$ \operatorname{sc}^{\text {copy }}\left(y_{t}\right)= \begin{cases}\max {k, \text { where } x{k}=y_{t}} r_{t, k}, & y_{t} \in \chi \ -i n f, & \text { otherwise }\end{cases} $$
:fire: Natural Question Generation with Reinforcement Learning Based Graph-to-Sequence Model,in ICLR 2020. [pdf] [torch]
:fire: Improving Question Generation With to the Point Context, in EMNLP 2019. [pdf]
三、多任务训练
Multi-Task Learning with Language Modeling for Question Generation, in EMNLP 2019. [pdf]
:fire: Improving Question Generation with Sentence-level Semantic Matching and Answer Position Inferring, in AAAI 2019. [pdf]
出发点是是解决生成错误的疑问词和copy原文中无关词的问题
作者认为生成错误词的原因是没有正确的利用answer position信息,copy无关词的原因是缺乏局部语义信息。
为了分别缓解这两个问题,作者也是设计了两个辅助任务:

:hammer_and_wrench: Varifocal Question Generation for Fact-checking, in EMNLP 2022. [pdf] [torch]
使用QG辅助做Fact-checking任务(很有用的一个应用场景),也有对应的数据集

:fire: :hammer_and_wrench: Generative Language Models for Paragraph-Level Question Generation, in EMNLP 2022 oral. [pdf] [demo]
大模型做QG的Benchmark
- 非常solid的一项工作
BLEU等:hammer_and_wrench: Educational Question Generation of Children Storybooks via Question Type Distribution Learning and Event-Centric Summarization, in ACL 2022. [pdf] [torch]
动机
hugh-cognitive-demand的问题很有意义low-dognitive-demand (LCD) 问题描述,比如说 他是谁?方法

:hammer_and_wrench: CQG: A Simple and Effective Controlled Generation Framework for Multi-hop Question Generation, in ACL 2022. [pdf] [torch]
动机

方法
先试用GAT抽取多文档之间联系,Standford corenlp toolkit来建立实体图
设计flag tag来约束chain of entities,保证问题生成的复杂程度(注意:并不是teacher forcig)
$$
\operatorname{flag}i^t= \begin{cases}0 & x_i \text { is not a constrain } \ 1 & x_i \text { does not appear in } y{1: t} \ 2 & x_i \text { appear in } y_{1: t}\end{cases}
$$


:hammer_and_wrench: [Meta Learning] DSM: Question Generation over Knowledge Base via Modeling Diverse Subgraphs with Meta-learner, in EMNLP 2022. [pdf] [torch]
COLING 2022 也有一篇CQG使用Meta Learning做QG的,和这篇很像
动机

方法模型,类似MAML的元学习器, MAML简单解析

:hammer_and_wrench: Modeling What-to-ask and How-to-ask for Answer-unaware Conversational Question Generation, in ACL 2023. [pdf] [torch]
有点教育学 + AI的意思,但实际上还是比较简单的做法

:hammer_and_wrench: Synthetic Question Value Estimation for Domain Adaptation of Question Answering, in ACL 2023. [pdf] [[torch](https://github. com/xiangyue9607/QVE)]


:hammer_and_wrench: Let the LLMs Talk: Simulating Human-to-Human Conversational QA via Zero-Shot LLM-to-LLM Interactions, in WSDM 2024. [pdf] [code]


:hammer_and_wrench: Diversify Question Generation with Retrieval-Augmented Style Transfer, in EMNLP 2023. [pdf] [torch]



:hammer_and_wrench: Diversity Enhanced Narrative Question Generation for StoryBooks, in EMNLP 2023. [pdf] [torch]



:fire: Prompting Large Language Models with Chain-of-Thought for Few-Shot Knowledge Base Question Generation, in EMNLP 2023. [pdf]

:triangular_ruler: Three Questions Concerning the Use of Large Language Models to Facilitate Mathematics Learning, in EMNLP Findings (Short Paper). [pdf]


:fire: Generating Natural Questions About an Image, in ACL 2022.
第一篇VQG的工作
动机


方法
:hammer_and_wrench: [No Visual] Entity Guided Question Generation with Contextual Structure and Sequence Information Capturing, in AAAI 2021. [pdf] [torch]
:white_check_mark: :hammer_and_wrench: Mixture Content Selection for Diverse Sequence Generation, in EMNLP 2019.[pdf] [torch]
:hammer_and_wrench: Radial Graph Convolutional Network for Visual Question Generation, in IEEE Transactions on Neural Networks and Learning Systems 2020. [pdf] [torch]
在2022年的今天,VQA任务不太可能从刷分的角度来入手了 [Blog链接]
VQA任务是什么
介绍之前的模型和方法
欢迎来到Transformer的时代
- 2019:尝试多模态表征
- 2020:拥抱多模态表征
- 2021:统一构架的探索
machine reading comprehension (MRC)和question answering (QA)的关系其实是相对独立的。Pure VQA任务一般是没有引入额外的文本内容,只是单纯的有${图, 问句, 回答}$。而Multimodal MRC任务,实际上就只是引入了额外的context作为VQA任务的知识,并且更加注重于自然语言的理解。MRC的主要任务类型一共有四种,分别为:
[非深度学习方法] Answer-Type Prediction for Visual Question Answering,in CVPR 2016. [pdf]
Differential Attention for Visual Question Answering, in CVPR 2018. [pdf]

:hammer_and_wrench: Multiple Objects-Aware Visual Question Generation, in ACM MM 2021. [pdf]

:hammer_and_wrench: Difficulty-Controllable Visual Question Generation, in APWeb-WAIM 2021. [pdf]
难度可控的问题生成:采用了教育学领域收集好的问题难度标签(DIF), 详见链接
在VQA2.0数据集的基础上构建了一个包含区分为容易和难的问题数据集

:hammer_and_wrench: Learning to Caption Images Through a Lifetime by Asking Questions, in ICCV 2019. [pdf] [torch]
:hammer_and_wrench: Inferential Visual Question Generation, in MM 2022. [pdf] [dataset & code]
无需模型训练的,就是人为规定函数模版规则的方法生成的。



:hammer_and_wrench: Multi-VQG: Generating Engaging Questions for Multiple Images, in EMNLP 2022. [pdf]] [dataset]
Baseline就是
VL-T5,详见论文

:hammer_and_wrench: Why Did the Chicken Cross the Road? Rephrasing and Analyzing Ambiguous Questions in VQA, in ACL 2023. [pdf] [dataset&torch]
![]() | ![]() |
|---|
:fire: :hammer_and_wrench: Location-Aware Visual Question Generation with Lightweight Models, in EMNLP 2023. [pdf] [torch]
很搞的一个动机


:fire: Advancing Large Multi-modal Models with Explicit Chain-of-Reasoning and Visual Question Generation, 2024-01. [pdf]



:hammer_and_wrench: ConVQG: Contrastive Visual Question Generation with Multimodal Guidance, in AAAI 2024. [pdf] [torch]



:fire: :hammer_and_wrench: Chain-of-Exemplar: Enhancing Distractor Generation for Multimodal Educational Question Generation, in ACL 2024. [pdf] [[torch](https://github.com/ Luohh5/Chain-of-Exemplar)]


Video Question Generation via Semantic Rich Cross-Modal Self-Attention Networks Learning, in ICASSP 2020. [pdf]
Multi-Turn Video Question Generation via Reinforced Multi-Choice Attention Network, in T-CSVT 2021.[pdf]
End-to-End Video Question-Answer Generation with Generator-Pretester Network, in T-CSVT 2021. [pdf]
:hammer_and_wrench: [2023.3未开源] Learning to Retrieve Videos by Asking Questions, in MM 2022. [pdf] [torch]
动机

和过去方法的不同


:hammer_and_wrench: MuKEA: Multimodal Knowledge Extraction and Accumulation for Knowledge-based Visual Question Answering, in CVPR 2022. [pdf] [torch]
动机

主要贡献

细节
三个损失函数的设计
Triplet TransE Loss: 保持embedding的结构(通过对比学习)$$ \mathcal{L}{\text {TransE }}=\sum{t^{+} \in \mathcal{A}^{+}} \sum_{t^{-} \in \mathcal{A}^{-}}\left[\gamma+\mathrm{d}\left(h+\boldsymbol{h}, \boldsymbol{t}^{+}\right)-\mathrm{d}\left(\boldsymbol{h}+\boldsymbol{r}, \boldsymbol{t}^{-}\right)\right]_{+} $$
Triplet Consistency Loss: 保证严格的拓扑关系$$ \mathcal{L}_{\mathrm{Tri}}=\operatorname{MSE}\left(h+r, t^{+}\right) $$
Semantic Consistency Loss: 保持在语义空间中的表达一致性$$ {P\left(t^{+}\right)=\operatorname{softmax}\left((T)^{T}(h+r)\right)} \{\mathcal{L}_{\mathrm{Sem}}=-\log \left(P\left(t^{+}\right)\right)} $$
预训练和微调策略
VQA 2.0数据集上进行预训练来收集视觉主导的知识KB-VQA数据集上进行微调关于尾部Entity
teacher-forcinglook up table $\mathbf{T}$的最小距离$$ \boldsymbol{t}{\inf f}=\underset{\boldsymbol{t}i \in T}{\arg \min } \mathrm{d}\left(\boldsymbol{h}{\text {inf } f}+\boldsymbol{r}{\text {inf } f}, \boldsymbol{t}_{\mathrm{i}}\right) $$

[VCR] Explicit Cross-Modal Representation Learning for Visual Commonsense Reasoning, in TMM 2022. [pdf
VCR任务的reasoning过程,不再那么隐式

:hammer_and_wrench: Knowledge-Grounded Self-Rationalization via Extractive and Natural Language Explanations, in ICML 2022. [pdf] [code (not released in 2022/11/20)]
注意这篇论文早于👇下面的两篇论文,所以结果上和下面两篇论文有明显的差距
动机
方法(这篇论文方法上文章没有做很详细的,描述,具体代码也没有公开,所以无法很详细的解析)

HardKuma吩咐的方法来学习latent selectors 选择合适的元素
:hammer_and_wrench: NLX-GPT: A Model for Natural Language Explanations in Vision and Vision-Language Tasks, in CVPR 2022. [pdf] [pdf]
A再生成E两步走, 缺乏对reasoning过程的考虑
CLIP
新的自动化评价指标
E和A的相关性

红色为小于0的值,都置为0,然后计算distance的平均值,值越小,模型的Bias越小
:hammer_and_wrench: [Viisual Explanation] Chunk-aware Alignment and Lexical Constraint for VisualEntailment with Natural Language Explanations, in MM 2022. [pdf] [Talk] [torch]

动机
方法(注意:使用了Oscar这个预训练模型)

:hammer_and_wrench: SwapMix: Diagnosing and Regularizing the Over-Reliance on Visual Context in Visual Question Answering, in CVPR 2022. [pdf] [torch]
动机

发现
贡献
方法

Weakly Supervised Relative Spatial Reasoning for Visual Question Answering, in ICCV 2021. [pdf]
动机

贡献
zero-shot能力,只需要10%的训练数据进行训练方法
预处理工作
抽取图像的深度(预训练好的AdaBins模型)
在上述两个任务中,预测都是实值向量。评估了这些任务的两个变体:
$$ b_c=\frac{1}{\lambda^{C-\left|c-\frac{C}{2}\right|+1}}-\frac{1}{\lambda^{C-\left|c-\frac{C}{2}\right|+2}} \forall c \in{0 . . C-1} $$
弱监督代理任务 $SR$
gt(预处理得到的) 进行计算Patches视觉信息

:fire: :hammer_and_wrench: An Empirical Study of GPT-3 for Few-Shot Knowledge-Based VQA, in AAAI 2022. [pdf] [torch] [博客链接]

PLM对知识的理解能力做VQA,不需要根据知识库进行检索few-shot的形式:fire: CLIP Models are Few-shot Learners: Empirical Studies on VQA and Visual Entailment, in ACL 2022. [pdf]
实验性的文章
动机
方法


question type $\times$ answer type 路(如果按照词汇表来分就太多了), $65 \times 3 = 195$ waysCLIP模型的一部分参数):fire: :hammer_and_wrench: A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge, in ECCV 2022. [home page]
OK-VQA的升级版本
动机
数据集知识类型

:hammer_and_wrench: [VCR] Heterogeneous Graph Learning for Visual Commonsense Reasoning, in NIPS 2019. [pdf] [torch]
VQA不太一样,R: 解释(Reason)


:hammer_and_wrench: [VCR] Connective Cognition Network for Directional Visual Commonsense Reasoning,in NIPS 2019. [pdf] [torch]
与上一篇论文思想比较类似,参考神经科学当中将神经元整合起来的思想
做法



:fire: [Knowledge-Based] KRISP: Integrating Implicit and Symbolic Knowledge for Open-Domain Knowledge-Based VQA, in CVPR 2021. [pdf] [博客链接]
动机
验证利用外部知识+隐式知识结合的做QA的能力
隐式知识可以从基于大规模语料预训练的模型有效地学习。
而显示的知识可以从知识库中的明确的、符号化的知识中学习。
将两种模型进行集成,即可同时结合隐式知识与显式知识进行推理。

方法
构建外部知识库,由于原来的知识太多了,根据规则对知识进行一定的筛选
DBPedia, ConceptNet , VisualGenome and hasPart KB ,进行节点过滤,只保留包含图像目标检测label的节点,最后的知识图包含8000节点和36000条边,采用RGCN作为卷积模型

模型设计(很简单)

:hammer_and_wrench: :fire: [多模态版本COT] *Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering, in NIPS 2022. [pdf] [博客] [data&code]
手速实在太快了
动机
贡献

数据集特点


方法
图像处理就简单的转成Caption,作为visual context
使用COT改进UnifiedQA (微调)
使用GPT-3+COT

实验结果(具体结果可以看论文或者博客)


:hammer_and_wrench: A Unified End-to-End Retriever-Reader Framework for Knowledge-based VQA, in MM 2022. [pdf] [torch]
这篇论文实验味道比较重,挺严谨的一个工作
动机
方法


推理时候
结论
:hammer_and_wrench: A Multi-Modal Context Reasoning Approach for Conditional Inference on Joint Textual and Visual Clues, in ACL 2023. [pdf] [torch]
动机

贡献
方法

:fire: :hammer_and_wrench: Combo of Thinking and Observing for Outside-Knowledge VQA, in ACL 2023. [pdf] [torch]



:hammer_and_wrench: [I-T Retrieval] CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not, in CVPR 2023. [pdf] [torch]



:hammer_and_wrench: [T-I Retrieval] Pic2Word: Mapping Pictures to Words for Zero-shot Composed Image Retrieval, in CVPR 2023. [pdf] [torch]


:hammer_and_wrench: Symbolic Replay: Scene Graph as Prompt for Continual Learning on VQA Task, in AAAI 2023. [pdf] [Dataset&torch]





:hammer_and_wrench: :fire: PROMPTCAP: Prompt-Guided Task-Aware Image Captioning, in ICCV 2023. [pdf] [torch]
这篇论文我关注很久了,估计是投CVPR没中,中了ICCV





:hammer_and_wrench: SlideVQA: A Dataset for Document Visual Question Answering on Multiple Images, in AAAI 2023. [pdf] [[torch](https://github. com/nttmdlab-nlp/SlideVQA)]


:hammer_and_wrench: :fire: Encyclopedic VQA: Visual questions about detailed properties of fine-grained categories, in ICCV 2023. [pdf] [dataset]

:hammer_and_wrench: :fire: Improving Zero-shot Visual Question Answering via Large Language Models with Reasoning Question Prompts, in MM 2023. [pdf] [torch]


:hammer_and_wrench: Analyzing Modular Approaches for Visual Question Decomposition, in EMNLP 2023. [pdf] [torch]


:hammer_and_wrench: From Wrong To Right: A Recursive Approach Towards Vision-Language Explanation, in EMNLP 2023. [pdf] [torch]

:hammer_and_wrench: :fire: Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback, in NeurIPS 2023. [project]
动机

方法


:hammer_and_wrench: Emergent Communication in Interactive Sketch Question Answering, in NeurIPS 2023. [pdf] [torch]



TOA: Task-oriented Active VQA, in NeurIPS 2023. [pdf]

:fire: :hammer_and_wrench: LoRA: A Logical Reasoning Augmented Dataset for Visual Question Answering, in NeurIPS 2023. [pdf] [torch]] [dataset]



:hammer_and_wrench: :fire: Exploring Question Decomposition for Zero-Shot VQA, in NeurIPS 2023. [project]



:hammer_and_wrench: One Self-Configurable Model to Solve Many Abstract Visual Reasoning Problems, in AAAI 2024. [pdf] [torch]



:hammer_and_wrench: Image Content Generation with Causal Reasoning, in AAAI 2024. [pdf]] [torch]



:fire: :hammer_and_wrench: Lecture Presentations Multimodal Dataset: Towards Understanding Multimodality in Educational Videos, in ICCV 2023. [pdf]] [torch&Dataset]
我认为这篇论文最大的问题就是Baseline对比的方法太过古老,在BLIP-2这种模型表现如何呢?数据集贡献确实大



:fire: :hammer_and_wrench: ECG-QA: A Comprehensive Question Answering Dataset Combined With Electrocardiogram, in NeurIPS 2023. [pdf] [dataset]



Object Attribute Matters in Visual Question Answering, in AAAI 2024. [pdf]
个人觉得很无聊

:fire: KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning, in AAAI 2024. [pdf]

:fire: :hammer_and_wrench: EQA-MX: Embodied Question Answering using Multimodal Expression, in ICLR 2024. [pdf]



:hammer_and_wrench: :fire: Towards More Faithful Natural Language Explanation Using Multi-Level Contrastive Learning in VQA, in AAAI 2024. [pdf] [torch]



:fire: :hammer_and_wrench: Bidirectional Contrastive Split Learning for Visual Question Answering, in AAAI 2024. [pdf]

:hammer_and_wrench: :fire: UniChart: A Universal Vision-language Pretrained Model for Chart Comprehension and Reasoning, in EMNLP 2023. [pdf] [[torch](https://github.com/vis- nlp/UniChart)]

:hammer_and_wrench: Can Pre-trained Vision and Language Models Answer Visual Information-Seeking Questions, in EMNLP 2023. [pdf] [torch]


:hammer_and_wrench: A Symbolic Character-Aware Model for Solving Geometry Problems, in MM 2023. [pdf] [torch]
详见论文

:hammer_and_wrench: VQAttack: Transferable Adversarial Attacks on Visual Question Answering via Pre-trained Models, in AAAI 2024. [pdf]


:fire: :hammer_and_wrench: [Question Answering] Commonsense for Generative Multi-Hop Question Answering Tasks, in EMNLP 2018. [pdf] [tensorflow]
:hammer_and_wrench: [Dialogue System] Improving Knowledge-aware Dialogue Generation via Knowledge Base Question Answering, in AAAI 2020. [pdf] [torch]
[Question Answering] Using Local Knowledge Graph Construction to Scale Seq2Seq Models to Multi-Document Inputs, in EMNLP 2019. [pdf]
:fire: :hammer_and_wrench: **[Question Answering] ** Improving Multi-hop Question Answering over Knowledge Graphs usingKnowledge Base Embeddings, in ACL 2020. [pdf] [torch]
:hammer_and_wrench: Found a Reason for me? Weakly-supervised Grounded Visual Question Answering using Capsules, in CVPR 2021. [pdf] [[torch](https://github.com/aurooj/ WeakGroundedVQA_Capsules.git)]
KQA Pro: A Dataset with Explicit Compositional Programs for Complex Question Answering over Knowledge Base, in ACL 2022. [pdf] [[project](https://github.com/shijx12/ KQAPro_Baselines)]

[自然语言推理] Generated Knowledge Prompting for Commonsense Reasoning, in ACL 2022. [pdf] [torch
这篇论文的思想和
few-shotCOT很像
动机
方法
让LM在few-shot情况下来生成解释的prompt

知识融合
concat操作$$ q_0=q, q_1=\left[k_1 | q\right], \ldots, q_M=\left[k_M | q\right] $$
最后选择最合适的对,扔进去下一个LM进行推理,最后生成答案
:hammer_and_wrench: GeoMLAMA: Geo-Diverse Commonsense Probing on Multilingual Pre-Trained Language Models, in EMNLP Oral. [pdf] [benchmark]
- Probe探究了神经网络的内部机制如何对auxiliary linguistic tasks (or probe tasks, or ancillary tasks)进行分类
具体地,以BERT举例,对于一个在训练在主任务上的大型神经网络,Probe是一个插入在其中间层的浅层神经网络,通常是一个分类器层。Probe有助于探查不同层捕获的信息。使用辅助任务对Probe进行训练和验证,以发现是否捕获了此类辅助信息。原文作者给了一个图片示例:


:hammer_and_wrench: :fire: Automatic Chain of Thought Prompting in Large Language Models, in ICLR 2023. [pdf] [torch] [rebuttal]
COT主要是两种
Zero-shot COT,就简单说一句“Let’s think step by step” 就让GPT-3直接生成Manual COT 人为手动标定一些Q-C的demonstration 做 in-context learningZero-shot COT 性能不够好,Manual COT 需要手工标注,并且对demonstration的领域很敏感(以往都是根据特定领域进行标注)
Zero-shot COT生成,但是实验过程中发现,demonstration对多样性的要求很高,所以考虑使用问题聚类的方式来先分类


Iteratively Prompt Pre-trained Language Models for Chain of Thought, in EMNLP 2022. [pdf] [torch]
动机
CoT的其中一种实现方式
方法

Inductive Relation Prediction with Logical Reasoning Using Contrastive Representations, in EMNLP 2022. [pdf


:hammer_and_wrench: Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models, in ACL 2023. [pdf] [code]
COT存在的问题


:fire: Think Twice: Measuring the Effificiency of Eliminating Prediction Shortcuts of Question Answering Models, in ACL 2023. [pdf]
:fire: :hammer_and_wrench: IfQA: A Dataset for Open-domain Question Answering under Counterfactual Presuppositions, in EMNLP 2023 Outstanding Paper. [pdf] [Benchmark]

:hammer_and_wrench: [Sentence Discrimination] Learning Semantic Sentence Embeddings using Sequential Pair-wise Discriminator,in COLING 2018. [pdf] [torch]
:hammer_and_wrench: [Hierarchical Sketch&Paraphrase Generation] Hierarchical Sketch Induction for Paraphrase Generation, in ACL 2022.[pdf] [torch]
:white_check_mark: :hammer_and_wrench: [Image Caption] Generating Diverse and Descriptive Image Captions Using Visual Paraphrases, in ICCV 2019. [pdf] [torch]
该论文研究了目前图像的文本描述的多样性和具体性缺乏的问题,提出了一种基于视觉复述的两阶段解码的模型。
给定图像输入,该模型首先生成初步的句子,再将其改写为内容更加多样和丰富的描述。在MS COCO图像描述数据集上的实验显示,方法可以显著提升文本描述的多样性和具体性。
重点探索visual paraphrases 角色 + scoring function
graph LR
与人类相比 --文章中有example--> 缺少多样性和具体性 --> 两阶段视觉复述方法 --> MSCOCO数据集
故事展开:
graph LR
标准 -->流畅+相关+多样+具体 --多样性--> 形容词
流畅+相关+多样+具体 --多样性--> 细节,with
形容词 --> Pa((Paraphrase))
细节,with --> Pa
Pa --> visual-paraphrase
visual-paraphrase --> sentence_pairs --> 两阶段编码
graph LR
相关工作 --caption--> 多caption.vs.单caption --paraphrases--> 未处理特征和视觉信息 --两阶段编码--> 中间seq.vs.2captions
模型方法:
graph LR
选择视觉复述caption对 --> 评分函数 --> 设计三个Attention操作,学习到多模态知识 --> 最后softmax输出
更多细节可见我个人的slide
:white_check_mark: ::fire: :hammer_and_wrench: [Text Generation & Image Caption] Show, Control and Tell: A Framework for Generating Controllable and Grounded Captions, in CVPR 2019. [pdf] [torch]


graph LR
外部信号控制 --> 图像中的一组区域块 --> core((核心))
core --> 改变chunk的顺序
core --> 改变图像的区域
model((模型)) --基于区域的特征与状态--> LSTM((LanguageModel,两层LSTM)) --第一层--> 计算attention --注意--> 所有区域的特征向量进行mean-pooling作为图像的总体特征I
LSTM --第二层--> 预测下一个单词
model --何时切换到下一个图像区域--> 块转移门 --计算gt--> 基于第一层LSTM的状态设立一个chunk-sentinel --> 类似计算ht对sc_t-rt的attention
model --视觉词or文本次--> AdaptiveAttention --> 设置一个visual-sentinel --> 类似计算ht对sv_t-rt的attention --> attention的结果,可以计算出当前时刻模型正在关注的上下文特征ct
model --无序集合排序--> 排序网络 --> R中包含N个区域集 --全连接层--> 每个区域集的特征映射为N维向量,然后拼接在一起 --Sinkhorn算子--> 软置换矩阵
每个区域集的特征映射为N维向量,然后拼接在一起 --> 最小化软置换与真实结果之间的均方误差
每个区域集的特征映射为N维向量,然后拼接在一起 --测试匈牙利算法进行匹配--> 软置换矩阵转化为最终的置换,以此来对R进行排序
:hammer_and_wrench: Length-Controllable Image Captioning, in ECCV 2020 by Qi Wu and Mingkui Tan. [pdf] [torch]

之前的SOTA方法可能会遗漏一些关键的信息,如果我想要更加细节点的描述,他们无法生成。
方法
过去由于方法是自回归的,所以计算复杂度会随着句子长度上升而上升。在这里提出了 non-autoregressive的方法。
获取句子长度信息(level -> $[L_{low}, L_{high}]$)做embedding
提出Decode 阶段 (non-autoregressive) LaBERT
使用位置信息来预测mask
使用长度信息来预测unmask
推理的时候鼓励生成更长的句子
exponentially decay: $p_i\left(s_i=[\mathrm{EOS}]\right) \leftarrow \gamma^{L_{\text {high }}-i} p_i\left(s_i=[\mathrm{EOS}]\right), \forall i \in\left[L_{\text {low }}, L_{\text {high }}\right]$

每一步都会对最低置信度的单词进行mask
:hammer_and_wrench: Human-like Controllable Image Captioning with Verb-specific Semantic Roles, in CVPR 2021. [pdf] [torch]

动机
事件兼容性,两个不兼容的事件不应该合在一起
采样的兼容性,不合理的采样不应该出现在句子当中
对于上面的case:
verb=sit, Arg1="thing sitting", Arg2="sitting position"
verb=read, Arg0="reader", Arg1="thing read"
方法上是先抽取出来约束的标签,再decoder

:star: MAGIC: Multimodal relAtional Graph adversarIal inferenCe for Diverse and Unpaired Text-Based Image Captioning, in AAAI 2022. [pdf]
GAN的思想)
:hammer_and_wrench: :fire: Show, Edit and Tell: A Framework for Editing Image Captions, in CVPR 2020. [pdf] [torch]
:hammer_and_wrench: Towards Accurate Text-based Image Captioning with Content Diversity Exploration, in CVPR 2021. [pdf] [torch]


Improving OCR-based Image Captioning by Incorporating Geometrical Relationship, in CVPR 2021. [pdf]

:hammer_and_wrench: Towards Unique and Informative Captioning of Images, in ECCV 2020. [pdf] [torch]

:hammer_and_wrench: Comprehensive Image Captioning via Scene Graph Decomposition, in ECCV 2020. [pdf] [torch]


:hammer_and_wrench: In Defense of Scene Graphs for Image Captioning, in ICCV 2021. [pdf] [[torch](https://github.com/ Kien085/SG2Caps)]


具体方法

:hammer_and_wrench: Beyond a Pre-Trained Object Detector: Cross-Modal Textual and Visual Context for Image Captioning, in CVPR 2022. [pdf] [torch]


:hammer_and_wrench: Comprehending and Ordering Semantics for Image Captioning, in CVPR. [pdf] [torch]


:hammer_and_wrench: DIFNet: Boosting Visual Information Flow for Image Captioning, in CVPR 2022. [[pdf](DIFNet: Boosting Visual Information Flow for Image Captioning)] [torch]


:hammer_and_wrench: Injecting Semantic Concepts into End-to-End Image Captioning, in CVPR 2022. [pdf] [torch]


:hammer_and_wrench: Show, Deconfound and Tell: Image Captioning with Causal Inference, in CVPR 2022. [pdf] [[torch](https: //github.com/CUMTGG/CIIC)]

主要为了解决两个Caption存在的问题
Encoder阶段(解决分类准确性的问题)





[因果关系 + 强化学习] Dependent Multi-Task Learning with Causal Intervention for Image Captioning, in IJCAI 2021. [pdf]
说实话这篇论文写作有点太复杂了,很难看懂


NOC-REK: Novel Object Captioning with Retrieved Vocabulary from External Knowledge, in CVPR 2022. [pdf]
任务描述

当前的方法
动机
方法

匹配loss的计算
为了鼓励模型引入新的类,我们将padding(和Region feature数量一致)的空类的15%随机替换成从外部知识库中随机挑选的词作为GT
计算$Hungarian$ loss (匈牙利损失) $$ \mathcal{L}{\mathrm{H}}(\mathcal{Y}, \mathcal{V})=\sum{i=1}^K-\log \operatorname{sim}\left(\mathbf{y}i, \hat{\mathbf{v}}{\hat{\sigma}(i)}\right) $$
Image Captioning with Novel Topics Guidance and Retrieval-based Topics Re-weighting, in TMM 2022. [pdf]
利用主题(Topic)模型来做image caption
什么是主题 链接
方法

NMF(NLTK中有)来从caption中提前获取的,数量为200。[主题 + Image Caption] Show, Rethink, And Tell: Image Caption Generation With Hierarchical Topic Cues, in ICME 2021. [pdf]
动机(caption本身就具备多个或者一个主题的特征)【目前觉得topic的作用就是生成更加多语义信息+生成更加精确】
方法

:hammer_and_wrench: Explicit Image Caption Editing, in ECCV 2022. [pdf] [torch]
任务:Image Caption编辑,修改得更加合理

数据集构建
方法 (模型是基于BERT)
引入编辑算子<ADD>, <DELETE>, <KEEP>


Case Study

:hammer_and_wrench: Object-Centric Unsupervised Image Captioning, in ECCV 2022. [pdf] [torch]
主要动机
贡献
如何无监督
BLEU4很低,只有6左右
:hammer_and_wrench: Human-Centric Image Captioning, in PR 2022. [pdf] [code & dataset]
动机【动机不是很强,说得不够有道理】
方法
标注数据集,已公开
人工标注了人的身体部位的boxes

背景物体和活动物体的特征区分是通过算和人之间的similarity来得到的
:hammer_and_wrench: Matching Visual Features to Hierarchical Semantic Topics for Image Paragraph Captioning, in IJCV 2022. [pdf] [torch]
引入Topic来做paragraph Caption的生成
动机
方法
end-to-end的方法**,即训练主题模型,又生成Caption,和以往的Two-stage不一样。

Case Study

:hammer_and_wrench: :fire: ClipCap: CLIP Prefix for Image Captioning, on arXiv (2021). [pdf] [torch]
模型非常简单,跑起来也很简单。一台1080 Ti 就可以训练了!
动机
$$ \max \theta \sum{i=1}^N \sum_{j=1}^{\ell} \log p_\theta\left(c_j^i \mid x^i, c_1^i, \ldots, c_{j-1}^i\right) $$
模型架构

Controllable Image Captioning via Prompting, in AAAI 2023. [pdf]
动机

方法极其简单

:hammer_and_wrench: Learning Distinct and Representative Modes for Image Captioning, in NIPS 2022. [pdf] [torch]


Incorporating Unlikely Negative Cues for Distinctive Image Captioning, in IJCAI 2023. [pdf]


:hammer_and_wrench: Transferable Decoding with Visual Entities for Zero-Shot Image Captioning, in ICCV 2023. [pdf]] [torch]



:hammer_and_wrench: With a Little Help from your own Past: Prototypical Memory Networks for Image Captioning, in ICCV 2023. [pdf] [torch]


:hammer_and_wrench: [多模态机器翻译] CLIPTrans: Transferring Visual Knowledge with Pre-trained Models for Multimodal Machine Translation, in ICCV 2023. [pdf] [torch]



:hammer_and_wrench: MultiCapCLIP: Auto-Encoding Prompts for Zero-Shot Multilingual Visual Captioning, in ACL 2023. [pdf] [torch]


:hammer_and_wrench: [视觉信息增强LM文本生成能力] Learning to Imagine: Visually-Augmented Natural Language Generation, in ACL 2023. [pdf] [torch]
⚠️事先声明:这篇论文实际上是做纯文本的,输入的是文本,输出的也是文本,只不过是想联想生成一些图片来增强模型文本生成的能力

:hammer_and_wrench: [多模态摘要生成] CFSum: A Coarse-to-Fine Contribution Network for Multimodal Summarization, in ACL 2023. [pdf] [torch]


:fire: :hammer_and_wrench: OxfordTVG-HIC: Can Machine Make Humorous Captions from Images?, in ICCV 2023. [pdf] [project]



:hammer_and_wrench: ALIP: Adaptive Language-Image Pre-training with Synthetic Caption, in ICCV 2023. [pdf] [torch]


:hammer_and_wrench: Noise-aware Learning from Web-crawled Image-Text Data for Image Captioning, in ICCV 2023. [pdf] [torch]


:fire: :hammer_and_wrench: Learning Descriptive Image Captioning via Semipermeable Maximum Likelihood Estimation, in NeurIPS 2023. [pdf] [torch]



:hammer_and_wrench: Exploring Diverse In-Context Configurations for Image Captioning, in NeurIPS 2023. [pdf] [torch]
探索In-context Learning如何做Image Caption,有点奇怪的论文🤔


:hammer_and_wrench: Dense and Aligned Captions (DAC) Promote Compositional Reasoning in VL Models, in NeurIPS 2023. [pdf]
看起来像是做Caption过滤的


:fire: :hammer_and_wrench: Caption Anything: Interactive Image Description with Diverse Multimodal Controls, 2023-05. [pdf] [torch]



:fire: :hammer_and_wrench: Attractive Storyteller: Stylized Visual Storytelling with Unpaired Text, in ACL 2023. [pdf]] [torch]


:hammer_and_wrench: :fire: Explore and Tell: Embodied Visual Captioning in 3D Environments, in ICCV 2023. [[project](https: //aim3-ruc.github.io/ExploreAndTell)]



Cycle-Consistency Learning for Captioning and Grounding, in AAAI 2024. [pdf]


:white_check_mark: :fire: :hammer_and_wrench: [TSN] Temporal Segment Networks: Towards Good Practices for Deep Action Recognition, in ECCV 2016. [pdf] [torch]
抽取所有帧是不现实的,TSN将其等间隔分为$K$个片段(i.e., $K=16$),在每个片段中谁寄抽取一帧作为输入
提供了非常常用的数据争强方式和一些训练时候的trick(主要包括location jittering, horizontal flipping, corner cropping, and scale jittering)
仍然利用双流的思路,让每个片段信息最后通过一个共识网络再Fusion

:white_check_mark: :fire: :hammer_and_wrench: [TRN] Temporal Relation Reasoning in Videos, in ECCV 2018. [pdf] [torch]

:white_check_mark: :fire: :hammer_and_wrench: [TSM] TSM: Temporal Shift Module for Efficient Video Understanding, in ICCV 2019. [pdf] [torch]
对某些通道shift,得到前一帧或者后一帧的特征

由于shift是有损失的,为此设计残差来进行弥补(原来的与残差的对比)

:white_check_mark: :fire: :hammer_and_wrench: [LRCN] Long-term Recurrent Convolutional Networks for Visual Recognition and Description, in CVPR 2015. [pdf] [torch]
LSTM得到每帧的时序特征关于视频特征抽取,下面讲一下
netvlad系列的结构,NextVlad就是专门针对视频帧融合来做的优化。
:fire: :hammer_and_wrench: [NetVLAD] NetVLAD: CNN architecture for weakly supervised place recognition, in CVPR 2016. [pdf] [torch (simple)]
VLAD算法(实际上就是Kmeans): $$ V(j, k)=\sum_{i=1}^{N} a_{k}\left(x_{i}\right)\left(x_{i}(j)-c_{k}(j)\right), \quad k \in K, j \in D $$
本文使用CNN模拟该VLAD算法的过程

:fire: :hammer_and_wrench: [NextVLAD] NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification, in ECCV workshop 2018. [pdf] [tensorflow]
:hammer_and_wrench: [PEFT&动作识别] AIM: Adapting Image Models for Efficient Video Action Recognition, in ICLR 2023. [pdf] [torch]
大模型时代下很好的尝试
Parameter Effectient Fine-tuning
动机

方法(很巧妙很简单)

:hammer_and_wrench: Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground, in ICCV 2023 Oral. [pdf] [torch]


我们使用一个十分经典的任务(Temporal Grounding)来看看视频的特征是如何利用的
:fire::hammer_and_wrench: [Video-NLP] Learning 2D Temporal Adjacent Networks for Moment Localization with Natural Language, in AAAI 2020. [pdf] [torch]

核心思想:
2D Temporal Feature Map Extraction所示Hadamard product)$$ \mathbf{F}=\left|\left(\mathbf{w}^{S} \cdot \mathbf{f}^{S} \cdot \mathbb{1}^{T}\right) \odot\left(\mathbf{W}^{M} \cdot \mathbf{F}^{M}\right)\right|_{F} $$
IoU进行一个scale变成监督信号$$ y_{i}= \begin{cases}0 & o_{i} \leq t_{\min } \ \frac{o_{i}-t_{\min }}{t_{\max }-t_{\min }} & t_{\min }<o_{i}<t_{\max } \ 1 & o_{i} \geq t_{\max }\end{cases} $$
$$ L o s s=\frac{1}{C} \sum_{i=1}^{C} y_{i} \log p_{i}+\left(1-y_{i}\right) \log \left(1-p_{i}\right) $$
:fire: :hammer_and_wrench: Negative Sample Matters: A Renaissance of Metric Learning for Temporal Grounding, in AAAI 2022. [pdf] [torch] [blog]
主干网络是沿用TDN
使用了metric learning的方法并且引入负样本来做Temporal Grounding的任务
IoU来标定监督信号yi,与2D-TAN一样处理得来的,记得scale一下)贡献
IoU来采样), 句子和视频对应的正负样本(负样本句子从别的视频抽取过来)Trick
DistilBERT来进行编码句子损失函数计算
2D-TDN一样的BCE_lossInfoNCE loss的设计对比损失$$ \begin{aligned} &p\left(i_{s} \mid v\right)=\frac{\exp \left(\left(\mathbf{f}{i}^{S T} \mathbf{f}^{V}-m\right) / \tau{v}\right)}{\exp \left(\left(\mathbf{f}{i}^{S T} \mathbf{f}^{V}-m\right) / \tau{v}\right)+\sum_{j \neq i}^{N_{s}} \exp \left(\mathbf{f}{j}^{S T} \mathbf{f}^{V} / \tau{v}\right)} \ &p\left(i_{v} \mid s\right)=\frac{\exp \left(\left(\mathbf{f}{i}^{V T} \mathbf{f}^{S}-m\right) / \tau{s}\right)}{\exp \left(\left(\mathbf{f}{i}^{V T} \mathbf{f}^{S}-m\right) / \tau{s}\right)+\sum_{j \neq i}^{N_{v}} \exp \left(\mathbf{f}{j}^{V T} \mathbf{f}^{S} / \tau{s}\right)} \ &L_{m m}=-\left(\sum_{i=1}^{N} \log p\left(i_{v} \mid s_{i}\right)+\sum_{i=1}^{N} \log p\left(i_{s} \mid v_{i}\right)\right) \end{aligned} $$
:fire: Event-Guided Procedure Planning from Instructional Videos with Text Supervision, in ICCV 2023. [pdf]



:fire: :hammer_and_wrench: Invariant Grounding for Video Question Answering, in CVPR 2022 Best Paper Finalist. [pdf] [torch]
这篇文章感觉是一篇很标准的
CVPR的中规中矩文章,写作用词上非常出色的
Casual)还有无关帧(补偿帧Complement)memory bank来存储所有样本 (因此要注意存储的特征维度不能太大)Video as Conditional Graph Hierarchy for Multi-Granular Question Answering,in AAAI 2022. [pdf] [torch]

:fire: :hammer_and_wrench: [交通事故QA数据集] SUTD-TraffificQA: A Question Answering Benchmark and an Effificient Network for Video Reasoning over Traffific Events, in CVPR 2021. [pdf] [project]

Cross-Modal Causal Relational Reasoning for Event-Level Visual Question Answering, in TPAMI 2022. [pdf]
这篇论文模型较为复杂,所以这里只讲诉其核心思想
动机
现有方法只关注了很简单的事件,比如说看电影,无法关注真正事件级的因果关系

语言和图像当中的干扰因素(Confounder)

方法

:hammer_and_wrench: :fire: Discovering Spatio-Temporal Rationales for Video Question Answering, in ICCV 2023. [pdf] [torch]


:fire: Redundancy-aware Transformer for Video Question Answering, in MM 2023. [pdf]
和上一篇论文动机很类似的,都是同一个作者,CVPR 22最佳论文候选那个作者

:fire: :hammer_and_wrench: Tem-adapter: Adapting Image-Text Pretraining for Video Question Answer, in ICCV 2023. [pdf] [torch]



:fire: :hammer_and_wrench: A-CQUIRED: A Dataset for Answering Counterfactual Questions In Real-Life Videos, in EMNLP 2023 [pdf] [[dataset&code](https: //github.com/PlusLabNLP/acquired)]


:fire: :hammer_and_wrench: Large Language Models are Temporal and Causal Reasoners for Video Question Answering, in EMNLP 2023. [pdf] [torch]

:fire: :hammer_and_wrench: LLCP: Learning Latent Causal Processes for Reasoning-based Video Question Answer, in ICLR 2024. [pdf]
说实话我不是很能看懂,假设前提有点太强了
详见原文



[Video Caption] VX2TEXT: End-to-End Learning of Video-Based Text Generation From Multimodal Inputs, in CVPR 2021. [pdf]
:hammer_and_wrench: :fire: [Video Caption] Robust Change Captioning, in ICCV 2019. [pdf] [torch]

:hammer_and_wrench: :fire: [Video Caption] Semantic Grouping Network for Video Captioning, in AAAI 2021. [pdf] [torch]

:hammer_and_wrench: :fire: Hierarchical Context-aware Network for Dense Video Event Captioning, in ACL 2021. [pdf] [torch]
Attention机制

graph LR
SG(Semantic-Grouping) --去掉冗余phrase--> 相似度计算
SG --attention机制 --> 对其phrase和frame --> 加入对比损失,计算没有包含negative的概率
对比损失$\mathcal{L}{c a}=\sum{(V, Y) \in \mathcal{D}} \sum_{t} \sum_{i}^{M_{t}}\left(-\log p_{c a}\left(s_{i, t}\right)\right)$, $p_{c a}\left(s_{i, t}\right)=\sum_{j=1}^{N} \alpha_{i, j, t}^{p o s}$ ($\alpha^{pos}$ 为正样本时候对齐注意力的权重)
:fire: :hammer_and_wrench: A New Comprehensive Benchmark for Semi-supervised Video Anomaly Detection and Anticipation, in CVPR 2023. [pdf] [torch]



:fire: Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model, in CVPR 2022. [pdf] [torch]
将CoOP(图像分类)论文思想放入OD当中
动机
模型方法(具体见论文)
连续型Prompt学习:


:fire: :hammer_and_wrench: Multi-Modal Classifiers for Open-Vocabulary Object Detection, in ICML 2023. [pdf] [code]


:fire: :star: Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond, in TACL 2022. [pdf] [blog]
:hammer_and_wrench: CauAIN: Causal Aware Interaction Network for Emotion Recognition in Conversations, in IJCAI 2022. [pdf] [torch]
在对话当中找情感相关的因果线索
动机:

方法
利用 ATOMIC 常识知识语料库,获得对话历史每句的 6 种因果线索

利用因果线索,建模对话历史

:hammer_and_wrench: Knowledge-Bridged Causal Interaction Network for Causal Emotion Entailment, in AAAI 2023. [pdf] [torch]
这份工作和上面👆的IJCAI 2022:CauAIN: Causal Aware Interaction Network for Emotion Recognition in Conversations那个工作idea是一样的,只不过做了两个不同的任务而已
主要动机都是找寻对话当中的因果线索

方法模型


:fire: :hammer_and_wrench: Everything Has a Cause: Leveraging Causal Inference in Legal Text Analysis, in NAACL 2021. [pdf] [torch]


:hammer_and_wrench: :fire: Deconfounded Video Moment Retrieval with Causal Intervention, in SIGIR 2021. [pdf] [torch]
动机
方法


:fire: :hammer_and_wrench: Interventional Video Grounding with Dual Contrastive Learning, in CVPR 2021. [pdf] [torch] (2023年1月仍为开源)
动机

方法

因果推理模块(通过loss来调整)


:fire: :hammer_and_wrench: Two Causal Principles for Improving Visual Dialog, in CVPR 2020. [pdf] [torch] [zhihu]
含老师团队第一篇因果推理方向的文章
一句话总结:
- 从因果图角度审视视觉对话任务,切断对话历史与答案的直接因果效应,添加混杂因子[用户偏好]构建描述真实世界的因果图。
动机
两个原则
$P_1$: H对话历史记录,不应该直接影响A

$P_2$: 用户偏好会导致回答的不一样

由于用户偏好不可观测,因此作者通过三种方式来近似采样。

构建两个原则的因果图:

具体后门调整及其概率运算公式,详见论文及其附录


:fire: :hammer_and_wrench: [因果关系] Visual Commonsense R-CNN, in CVPR 2020. [pdf] [torch] [blog]
出自MReal, 张含望老师团队的工作,非常Solid的一篇工作
- 目标是训练基于
Faster-RCNN训练一个更强的feature extractor可以捕获视觉上的常识信息。- 这篇论文实在太多细节和推理了,建议看我自己的GoodNote上的笔记!
动机
现在的模型无法学习到视觉常识(Commonsense):人和椅子 -> 人可以坐在椅子上。但在NLP中,常识的信息已经放在特征里面了

数据集的偏差会导致无法捕捉到常识信息
因果理论就是用来发现==现象背后的不变规律==的,是一种鲁棒的预测。这与常识本身不就很相似吗,我们人类也是从生活中不断总结积累这些不变的、鲁棒的经验或者因果规律,并把他们叫做常识。 比如,看见凳子知道可以坐,看见pizza知道可以吃。
Association 和 Intervention(分层)的计算 $$ \begin{gathered} P(Y \mid X)=\sum_z P(Y \mid X, z) P(z \mid X)=\frac{P(Y, X)}{P(X)} \ P(Y \mid d o(X))=\sum_z P(Y \mid X, z) P(z)=\sum_z \frac{P(Y, X, z) P(z)}{P(X, z)} \end{gathered} $$ 其中 $X, Y, z$分别代表了图片中的object label,同时这里我们用物体出现的频率来代替概率,比如 $P(Sink|Hair drier)$就是用“含有$Sink$和$Hair drier$两者的图片数”比上“只含有Hair drier的图片数”计算得到的。画出两者计算结果差异的对比图(只标明了20类):

两个Case的分析
方法(因果干预Intervention)
代理任务(无监督学习):给定RoI X的feature去预测RoI Y的类别
包括很多潜在的混杂因子,如果直接预测周围物体Y就不可避免的会被上文提到的混杂因子confounder所影响。根据我们刚刚介绍的**“do算子”的理论,解决的办法也不难,只要能找到confounder然后对他们使用backdoor理论**进行控制即可。
混杂因子是什么? 我们直接把整个数据集上的object RoI特征(Faster RCNN中来)在每个类别上取平均,当作这个类别的表示,进而构建出一个 类别数x1024 的confounder字典作为$Z$(比如MSCOCO有80类,就是 80x1024),它包含着所有可能的混杂因子。
后门调整

模型

注意:VC R-CNN的实现和原先的Faster R-CNN相比,去除了RPN网络(Region Proposal Network),不再训练网络propose边界框,而是直接将数据集ground-truth的bounding box坐标输入到其中,直接提取region的特征。而在训练完成后的feature提取阶段,相对应的,只要给定图片和bounding box坐标,都可以获得对应的VC特征。就这样,我们利用bottomup特征已有的边界框坐标提取VC特征后,将其并在先前的bottomup特征上作为新的特征。我们在传统的 Vision&Language 三大任务上挑选了经典model和SOTA model进行了测试,发现在各个任务上都取得了明显的提升,尤其是在image captioning上的提升尤其大。同时为了验证性能的提升不是由于参数增多带来的,我们还在原有特征上并上了ablative的特征(单独object特征,用correlation计算的特征),具体可以参考论文的实验部分。
:hammer_and_wrench: [指代表达] Deconfounded Visual Grounding, in AAAI 2022. [pdf] [torch] (2023.1.17未开源)
动机

方法
构建因果图

R语言信息,因为其实视觉上的bias其他一些方向论文都有所解决了由于$G$不可观测,因此需要近似采样

:fire: :hammer_and_wrench: Unbiased Scene Graph Generation from Biased Training, in CVPR 2020. [pdf] [torch] [zhihu]
动机
贡献

问题定义(有偏见的数据标注的原因)
无偏的思想(content:内因,context:外因)
人类在有偏见的大自然中生长,在拥抱好的context的同时,避免不好的context,并与content一起做出无偏见的决定。
其潜在的机制是基于因果关系的(causality-based):决策是通过追求由content引起的主要因果效应,而不是追求由context引起的副作用来做出的。然而,机器是基于可能性的(likelihood-based),会产生有偏结果。
故论文认为,无偏预测的关键是教会机器如何区分主要作用(main effect)和副作用
为了在无偏预测中追求主要作用,论文提出赋予机器反事实思维(counterfactual thinking): If i had not seen the content, would I still make the same prediction?
反事实思维:事实与反事实之间的比较,将会自然地消除context偏差的影响,因为context是两者之间唯一不变的东西。
如图,左侧图片是所谓的事实场景,也可以说是原始场景;右侧图片是反事实场景,就是**将原始场景中content(狗和冲浪板的视觉特征)去除,其他部分(如scene和object classes)保持不变,就像object的视觉特征从未出现过。**通过这两者的比较,我们可以专注于关系的主要视觉影响,同时也不丢失context。

方法


TDE(Total Direct Effect)方法没有引入任何额外的参数,也可以说没有针对模型的有偏训练进行任何改动,其使用原始SGG模型进行了两次预测,将两次预测的结果进行差值运算,最终得到无偏见的预测。所以TDE方法是模型“不可见”的,广泛适用于各种SGG模型。
:fire: :hammer_and_wrench: Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect, in NIPS 2020. [pdf] [torch] [zhihu]
动机
核心思想
需要利用原始的长尾分布来学习特征提取的原因在于,大量的尾部类别其实不足以提供足够的样本来学习鲁棒的特征表达。人类描述罕见的物体时,往往是通过和已知常见类的比较,比如会说狮鹫是有着狮子的身体,鹰的翅膀和头的生物,而不必要单独拿一堆狮鹫的图片出来,让你死记硬背住狮鹫的长相。

发现优化器的动量项时,这货不就是在训练数据时引入数据分布,从而产生shortcut的元凶么。
方法,详细的推到有点复杂,相近论文或者知乎

方法速成步骤

:fire: :hammer_and_wrench: Interventional Few-Shot Learning, in NIPS 2020. [pdf] [torch] [blog]
博客讲得很全面,具体可以见博客对本文的讲解
- 方法挺会玩的!
动机


问题建模(探索Many-shot Learning 和 Few-shot在因果关系上的不一致地方)


$$ P(Y \mid d o(X=\boldsymbol{x}))=\sum_d P(Y \mid X=\boldsymbol{x}, D=d, C=g(\mathbf{x}, d)) P(D=d) $$

:fire: :hammer_and_wrench: Introspective Distillation for Robust Question Answering, in NIPS 2021. [pdf] [torch] [zhihu]

:fire: :hammer_and_wrench: Distilling Causal Effect of Data in Class-Incremental Learning, in CVPR 2021. [pdf] [torch]
动机
对撞节点的存在使得模型对新数据会产生灾难性遗忘
过去的方法当中


思路
文章细节详见开头的博客链接
:fire: :hammer_and_wrench: Causal Attention for Vision-Language Tasks, in CVPR 2021. [pdf] [torch] [Blog]
我的想法和博客最后说的类似,似乎本文推理推得不是特别彻底,但是仍然是一篇优秀的论文
动机:解决当前VL模型当中训练数据集当中bias

思想

方法(前门调整机制做Causal Attention,此处不做展开讨论,详见论文)


结果case

:fire: :hammer_and_wrench: Counterfactual Zero-Shot and Open-Set Visual Recognition, in CVPR 2021. [pdf] [torch] [zhihu]
生成式的因果模型,博客很详细,详见知乎!
方法
样本特征和类别特征之间解耦


:fire: :hammer_and_wrench: Counterfactual VQA: A Cause-Effect Look at Language Bias, in CVPR 2021. [pdf] [torch]
动机
方法(基于ensemble的VQA模型做的一个推理框架,相当简单)


因果推理角度解释(具体见论文)

Learning Causal Effects on Hypergraphs, Best Paper of KDD 2022. [pdf]
Entropic Causal Inference: Graph Identifiability, in ICML 2022. [pdf]
新的因果推断架构,可以不借助intervention
理论性很强的一篇文章
熵因果推断:
通过寻找数据的信息-理论上最简单的结构解释,即最小熵模型,从观测数据中学习两个变量之间的因果图。
在这个工作中,首先推广了松弛假设下的因果图可辨识性结果。
然后,我们展示了第一个可识别的结果,使用熵的方法学习超过两个节点的因果图。

:hammer_and_wrench: [2023.01未开源] Disentangle and Remerge: Interventional Knowledge Distillation for Few-Shot Object Detection from A Conditional Causal Perspective, in AAAI 2023. [pdf] [torch]
本文方法上和实验上还有很多亮点,这里主要就说因果推理


Learning to Imagine: Integrating Counterfactual Thinking in Neural Discrete Reasoning, in ACL 2022. [pdf]
动机

方法

:hammer_and_wrench: :star2: Causality Inspired Representation Learning for Domain Generalization, in CVPR 2022 Oral. [pdf] [torch]
傅立叶变换结合因果推理,超级solid的工作
动机
方法
前提:傅立叶变换当中

核心思想与流程概览,具体见论文

$$ \min {\hat{g}, \hat{h}{1}, \hat{h}{2}} \mathcal{L}{c l s}^{s u p}+\mathcal{L}{c l s}^{i n f}+\tau \mathcal{L}{F a c}, \quad \min {\hat{w}} \mathcal{L}{c l s}^{s u p}-\mathcal{L}_{c l s}^{i n f}, $$
:fire: [CLIP] EI-CLIP: Entity-aware Interventional Contrastive Learning for E-commerce Cross-modal Retrieval, in CVPR 2022. [pdf]
动机


方法

[领域泛化] GCISG: Guided Causal Invariant Learning for Improved Syn-to-real Generalization, in ECCV 2022. [pdf]
个人认为这篇论文写得有点夸张了,实际上就那么回事。
动机
因果发现

方法(很简单)

:hammer_and_wrench: Causality-aware Concept Extraction based on Knowledge-guided Prompting, in ACL 2023. [pdf] [torch]
方法极其简单
动机


因果分析

方法(极其简单,加多一个分类出来的Topic)

:hammer_and_wrench: COLA: Contextualized Commonsense Causal Reasoning from the Causal Inference Perspective, in ACL 2023. [pdf] [[torch](https://github. com/HKUST- KnowComp/COLA)]
和2022年那篇ICML论文ROCK做的任务是一样的,就是纯文本当中的因果事件检测
动机

方法(理论依据详见论文)

:fire: CFL: Causally Fair Language Models Through Token-level Attribute Controlled Generation, in ACL Finding 2023. [pdf]
有点晦涩写的,不是很看得懂


模型方法


:fire: :hammer_and_wrench: Causal-Debias: Unifying Debiasing in Pretrained Language Models and Fine-tuning via Causal Invariant Learning, in ACL 2023. [pdf] [torch]
通过学习因果不变性,来缓解PLM在fine-tune时候的bias问题


$do(N=n)$求invariant loss
$$
\min \mathcal{L}_{\text {invariant }}=\mathbb{E}_n(\mathcal{R})+\operatorname{Var}_n(\mathcal{R})
$$

:fire: :hammer_and_wrench: Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference, in ACL 2023 Oral. [pdf] [torch]



[VQA] Reducing Vision-Answer Biases for Multiple-Choice VQA, in TIP 2023. [pdf]


Causal Intervention and Counterfactual Reasoning for Multi-modal Fake News Detection, in ACL 2023. [pdf]



:hammer_and_wrench: :fire: A Causal Framework to Quantify the Robustness of Mathematical Reasoning with Language Models, in ACL 2023. [pdf] [code]
非常Perfect的关于因果推理 + LLM + 数学推理的Empirical Study


:hammer_and_wrench: Random Boxes Are Open-world Object Detectors, in CVPR 2023. [pdf] [torch]





:fire: :hammer_and_wrench: Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning, in ICCV 2023. [pdf] [torch]



:fire: Variational Causal Inference Network for Explanatory Visual Question Answering, in ICCV 2023. [pdf]
本文公式推导较为复杂,具体可以见论文




:fire: A Multi-modal Debiasing Model with Dynamical Constraint for Robust Visual Question Answering, in ACL Findings 2023. [pdf]



:fire: Neuro-Symbolic Procedural Planning with Commonsense Prompting, in ICLR 2023. [pdf]
A Causal Inference Look at Unsupervised Video Anomaly Detection, in AAAI 2022. [pdf]

:hammer_and_wrench: Deconfounded Multimodal Learning for Spatio-temporal Video Grounding, in MM 2023. [pdf] [torch]



Understanding Chat Messages for Sticker Recommendation in Messaging Apps, in AAAI 2020. [pdf]

:hammer_and_wrench: :fire: Learning to Respond with Stickers: A Framework of Unifying Multi-Modality in Multi-Turn Dialog, in WWW 2020. [pdf] [torch]


SER30K: A Large-Scale Dataset for Sticker Emotion Recognition, in MM 2022 Oral. [pdf] [torch]


数据集特点(sticker emotion recognition)

方法(Baseline)


:hammer_and_wrench: :fire: Selecting Stickers in Open-Domain Dialogue through Multitask Learning, in ACL Finding 2022. [pdf] [torch]


Named Entity Recognition
当前竞赛NER任务的baseline:
:fire: Bidirectional LSTM-CRF Models for Sequence Tagging, in 2015. [pdf] [code
:hammer_and_wrench: :fire: Fast and Accurate Entity Recognition with Iterated Dilated Convolutions, in EMNLP 2017. [pdf] [tensorflow]

核心思想

Bi-LSTM-CRF快了非常多,而且精度没有下滑:hammer_and_wrench: :fire: BOND: BERT-Assisted Open-Domain Named Entity Recognition with Distant Supervision , in KDD 2020. [pdf] [torch]
远距离监督的问题
trade-off 在标注准确度和覆盖范围之间想法
第一阶段使用RoBERTa微调,适应NER任务
Early stopping 方法防止数据过拟合还有对未知数据增强泛化能力POS识别潜在实体,然后通过语料库计算最小损失确定实体
第二阶段自我学习框架 (teacher-student模型2)

两阶段的BOND框架
:hammer_and_wrench: :fire: A Boundary-aware Neural Model for Nested Named Entity Recognition , in EMNLP 2019. [pdf] [torch]
Nested NER的问题:hammer_and_wrench: :fire: Cross-Domain NER using Cross-Domain Language , in ACL 2020. [pdf] [torch]

最底下的一层是数据层,标准情况一下总共有四份语料,分别对应两个domain下的两个task(NER和语言建模)。其中Source Domain(即保证有标记数据用于NER的domain)对应之前提到的News Domain,因为论文中Source Domain使用的是新闻数据。另外如果是无监督抽取Target Domain数据则只有三份语料。
由底向上第二层是Word Embedding层,论文中的Word Embedding结合了词级别和字符级别的向量表示。即把词向量和一个词的字符序列形成的矩阵经过CNN处理后的向量concatenate起来。
第三层是双向LSTM,用于序列处理第二层的数据,生成前后向hidden state。
第四层LM和CRF,即task model层。我们可以看到第四层有三个模块,两个是用于NER的CRF模型,分别对应Source Domain和Target Domain。另一个是基于第三层BiLSTM的语言模型,NSSoftmax是指这个语言模型利用Negative Sampling Softmax的方式进行训练。
Bi-LSTM的参数是生成的,不同domain的不同task需要的LSTM的参数和$I$有关$$ \begin{equation} \theta_{\mathrm{LSTM}}^{d, t}=\mathbf{W} \otimes \mathbf{I}{d}^{D} \otimes \mathbf{I}{t}^{T}, \end{equation} $$
NER的未来
既然模型打不动了,然后我找了找 ACL2020做NER的论文,看看现在的NER还在做哪些事情,主要分几个方面
作者:王岳王院长 链接:https://zhuanlan.zhihu.com/p/166496466 来源:知乎 著作权归作者所有。商业转载请联系作者获得授权,非商业转载请注明出处。
:fire: :hammer_and_wrench: Knowledgeable Prompt-tuning: Incorporating Knowledge into Prompt Verbalizer for Text Classifification, in ACL 2022. [pdf] [torch]] [知乎博客]
将外部知识融入当prompt-tuning当中做TC任务

:fire: [层次化主题] Deep Latent Dirichlet Allocation with Topic-Layer-Adaptive Stochastic Gradient Riemannian MCMC, in ICML 2017. [pdf]
理论性非常强的论文
- 详细的理论推导请见原文。

:fire: :hammer_and_wrench: Quark: Controllable Text Generation with Reinforced [Un]learning, in NIPS 2022. [pdf] [torch]
InstructGPT很类似,但是不再训练奖励模型,而是使用prompt来取代(评判生成的是否是好句子)
:fire: :hammer_and_wrench: Large Language Models Are Human-Level Prompt Engineers, in ICLR 2023. [pdf] [torch] [zhihu]

:fire: [2023.5.6未开源] SCOTT: Self-Consistent Chain-of-Thought Distillation, in ACL 2023. [pdf]
动机

方法【teacher:GPT-neox20B student:T5 3B】

细节


:fire: Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes
方法也太简单了,其实就是当前LLM的范式,用大语言模型生成的东西来蒸馏学习小模型

:fire: Poisoning Language Models During Instruction Tuning, in ICML 2023. [pdf] [torch]
动机

:fire: :hammer_and_wrench: Outline, Then Details: Syntactically Guided Coarse-To-Fine Code Generation, in ICML 2023. [pdf] [torch]
动机
最大贡献
方法

$S3$和$S4$是重点,tokenizer的encode伪代码如下:

模型(预训练+微调)

:hammer_and_wrench: Distilling Script Knowledge from Large Language Models for Constrained Language Planning, in ACL 2023. [pdf] [torch]


:fire: :hammer_and_wrench: BadPrompt: Backdoor Attacks on Continuous Prompts, in NIPS 2022. [pdf] [torch]


:fire: :hammer_and_wrench: RL4F: Generating Natural Language Feedback with Reinforcement Learning for Repairing Model Outputs, in ACL 2023. [pdf] [torch]
太棒的工作了!

T5-large (0.77M)】
:fire: :hammer_and_wrench: Are You Copying My Model? Protecting the Copyright of Large Language Models for EaaS via Backdoor Watermark, in ACL 2023. [pdf]] [[torch](https: //github.com/yjw1029/EmbMarker)]


:hammer_and_wrench: :fire: Tree of Thoughts: Deliberate Problem Solving with Large Language Models, in 2023 05. [pdf] [torch]




:hammer_and_wrench: Plug-and-Play Knowledge Injection for Pre-trained Language Models, in ACL 2023. [pdf] [[torch](https://github.com/ THUNLP/Knowledge-Plugin)]

:hammer_and_wrench: :fire: Mixture-of-Domain-Adapters: Decoupling and Injecting Domain Knowledge to Pre-trained Language Models’ Memories, in ACL 2023. [pdf] [[torch](https://github.com/ Amano-Aki/Mixture-of-Domain-Adapters)]

:fire: :hammer_and_wrench: Grounding Language Models to Images for Multimodal Inputs and Outputs, in ICML 2023. [pdf] [torch]
单卡A6000进行训练,很友好



:hammer_and_wrench: [COT] Element-aware Summarization with Large Language Models: Expert-aligned Evaluation and Chain-of-Thought Method, in ACL 2023. [pdf] [code]
相当于重构了现在的Summarization数据集


:fire: :hammer_and_wrench: Synthetic Text Generation with Differential Privacy: A Simple and Practical Recipe, in ACL 2023. [pdf] [torch]


:fire: :hammer_and_wrench: Editing Large Language Models: Problems, Methods, and Opportunities, in 2023.05. [pdf] [torch]

:fire: :hammer_and_wrench: LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models, in ICCV 2023. [pdf] [torch]


:fire: :hammer_and_wrench: CoTDet: Affordance Knowledge Prompting for Task Driven Object Detection, in ICCV 2023. [pdf] [torch]


:hammer_and_wrench: :fire: K2: A Foundation Language Model for Geoscience Knowledge Understanding and Utilization, in WSDM 2024. [pdf] [dataset&code]

:hammer_and_wrench: :fire: Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning, in EMNLP 2023 Best Paper. [pdf] [[torch](https://github.com/lancopku/ label-words-are-anchors)] [zhihu]

✨ 标签词作为锚的信息流$H_1$ :在LLM的浅层,标签词聚集示例具体内容的信息,为深层形成语义表示。 $H_2$:在LLM的深层,模型从聚集后的标签词位置中提取信息,形成最终预测。


Q1: Analysis of different ICL formats on the final prediction like random labels, reversed labels (e.g.,True->False, False->True), and label agencies (replace labels with meaningless characters) may be helpful.
缺少对于不同的 ICL 格式对最终预测的影响的分析,如随机标签、反向标签(如真->假、假->真)和标签机构(用无意义字符替换标签),可能会有所帮助。
分析:粗略的看,本文的结论和之前的工作得到的结论“ICL中的标签正确性对于结果似乎影响不大”[1] 可能有矛盾之处。如果本文结论合理,在深层,ICL机制是进行对label words处的聚合,那在示例中标签是错误的时候,label words处聚合到的信息(或者说聚合到的文本到标签的映射关系信息)有可能是有误导性的,进而导致性能下降;但是[1]中实验表明“ICL中的标签正确性对于结果似乎影响不大”,似乎矛盾。
对此可能合理的解释是:深层聚合时,不止参考了label words处的聚合信息,还参考了待预测样本的本身信息;对于能力较强的LLM(如gpt-3.5),有一定排除前面聚合的误导信息的能力,而只参考前面的格式;对于能力较弱的(如本文用的gpt2-xl,gpt-j-6b),这种“排除误导信息”的能力较弱,因此会表现出“icl中给随机标签得到的结果和随即猜测差不多”(也就是作者在rebuttal回复的)。综上,本文所分析的icl机制也许还存在一些待补全的部分,即对于待预测样本本身文本信息的参考程度。
[1] EMNLP 2022. Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?
:hammer_and_wrench: :fire: Can Large Language Models Infer Causation from Correlation?, in ICLR 2024. [pdf]] [dataset] [torch]




:fire: :hammer_and_wrench: Knowledge Rumination for Pre-trained Language Models, in EMNLP 2023. [pdf] [torch] [zhihu]
具体见论文,是知识编辑方面的工作


:fire: :hammer_and_wrench: When Do Program-of-Thought Works for Reasoning?, in AAAI 2024. [pdf] [torch]


:fire: :hammer_and_wrench: INSTRUCTSCORE: Explainable Text Generation Evaluation with Fine-grained Feedback, in EMNLP 2023. [pdf] [torch]



:fire: :hammer_and_wrench: GENOME: GenerativE Neuro-symbOlic visual reasoning by growing and reusing ModulEs, in ICLR 2024. [pdf] [project]


:fire: :hammer_and_wrench: Learning From Mistakes Makses LLM Better Reasoner, in 2024-02. [pdf] [torch]




:fire: :hammer_and_wrench: Safe RLHF: Safe Reinforcement Learning from Human Feedback, in ICLR 2024. [pdf] [torch] [zhihu]
理论和代码可以见知乎博客,很详细

:fire: :hammer_and_wrench: Digital Socrates: Evaluating LLMs through Explanation Critiques, in ACL 2024. [[project](https://allenai.org/data/ digital-socrates)]

:fire: :hammer_and_wrench: Navigating Dataset Documentations in AI: A Large-Scale Analysis of Dataset Cards on Hugging Face, in ICLR 2024. [project]

:fire: :hammer_and_wrench: What's documented in AI? Systematic Analysis of 32K AI Model Cards, in Nature Machine Intelligence 2024. [pdf] [Code]
:hammer_and_wrench: :fire: ModelGPT: Unleashing LLM’s Capabilities for Tailored Model Generation, in 2024-02. [project]


:hammer_and_wrench: :fire: What is the Best Way for ChatGPT to Translate Poetry?, in ACL 2024. [project]

:fire: :hammer_and_wrench: INCHARACTER: Evaluating Personality Fidelity in Role-Playing Agents through Psychological Interviews, in ACL 2024. [project]


:fire: :hammer_and_wrench: Generating Summaries with Controllable Readability Levels, in EMNLP 2023. [pdf] [[torch](https://github.com/amazon-science/ controllable-readability-summarization)]


:fire: Dr.Academy: A Benchmark for Evaluating Questioning Capability in Education for Large Language Models, in ACL 2024. [pdf]


:hammer_and_wrench: OceanGPT: A Large Language Model for Ocean Science Tasks, in ACL 2024. [project]



:fire: :hammer_and_wrench: Aligning Large Language Models with Human Preferences through Representation Engineering, in ACL 2024. [pdf] [torch]


:fire: :hammer_and_wrench: LexEval: A Comprehensive Chinese Legal Benchmark for Evaluating Large Language Models, in NeurIPS 2024. [project]

:hammer_and_wrench: AST-Trans: Code Summarization with Efficient Tree-Structured Attention, in ICSE 2022. [pdf] [torch]
代码摘要的生成

方法

实验
:fire: :hammer_and_wrench: Multi-target Backdoor Attacks for Code Pre-trained Models, in ACL 2023. [pdf] [torch]
代码仓库很完善,但这篇论文写作一般般吧

:hammer_and_wrench: EGFE: End-to-end Grouping of Fragmented Elements in UI Designs with Multimodal Learning, in ICSE 2024. [pdf] [Dataset&Code]
但这个数据集好像只是组合元素,不是生成前端代码



:fire: Android in the Wild: A Large-Scale Dataset for Android Device Control, in NeurIPS 2023. [pdf] [[dataset](https://github.com/google-research/google-research/ tree/master/android_in_the_wild)]

:hammer_and_wrench: Pairwise GUI Dataset Construction Between Android Phones and Tablets, in NeurIPS 2023. [pdf] [dataset]


多模态预训练
:fire: :hammer_and_wrench: [Cross-Modal&Contrastive Learning] UNIMO: Towards Unified-Modal Understanding and Generation via Cross-Modal Contrastive Learning, in ACL(long paper) 2021. [pdf] [project from Baidu]
:hammer_and_wrench: [MultiModal] UniT: Multimodal Multitask Learning with a Unified Transformer, ICCV 2021. [pdf] [project from Fair]
:fire: :hammer_and_wrench: Leveraging Visual Knowledge in Language Tasks: An Empirical Study on Intermediate Pre-training for Cross-modal Knowledge Transfer, in ACL 2022. [pdf] [torch (2022.12.01未开源)]
本文是一篇实验性的文章,实验的方法写得不错!
什么叫做:intermediate pre-training?
- 在预训练好的模型上,补充一些数据集或者语料库继续预训练
动机

方法(以下两种方法来弥补PLM在这方面的缺陷)
text knowledge transfer: 即使用image caption进行知识迁移,设计了如下两个训练目标:
cross-modal knowledge transfer: 即使用图片和文本以及V&L的训练来进行知识迁移,设计了如下几个训练目标:
Voken Classification: voken采用token层面的text2image检索来迁移视觉知识,它假设每一个token都有一个视觉域的voken与其对应,训练目标就是在所有预先设定好的voken中将正确的voken检索出来
Masked Language Modeling with Visual Clues: 给定图片作为线索,预测mask掉的token,比MLM多了图片作为输入,目标函数是一样的
Cross-Modal Contrastive Learning (CMCL): 和CLIP一样(不同在于这里把视觉网络给冻住),是跨模态的对比学习

Cross-Modal Knowledge Distillation (CMKD): 将在MSCOCO数据集上进行对比学习的多模态模型作为teacher model,将一个语言模型作为student,在纯文本语料Wiki103上进行知识蒸馏
模型示意图

结论
:hammer_and_wrench: :fire:[Relation CLIP] RelCLIP: Adapting Language-Image Pretraining for Visual Relationship Detection via Relational Contrastive Learning, in EMNLP 2022. [pdf] [torch]
动机
CLIP可以识别对象之间的relation方法(很简单)
Commonsense Knowledge 就是Conceptual Caption数据中来

负样本收集:

:star: [MM-KG + CLIP] Contrastive Language-Image Pre-Training with Knowledge Graphs, in NIPS 2022. [pdf] [rebuttal]
动机


:hammer_and_wrench: PuMer: Pruning and Merging Tokens for Efficient Vision Language Models, in ACL 2023. [pdf] [[torch](https://github.com/ csarron/PuMer)]

:fire: Learning to Estimate Shapley Values with Vision Transformers, in ICLR 2023. [pdf] [torch]
很棒的一篇Shapley Values 结合 ViT的论文

:fire: Knowledge-Aware Prompt Tuning for Generalizable Vision-Language Models, in ICCV 2023. [pdf]
动机
传统的CLIP对于一些unseen class 泛化能力不是很好
CLIP + 外部知识 (个人觉得 CVPR的那几篇更好)


:hammer_and_wrench: [多语言CLIP] mCLIP: Multilingual CLIP via Cross-lingual Transfer, in ACL 2023. [pdf] [[torch](https://github.com/ huawei-noah/noah-research/NLP/mclip)]


:hammer_and_wrench: Rethinking Multimodal Entity and Relation Extraction from a Translation Point of View, in ACL 2023. [pdf] [torch]




:fire: :hammer_and_wrench: Preserving Modality Structure Improves Multi-Modal Learning, in ICCV 2023. [pdf] [[torch](https://github. com/Swetha5/Multi_Sinkhorn_Knopp)]

:hammer_and_wrench: [复杂的图像检索任务] A Neural Divide-and-Conquer Reasoning Framework for Image Retrieval from Linguistically Complex Text, in ACL 2023. [pdf] [torch]




:hammer_and_wrench: :fire: Bayesian Prompt Learning for Image-Language Model Generalization, in ICCV 2023. [pdf] [[torch](https://github.com/saic-fi/Bayesian- Prompt-Learning)]


:fire: :hammer_and_wrench: A Unified Objective for Novel Class Discovery, in ICCV 2021 Oral. [pdf] [torch] [blog]

:fire: :hammer_and_wrench: E2VPT: An Effective and Efficient Approach for Visual Prompt Tuning, in ICCV 2023. [pdf] [torch]

:fire: SINC: Self-Supervised In-Context Learning for Vision-Language Tasks, in ICCV 2023. [pdf]



:fire: :hammer_and_wrench: Make the U in UDA Matter: Invariant Consistency Learning for Unsupervised Domain Adaptation, in NeurIPS 2023. [pdf] [torch]


:hammer_and_wrench: ViStruct: Visual Structural Knowledge Extraction via Curriculum Guided Code-Vision Representation, in EMNLP 2023. [pdf] [[torch](https://github.com/ Yangyi-Chen/vi-struct)]



:hammer_and_wrench: :fire: DDCoT: Duty-Distinct Chain-of-Thought Prompting for Multimodal Reasoning in Language Models, in NeurIPS 2023. [project]


:fire: :hammer_and_wrench: See and Think: Embodied Agent in Virtual Environment, 2023-12. [project]


:hammer_and_wrench: [表情包] MEMECAP: A Dataset for Captioning and Interpreting Memes, in EMNLP 2023. [pdf] [torch]




:fire: :hammer_and_wrench: CLOVA: A Closed-LOop Visual Assistant with Tool Usage and Update, in 2023-12. [pdf] [torch] [知乎]

:hammer_and_wrench: :fire: [2024-01 刚放榜,未开源] ADAPTING TO DISTRIBUTION SHIFT BY VISUAL DOMAIN PROMPT GENERATION, in ICLR 2024. [pdf]


:hammer_and_wrench: DOC2PPT: Automatic Presentation Slides Generation from Scientific Documents, in AAAI 2022. [pdf] [torch]



:fire: :hammer_and_wrench: VLIS: Unimodal Language Models Guide Multimodal Language Generation, in EMNLP 2023. [pdf] [torch]


:fire: Multitask Multimodal Prompted Training for Interactive Embodied Task Completion, in EMNLP 2023. [pdf] [torch]

:fire: :hammer_and_wrench: :star2: Fine-tuning Multimodal LLMs to Follow Zero-shot Demonstrative Instructions, in ICLR 2024. [pdf] [torch]





:fire: :hammer_and_wrench: UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model, in EMNLP 2023. [pdf] [torch]] [zhihu]
依托达摩院mPLUG系列大模型的



:fire: :hammer_and_wrench: mPLUG-PaperOwl: Scientific Diagram Analysis with the Multimodal Large Language Model, in 2023-11. [pdf] [[torch](https://github.com/X-PLUG/mPLUG-DocOwl/ tree/main/PaperOwl)]


:fire: :hammer_and_wrench: Too Large; Data Reduction for Vision-Language Pre-Training, in ICCV 2023. [pdf] [code]

:hammer_and_wrench: :fire: MathVista Evaluating Math Reasoning in Visual Contexts, in ICLR 2024 Oral. [project]
动机
表现


:fire: :hammer_and_wrench: BLIVA: A Simple Multimodal LLM for Better Handling of Text-Rich Visual, in AAAI 2024. [pdf] [torch]


:fire: :hammer_and_wrench: [2024-01 未开源] Imagine That! Abstract-to-Intricate Text-to-Image Synthesis with Scene Graph Hallucination Diffusion, in NeurIPS 2023. [pdf] [torch]


:fire: :hammer_and_wrench: Generating Explanations for Embodied Action Decision from Visual Observation, in MM 2023. [pdf]


:fire: :hammer_and_wrench: InstructDoc: A Dataset for Zero-Shot Generalization of Visual Document Understanding with Instructions, in AAAI 2024. [pdf] [[torch](https://github. com/nttmdlab-nlp/InstructDoc)]



:fire: :hammer_and_wrench: Improving CLIP Training with Language Rewrites, in NeurIPS 2023. [pdf] [torch]
用ChatGPT改写CLIP文本端的Prompt

:fire: :hammer_and_wrench: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation, in 2023-12. [pdf] [torch]
动机
模型方法(单卡4090~2.5天就可以收敛)


:fire: :hammer_and_wrench: FROSTER: Frozen CLIP is A Strong Teacher for Open-Vocabulary Action Recognition, in ICLR 2024. [pdf]


:fire: :hammer_and_wrench: A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation, in ICLR 2024. [pdf]



:fire: :hammer_and_wrench: Small Language Model Meets with Reinforced Vision Vocabulary, in 2024-02. [pdf] [torch]

:fire: :hammer_and_wrench: MoE-LLaVA: Mixture of Experts for Large Vision-Language Models, in 2024-02. [pdf] [torch]


:fire: :hammer_and_wrench: Cheap and Quick: Efficient Vision-Language Instruction Tuning for Large Language Models, in NeurIPS 2023. [pdf] [project] [[INT4量
Truncated — view the full README on GitHub.
141 commits
😎 Awesome list of papers about knowledge-enhanced Question generation with notes.
:white_check_mark: : already reading carefully
:fire:: high citation in recent years
:hammer_and_wrench:: available code
Content
[TOC]
一、利用答案和语言特征
:white_check_mark: :fire: Neural question generation from text: A preliminary study, in EMNLP 2017. [pdf]
graph LR
en((encoder)) --bi-GRU--> fe((feature-Rich)) --> word-vecotr
fe --> lexcial-feature-embedding-vectors --> POS+NER
fe --> answer-position-embedding --> BIO-tagging
word-vecotr --> 双向的隐藏层
POS+NER --> 双向的隐藏层
BIO-tagging --> 双向的隐藏层
de((decoder)) --带注意力机制,使用加性注意力--> maxout-hidden+具体需要看reference论文
de --> GRU
de --> Copy-Mechanism,一样使用加性注意力 --> 计算出概率从source句子中直接copy单词
:white_check_mark: :fire: :hammer_and_wrench: Learning to Ask: Neural Question Generation for Reading Comprehension, in ACL 2017. [pdf] [official code (torch)]
graph LR
任务难点 --更加接近于人类--> 同义词替换+知识引入 --> 相关工作 --> 过去:rule-based
相关工作 --> 其他数据映射自然语言
Seq2Seq --> en((encoder)) --bidirectional--> soft计算注意力分数 --> lstm((LSTM)) --> only-sentence
lstm --> sentence+paragraph --> truncate截断,当然更好的方法是切片
Seq2Seq --> de((decoder)) --word-level-prediction--> LSTM((LSTM)) --> 隐藏层初始化 --basic-model --> 句子encoder的最后隐藏层
LSTM --oours--> 句子+段落的encoder输出
:white_check_mark: :fire: Improving Neural Question Generation using Answer Separation, in AAAI 2019. [pdf]
传统的有POS(词性标注)和NER(命名实体识别)。后续还有一些更加细微的处理
:fire: Learning to Generate Questions by Learning What not to Generate, in WWW 2019. [pdf]


Question-type Driven Question Generation, in EMNLP 2019. [pdf]




二、段落级别特征
:fire: :hammer_and_wrench: Paragraph-level Neural Question Generation with Maxout Pointer and Gated Self-attention Networks, in EMNLP 2018. [pdf] [torch]
主要贡献都在模型上面,基于seq2seq设计:
gate self-attention: 个人觉得是一套很常用的框架,可以学习一下,也非常简单
Maxout ==Pointer== & Decoding 全新的处理 copy 机制 (有空可以自行去看看代码!)
之前copy得分:$\operatorname{sc}^{\text {copy }}\left(y_{t}\right)=\left{\begin{array}{l}\sum_{k, \text { where } x_{k}=y_{t}} r_{t, k}, \quad y_{t} \in \chi \ -i n f, \text{otherwise}\end{array} \quad\right.$ , 问题在于若文章中某个单词重复出现多次,则对该单词copy也会多,影响语句通顺。
为此改进为Maxout Pointer: $$ \operatorname{sc}^{\text {copy }}\left(y_{t}\right)= \begin{cases}\max {k, \text { where } x{k}=y_{t}} r_{t, k}, & y_{t} \in \chi \ -i n f, & \text { otherwise }\end{cases} $$
:fire: Natural Question Generation with Reinforcement Learning Based Graph-to-Sequence Model,in ICLR 2020. [pdf] [torch]
:fire: Improving Question Generation With to the Point Context, in EMNLP 2019. [pdf]
三、多任务训练
Multi-Task Learning with Language Modeling for Question Generation, in EMNLP 2019. [pdf]
:fire: Improving Question Generation with Sentence-level Semantic Matching and Answer Position Inferring, in AAAI 2019. [pdf]
出发点是是解决生成错误的疑问词和copy原文中无关词的问题
作者认为生成错误词的原因是没有正确的利用answer position信息,copy无关词的原因是缺乏局部语义信息。
为了分别缓解这两个问题,作者也是设计了两个辅助任务:

:hammer_and_wrench: Varifocal Question Generation for Fact-checking, in EMNLP 2022. [pdf] [torch]
使用QG辅助做Fact-checking任务(很有用的一个应用场景),也有对应的数据集

:fire: :hammer_and_wrench: Generative Language Models for Paragraph-Level Question Generation, in EMNLP 2022 oral. [pdf] [demo]
大模型做QG的Benchmark
- 非常solid的一项工作
BLEU等:hammer_and_wrench: Educational Question Generation of Children Storybooks via Question Type Distribution Learning and Event-Centric Summarization, in ACL 2022. [pdf] [torch]
动机
hugh-cognitive-demand的问题很有意义low-dognitive-demand (LCD) 问题描述,比如说 他是谁?方法

:hammer_and_wrench: CQG: A Simple and Effective Controlled Generation Framework for Multi-hop Question Generation, in ACL 2022. [pdf] [torch]
动机

方法
先试用GAT抽取多文档之间联系,Standford corenlp toolkit来建立实体图
设计flag tag来约束chain of entities,保证问题生成的复杂程度(注意:并不是teacher forcig)
$$
\operatorname{flag}i^t= \begin{cases}0 & x_i \text { is not a constrain } \ 1 & x_i \text { does not appear in } y{1: t} \ 2 & x_i \text { appear in } y_{1: t}\end{cases}
$$


:hammer_and_wrench: [Meta Learning] DSM: Question Generation over Knowledge Base via Modeling Diverse Subgraphs with Meta-learner, in EMNLP 2022. [pdf] [torch]
COLING 2022 也有一篇CQG使用Meta Learning做QG的,和这篇很像
动机

方法模型,类似MAML的元学习器, MAML简单解析

:hammer_and_wrench: Modeling What-to-ask and How-to-ask for Answer-unaware Conversational Question Generation, in ACL 2023. [pdf] [torch]
有点教育学 + AI的意思,但实际上还是比较简单的做法

:hammer_and_wrench: Synthetic Question Value Estimation for Domain Adaptation of Question Answering, in ACL 2023. [pdf] [[torch](https://github. com/xiangyue9607/QVE)]


:hammer_and_wrench: Let the LLMs Talk: Simulating Human-to-Human Conversational QA via Zero-Shot LLM-to-LLM Interactions, in WSDM 2024. [pdf] [code]


:hammer_and_wrench: Diversify Question Generation with Retrieval-Augmented Style Transfer, in EMNLP 2023. [pdf] [torch]



:hammer_and_wrench: Diversity Enhanced Narrative Question Generation for StoryBooks, in EMNLP 2023. [pdf] [torch]



:fire: Prompting Large Language Models with Chain-of-Thought for Few-Shot Knowledge Base Question Generation, in EMNLP 2023. [pdf]

:triangular_ruler: Three Questions Concerning the Use of Large Language Models to Facilitate Mathematics Learning, in EMNLP Findings (Short Paper). [pdf]


:fire: Generating Natural Questions About an Image, in ACL 2022.
第一篇VQG的工作
动机


方法
:hammer_and_wrench: [No Visual] Entity Guided Question Generation with Contextual Structure and Sequence Information Capturing, in AAAI 2021. [pdf] [torch]
:white_check_mark: :hammer_and_wrench: Mixture Content Selection for Diverse Sequence Generation, in EMNLP 2019.[pdf] [torch]
:hammer_and_wrench: Radial Graph Convolutional Network for Visual Question Generation, in IEEE Transactions on Neural Networks and Learning Systems 2020. [pdf] [torch]
在2022年的今天,VQA任务不太可能从刷分的角度来入手了 [Blog链接]
VQA任务是什么
介绍之前的模型和方法
欢迎来到Transformer的时代
- 2019:尝试多模态表征
- 2020:拥抱多模态表征
- 2021:统一构架的探索
machine reading comprehension (MRC)和question answering (QA)的关系其实是相对独立的。Pure VQA任务一般是没有引入额外的文本内容,只是单纯的有${图, 问句, 回答}$。而Multimodal MRC任务,实际上就只是引入了额外的context作为VQA任务的知识,并且更加注重于自然语言的理解。MRC的主要任务类型一共有四种,分别为:
[非深度学习方法] Answer-Type Prediction for Visual Question Answering,in CVPR 2016. [pdf]
Differential Attention for Visual Question Answering, in CVPR 2018. [pdf]

:hammer_and_wrench: Multiple Objects-Aware Visual Question Generation, in ACM MM 2021. [pdf]

:hammer_and_wrench: Difficulty-Controllable Visual Question Generation, in APWeb-WAIM 2021. [pdf]
难度可控的问题生成:采用了教育学领域收集好的问题难度标签(DIF), 详见链接
在VQA2.0数据集的基础上构建了一个包含区分为容易和难的问题数据集

:hammer_and_wrench: Learning to Caption Images Through a Lifetime by Asking Questions, in ICCV 2019. [pdf] [torch]
:hammer_and_wrench: Inferential Visual Question Generation, in MM 2022. [pdf] [dataset & code]
无需模型训练的,就是人为规定函数模版规则的方法生成的。



:hammer_and_wrench: Multi-VQG: Generating Engaging Questions for Multiple Images, in EMNLP 2022. [pdf]] [dataset]
Baseline就是
VL-T5,详见论文

:hammer_and_wrench: Why Did the Chicken Cross the Road? Rephrasing and Analyzing Ambiguous Questions in VQA, in ACL 2023. [pdf] [dataset&torch]
![]() | ![]() |
|---|
:fire: :hammer_and_wrench: Location-Aware Visual Question Generation with Lightweight Models, in EMNLP 2023. [pdf] [torch]
很搞的一个动机


:fire: Advancing Large Multi-modal Models with Explicit Chain-of-Reasoning and Visual Question Generation, 2024-01. [pdf]



:hammer_and_wrench: ConVQG: Contrastive Visual Question Generation with Multimodal Guidance, in AAAI 2024. [pdf] [torch]



:fire: :hammer_and_wrench: Chain-of-Exemplar: Enhancing Distractor Generation for Multimodal Educational Question Generation, in ACL 2024. [pdf] [[torch](https://github.com/ Luohh5/Chain-of-Exemplar)]


Video Question Generation via Semantic Rich Cross-Modal Self-Attention Networks Learning, in ICASSP 2020. [pdf]
Multi-Turn Video Question Generation via Reinforced Multi-Choice Attention Network, in T-CSVT 2021.[pdf]
End-to-End Video Question-Answer Generation with Generator-Pretester Network, in T-CSVT 2021. [pdf]
:hammer_and_wrench: [2023.3未开源] Learning to Retrieve Videos by Asking Questions, in MM 2022. [pdf] [torch]
动机

和过去方法的不同


:hammer_and_wrench: MuKEA: Multimodal Knowledge Extraction and Accumulation for Knowledge-based Visual Question Answering, in CVPR 2022. [pdf] [torch]
动机

主要贡献

细节
三个损失函数的设计
Triplet TransE Loss: 保持embedding的结构(通过对比学习)$$ \mathcal{L}{\text {TransE }}=\sum{t^{+} \in \mathcal{A}^{+}} \sum_{t^{-} \in \mathcal{A}^{-}}\left[\gamma+\mathrm{d}\left(h+\boldsymbol{h}, \boldsymbol{t}^{+}\right)-\mathrm{d}\left(\boldsymbol{h}+\boldsymbol{r}, \boldsymbol{t}^{-}\right)\right]_{+} $$
Triplet Consistency Loss: 保证严格的拓扑关系$$ \mathcal{L}_{\mathrm{Tri}}=\operatorname{MSE}\left(h+r, t^{+}\right) $$
Semantic Consistency Loss: 保持在语义空间中的表达一致性$$ {P\left(t^{+}\right)=\operatorname{softmax}\left((T)^{T}(h+r)\right)} \{\mathcal{L}_{\mathrm{Sem}}=-\log \left(P\left(t^{+}\right)\right)} $$
预训练和微调策略
VQA 2.0数据集上进行预训练来收集视觉主导的知识KB-VQA数据集上进行微调关于尾部Entity
teacher-forcinglook up table $\mathbf{T}$的最小距离$$ \boldsymbol{t}{\inf f}=\underset{\boldsymbol{t}i \in T}{\arg \min } \mathrm{d}\left(\boldsymbol{h}{\text {inf } f}+\boldsymbol{r}{\text {inf } f}, \boldsymbol{t}_{\mathrm{i}}\right) $$

[VCR] Explicit Cross-Modal Representation Learning for Visual Commonsense Reasoning, in TMM 2022. [pdf
VCR任务的reasoning过程,不再那么隐式

:hammer_and_wrench: Knowledge-Grounded Self-Rationalization via Extractive and Natural Language Explanations, in ICML 2022. [pdf] [code (not released in 2022/11/20)]
注意这篇论文早于👇下面的两篇论文,所以结果上和下面两篇论文有明显的差距
动机
方法(这篇论文方法上文章没有做很详细的,描述,具体代码也没有公开,所以无法很详细的解析)

HardKuma吩咐的方法来学习latent selectors 选择合适的元素
:hammer_and_wrench: NLX-GPT: A Model for Natural Language Explanations in Vision and Vision-Language Tasks, in CVPR 2022. [pdf] [pdf]
A再生成E两步走, 缺乏对reasoning过程的考虑
CLIP
新的自动化评价指标
E和A的相关性

红色为小于0的值,都置为0,然后计算distance的平均值,值越小,模型的Bias越小
:hammer_and_wrench: [Viisual Explanation] Chunk-aware Alignment and Lexical Constraint for VisualEntailment with Natural Language Explanations, in MM 2022. [pdf] [Talk] [torch]

动机
方法(注意:使用了Oscar这个预训练模型)

:hammer_and_wrench: SwapMix: Diagnosing and Regularizing the Over-Reliance on Visual Context in Visual Question Answering, in CVPR 2022. [pdf] [torch]
动机

发现
贡献
方法

Weakly Supervised Relative Spatial Reasoning for Visual Question Answering, in ICCV 2021. [pdf]
动机

贡献
zero-shot能力,只需要10%的训练数据进行训练方法
预处理工作
抽取图像的深度(预训练好的AdaBins模型)
在上述两个任务中,预测都是实值向量。评估了这些任务的两个变体:
$$ b_c=\frac{1}{\lambda^{C-\left|c-\frac{C}{2}\right|+1}}-\frac{1}{\lambda^{C-\left|c-\frac{C}{2}\right|+2}} \forall c \in{0 . . C-1} $$
弱监督代理任务 $SR$
gt(预处理得到的) 进行计算Patches视觉信息

:fire: :hammer_and_wrench: An Empirical Study of GPT-3 for Few-Shot Knowledge-Based VQA, in AAAI 2022. [pdf] [torch] [博客链接]

PLM对知识的理解能力做VQA,不需要根据知识库进行检索few-shot的形式:fire: CLIP Models are Few-shot Learners: Empirical Studies on VQA and Visual Entailment, in ACL 2022. [pdf]
实验性的文章
动机
方法


question type $\times$ answer type 路(如果按照词汇表来分就太多了), $65 \times 3 = 195$ waysCLIP模型的一部分参数):fire: :hammer_and_wrench: A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge, in ECCV 2022. [home page]
OK-VQA的升级版本
动机
数据集知识类型

:hammer_and_wrench: [VCR] Heterogeneous Graph Learning for Visual Commonsense Reasoning, in NIPS 2019. [pdf] [torch]
VQA不太一样,R: 解释(Reason)


:hammer_and_wrench: [VCR] Connective Cognition Network for Directional Visual Commonsense Reasoning,in NIPS 2019. [pdf] [torch]
与上一篇论文思想比较类似,参考神经科学当中将神经元整合起来的思想
做法



:fire: [Knowledge-Based] KRISP: Integrating Implicit and Symbolic Knowledge for Open-Domain Knowledge-Based VQA, in CVPR 2021. [pdf] [博客链接]
动机
验证利用外部知识+隐式知识结合的做QA的能力
隐式知识可以从基于大规模语料预训练的模型有效地学习。
而显示的知识可以从知识库中的明确的、符号化的知识中学习。
将两种模型进行集成,即可同时结合隐式知识与显式知识进行推理。

方法
构建外部知识库,由于原来的知识太多了,根据规则对知识进行一定的筛选
DBPedia, ConceptNet , VisualGenome and hasPart KB ,进行节点过滤,只保留包含图像目标检测label的节点,最后的知识图包含8000节点和36000条边,采用RGCN作为卷积模型

模型设计(很简单)

:hammer_and_wrench: :fire: [多模态版本COT] *Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering, in NIPS 2022. [pdf] [博客] [data&code]
手速实在太快了
动机
贡献

数据集特点


方法
图像处理就简单的转成Caption,作为visual context
使用COT改进UnifiedQA (微调)
使用GPT-3+COT

实验结果(具体结果可以看论文或者博客)


:hammer_and_wrench: A Unified End-to-End Retriever-Reader Framework for Knowledge-based VQA, in MM 2022. [pdf] [torch]
这篇论文实验味道比较重,挺严谨的一个工作
动机
方法


推理时候
结论
:hammer_and_wrench: A Multi-Modal Context Reasoning Approach for Conditional Inference on Joint Textual and Visual Clues, in ACL 2023. [pdf] [torch]
动机

贡献
方法

:fire: :hammer_and_wrench: Combo of Thinking and Observing for Outside-Knowledge VQA, in ACL 2023. [pdf] [torch]



:hammer_and_wrench: [I-T Retrieval] CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not, in CVPR 2023. [pdf] [torch]



:hammer_and_wrench: [T-I Retrieval] Pic2Word: Mapping Pictures to Words for Zero-shot Composed Image Retrieval, in CVPR 2023. [pdf] [torch]


:hammer_and_wrench: Symbolic Replay: Scene Graph as Prompt for Continual Learning on VQA Task, in AAAI 2023. [pdf] [Dataset&torch]





:hammer_and_wrench: :fire: PROMPTCAP: Prompt-Guided Task-Aware Image Captioning, in ICCV 2023. [pdf] [torch]
这篇论文我关注很久了,估计是投CVPR没中,中了ICCV





:hammer_and_wrench: SlideVQA: A Dataset for Document Visual Question Answering on Multiple Images, in AAAI 2023. [pdf] [[torch](https://github. com/nttmdlab-nlp/SlideVQA)]


:hammer_and_wrench: :fire: Encyclopedic VQA: Visual questions about detailed properties of fine-grained categories, in ICCV 2023. [pdf] [dataset]

:hammer_and_wrench: :fire: Improving Zero-shot Visual Question Answering via Large Language Models with Reasoning Question Prompts, in MM 2023. [pdf] [torch]


:hammer_and_wrench: Analyzing Modular Approaches for Visual Question Decomposition, in EMNLP 2023. [pdf] [torch]


:hammer_and_wrench: From Wrong To Right: A Recursive Approach Towards Vision-Language Explanation, in EMNLP 2023. [pdf] [torch]

:hammer_and_wrench: :fire: Divide, Evaluate, and Refine: Evaluating and Improving Text-to-Image Alignment with Iterative VQA Feedback, in NeurIPS 2023. [project]
动机

方法


:hammer_and_wrench: Emergent Communication in Interactive Sketch Question Answering, in NeurIPS 2023. [pdf] [torch]



TOA: Task-oriented Active VQA, in NeurIPS 2023. [pdf]

:fire: :hammer_and_wrench: LoRA: A Logical Reasoning Augmented Dataset for Visual Question Answering, in NeurIPS 2023. [pdf] [torch]] [dataset]



:hammer_and_wrench: :fire: Exploring Question Decomposition for Zero-Shot VQA, in NeurIPS 2023. [project]



:hammer_and_wrench: One Self-Configurable Model to Solve Many Abstract Visual Reasoning Problems, in AAAI 2024. [pdf] [torch]



:hammer_and_wrench: Image Content Generation with Causal Reasoning, in AAAI 2024. [pdf]] [torch]



:fire: :hammer_and_wrench: Lecture Presentations Multimodal Dataset: Towards Understanding Multimodality in Educational Videos, in ICCV 2023. [pdf]] [torch&Dataset]
我认为这篇论文最大的问题就是Baseline对比的方法太过古老,在BLIP-2这种模型表现如何呢?数据集贡献确实大



:fire: :hammer_and_wrench: ECG-QA: A Comprehensive Question Answering Dataset Combined With Electrocardiogram, in NeurIPS 2023. [pdf] [dataset]



Object Attribute Matters in Visual Question Answering, in AAAI 2024. [pdf]
个人觉得很无聊

:fire: KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning, in AAAI 2024. [pdf]

:fire: :hammer_and_wrench: EQA-MX: Embodied Question Answering using Multimodal Expression, in ICLR 2024. [pdf]



:hammer_and_wrench: :fire: Towards More Faithful Natural Language Explanation Using Multi-Level Contrastive Learning in VQA, in AAAI 2024. [pdf] [torch]



:fire: :hammer_and_wrench: Bidirectional Contrastive Split Learning for Visual Question Answering, in AAAI 2024. [pdf]

:hammer_and_wrench: :fire: UniChart: A Universal Vision-language Pretrained Model for Chart Comprehension and Reasoning, in EMNLP 2023. [pdf] [[torch](https://github.com/vis- nlp/UniChart)]

:hammer_and_wrench: Can Pre-trained Vision and Language Models Answer Visual Information-Seeking Questions, in EMNLP 2023. [pdf] [torch]


:hammer_and_wrench: A Symbolic Character-Aware Model for Solving Geometry Problems, in MM 2023. [pdf] [torch]
详见论文

:hammer_and_wrench: VQAttack: Transferable Adversarial Attacks on Visual Question Answering via Pre-trained Models, in AAAI 2024. [pdf]


:fire: :hammer_and_wrench: [Question Answering] Commonsense for Generative Multi-Hop Question Answering Tasks, in EMNLP 2018. [pdf] [tensorflow]
:hammer_and_wrench: [Dialogue System] Improving Knowledge-aware Dialogue Generation via Knowledge Base Question Answering, in AAAI 2020. [pdf] [torch]
[Question Answering] Using Local Knowledge Graph Construction to Scale Seq2Seq Models to Multi-Document Inputs, in EMNLP 2019. [pdf]
:fire: :hammer_and_wrench: **[Question Answering] ** Improving Multi-hop Question Answering over Knowledge Graphs usingKnowledge Base Embeddings, in ACL 2020. [pdf] [torch]
:hammer_and_wrench: Found a Reason for me? Weakly-supervised Grounded Visual Question Answering using Capsules, in CVPR 2021. [pdf] [[torch](https://github.com/aurooj/ WeakGroundedVQA_Capsules.git)]
KQA Pro: A Dataset with Explicit Compositional Programs for Complex Question Answering over Knowledge Base, in ACL 2022. [pdf] [[project](https://github.com/shijx12/ KQAPro_Baselines)]

[自然语言推理] Generated Knowledge Prompting for Commonsense Reasoning, in ACL 2022. [pdf] [torch
这篇论文的思想和
few-shotCOT很像
动机
方法
让LM在few-shot情况下来生成解释的prompt

知识融合
concat操作$$ q_0=q, q_1=\left[k_1 | q\right], \ldots, q_M=\left[k_M | q\right] $$
最后选择最合适的对,扔进去下一个LM进行推理,最后生成答案
:hammer_and_wrench: GeoMLAMA: Geo-Diverse Commonsense Probing on Multilingual Pre-Trained Language Models, in EMNLP Oral. [pdf] [benchmark]
- Probe探究了神经网络的内部机制如何对auxiliary linguistic tasks (or probe tasks, or ancillary tasks)进行分类
具体地,以BERT举例,对于一个在训练在主任务上的大型神经网络,Probe是一个插入在其中间层的浅层神经网络,通常是一个分类器层。Probe有助于探查不同层捕获的信息。使用辅助任务对Probe进行训练和验证,以发现是否捕获了此类辅助信息。原文作者给了一个图片示例:


:hammer_and_wrench: :fire: Automatic Chain of Thought Prompting in Large Language Models, in ICLR 2023. [pdf] [torch] [rebuttal]
COT主要是两种
Zero-shot COT,就简单说一句“Let’s think step by step” 就让GPT-3直接生成Manual COT 人为手动标定一些Q-C的demonstration 做 in-context learningZero-shot COT 性能不够好,Manual COT 需要手工标注,并且对demonstration的领域很敏感(以往都是根据特定领域进行标注)
Zero-shot COT生成,但是实验过程中发现,demonstration对多样性的要求很高,所以考虑使用问题聚类的方式来先分类


Iteratively Prompt Pre-trained Language Models for Chain of Thought, in EMNLP 2022. [pdf] [torch]
动机
CoT的其中一种实现方式
方法

Inductive Relation Prediction with Logical Reasoning Using Contrastive Representations, in EMNLP 2022. [pdf


:hammer_and_wrench: Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models, in ACL 2023. [pdf] [code]
COT存在的问题


:fire: Think Twice: Measuring the Effificiency of Eliminating Prediction Shortcuts of Question Answering Models, in ACL 2023. [pdf]
:fire: :hammer_and_wrench: IfQA: A Dataset for Open-domain Question Answering under Counterfactual Presuppositions, in EMNLP 2023 Outstanding Paper. [pdf] [Benchmark]

:hammer_and_wrench: [Sentence Discrimination] Learning Semantic Sentence Embeddings using Sequential Pair-wise Discriminator,in COLING 2018. [pdf] [torch]
:hammer_and_wrench: [Hierarchical Sketch&Paraphrase Generation] Hierarchical Sketch Induction for Paraphrase Generation, in ACL 2022.[pdf] [torch]
:white_check_mark: :hammer_and_wrench: [Image Caption] Generating Diverse and Descriptive Image Captions Using Visual Paraphrases, in ICCV 2019. [pdf] [torch]
该论文研究了目前图像的文本描述的多样性和具体性缺乏的问题,提出了一种基于视觉复述的两阶段解码的模型。
给定图像输入,该模型首先生成初步的句子,再将其改写为内容更加多样和丰富的描述。在MS COCO图像描述数据集上的实验显示,方法可以显著提升文本描述的多样性和具体性。
重点探索visual paraphrases 角色 + scoring function
graph LR
与人类相比 --文章中有example--> 缺少多样性和具体性 --> 两阶段视觉复述方法 --> MSCOCO数据集
故事展开:
graph LR
标准 -->流畅+相关+多样+具体 --多样性--> 形容词
流畅+相关+多样+具体 --多样性--> 细节,with
形容词 --> Pa((Paraphrase))
细节,with --> Pa
Pa --> visual-paraphrase
visual-paraphrase --> sentence_pairs --> 两阶段编码
graph LR
相关工作 --caption--> 多caption.vs.单caption --paraphrases--> 未处理特征和视觉信息 --两阶段编码--> 中间seq.vs.2captions
模型方法:
graph LR
选择视觉复述caption对 --> 评分函数 --> 设计三个Attention操作,学习到多模态知识 --> 最后softmax输出
更多细节可见我个人的slide
:white_check_mark: ::fire: :hammer_and_wrench: [Text Generation & Image Caption] Show, Control and Tell: A Framework for Generating Controllable and Grounded Captions, in CVPR 2019. [pdf] [torch]


graph LR
外部信号控制 --> 图像中的一组区域块 --> core((核心))
core --> 改变chunk的顺序
core --> 改变图像的区域
model((模型)) --基于区域的特征与状态--> LSTM((LanguageModel,两层LSTM)) --第一层--> 计算attention --注意--> 所有区域的特征向量进行mean-pooling作为图像的总体特征I
LSTM --第二层--> 预测下一个单词
model --何时切换到下一个图像区域--> 块转移门 --计算gt--> 基于第一层LSTM的状态设立一个chunk-sentinel --> 类似计算ht对sc_t-rt的attention
model --视觉词or文本次--> AdaptiveAttention --> 设置一个visual-sentinel --> 类似计算ht对sv_t-rt的attention --> attention的结果,可以计算出当前时刻模型正在关注的上下文特征ct
model --无序集合排序--> 排序网络 --> R中包含N个区域集 --全连接层--> 每个区域集的特征映射为N维向量,然后拼接在一起 --Sinkhorn算子--> 软置换矩阵
每个区域集的特征映射为N维向量,然后拼接在一起 --> 最小化软置换与真实结果之间的均方误差
每个区域集的特征映射为N维向量,然后拼接在一起 --测试匈牙利算法进行匹配--> 软置换矩阵转化为最终的置换,以此来对R进行排序
:hammer_and_wrench: Length-Controllable Image Captioning, in ECCV 2020 by Qi Wu and Mingkui Tan. [pdf] [torch]

之前的SOTA方法可能会遗漏一些关键的信息,如果我想要更加细节点的描述,他们无法生成。
方法
过去由于方法是自回归的,所以计算复杂度会随着句子长度上升而上升。在这里提出了 non-autoregressive的方法。
获取句子长度信息(level -> $[L_{low}, L_{high}]$)做embedding
提出Decode 阶段 (non-autoregressive) LaBERT
使用位置信息来预测mask
使用长度信息来预测unmask
推理的时候鼓励生成更长的句子
exponentially decay: $p_i\left(s_i=[\mathrm{EOS}]\right) \leftarrow \gamma^{L_{\text {high }}-i} p_i\left(s_i=[\mathrm{EOS}]\right), \forall i \in\left[L_{\text {low }}, L_{\text {high }}\right]$

每一步都会对最低置信度的单词进行mask
:hammer_and_wrench: Human-like Controllable Image Captioning with Verb-specific Semantic Roles, in CVPR 2021. [pdf] [torch]

动机
事件兼容性,两个不兼容的事件不应该合在一起
采样的兼容性,不合理的采样不应该出现在句子当中
对于上面的case:
verb=sit, Arg1="thing sitting", Arg2="sitting position"
verb=read, Arg0="reader", Arg1="thing read"
方法上是先抽取出来约束的标签,再decoder

:star: MAGIC: Multimodal relAtional Graph adversarIal inferenCe for Diverse and Unpaired Text-Based Image Captioning, in AAAI 2022. [pdf]
GAN的思想)
:hammer_and_wrench: :fire: Show, Edit and Tell: A Framework for Editing Image Captions, in CVPR 2020. [pdf] [torch]
:hammer_and_wrench: Towards Accurate Text-based Image Captioning with Content Diversity Exploration, in CVPR 2021. [pdf] [torch]


Improving OCR-based Image Captioning by Incorporating Geometrical Relationship, in CVPR 2021. [pdf]

:hammer_and_wrench: Towards Unique and Informative Captioning of Images, in ECCV 2020. [pdf] [torch]

:hammer_and_wrench: Comprehensive Image Captioning via Scene Graph Decomposition, in ECCV 2020. [pdf] [torch]


:hammer_and_wrench: In Defense of Scene Graphs for Image Captioning, in ICCV 2021. [pdf] [[torch](https://github.com/ Kien085/SG2Caps)]


具体方法

:hammer_and_wrench: Beyond a Pre-Trained Object Detector: Cross-Modal Textual and Visual Context for Image Captioning, in CVPR 2022. [pdf] [torch]


:hammer_and_wrench: Comprehending and Ordering Semantics for Image Captioning, in CVPR. [pdf] [torch]


:hammer_and_wrench: DIFNet: Boosting Visual Information Flow for Image Captioning, in CVPR 2022. [[pdf](DIFNet: Boosting Visual Information Flow for Image Captioning)] [torch]


:hammer_and_wrench: Injecting Semantic Concepts into End-to-End Image Captioning, in CVPR 2022. [pdf] [torch]


:hammer_and_wrench: Show, Deconfound and Tell: Image Captioning with Causal Inference, in CVPR 2022. [pdf] [[torch](https: //github.com/CUMTGG/CIIC)]

主要为了解决两个Caption存在的问题
Encoder阶段(解决分类准确性的问题)





[因果关系 + 强化学习] Dependent Multi-Task Learning with Causal Intervention for Image Captioning, in IJCAI 2021. [pdf]
说实话这篇论文写作有点太复杂了,很难看懂


NOC-REK: Novel Object Captioning with Retrieved Vocabulary from External Knowledge, in CVPR 2022. [pdf]
任务描述

当前的方法
动机
方法

匹配loss的计算
为了鼓励模型引入新的类,我们将padding(和Region feature数量一致)的空类的15%随机替换成从外部知识库中随机挑选的词作为GT
计算$Hungarian$ loss (匈牙利损失) $$ \mathcal{L}{\mathrm{H}}(\mathcal{Y}, \mathcal{V})=\sum{i=1}^K-\log \operatorname{sim}\left(\mathbf{y}i, \hat{\mathbf{v}}{\hat{\sigma}(i)}\right) $$
Image Captioning with Novel Topics Guidance and Retrieval-based Topics Re-weighting, in TMM 2022. [pdf]
利用主题(Topic)模型来做image caption
什么是主题 链接
方法

NMF(NLTK中有)来从caption中提前获取的,数量为200。[主题 + Image Caption] Show, Rethink, And Tell: Image Caption Generation With Hierarchical Topic Cues, in ICME 2021. [pdf]
动机(caption本身就具备多个或者一个主题的特征)【目前觉得topic的作用就是生成更加多语义信息+生成更加精确】
方法

:hammer_and_wrench: Explicit Image Caption Editing, in ECCV 2022. [pdf] [torch]
任务:Image Caption编辑,修改得更加合理

数据集构建
方法 (模型是基于BERT)
引入编辑算子<ADD>, <DELETE>, <KEEP>


Case Study

:hammer_and_wrench: Object-Centric Unsupervised Image Captioning, in ECCV 2022. [pdf] [torch]
主要动机
贡献
如何无监督
BLEU4很低,只有6左右
:hammer_and_wrench: Human-Centric Image Captioning, in PR 2022. [pdf] [code & dataset]
动机【动机不是很强,说得不够有道理】
方法
标注数据集,已公开
人工标注了人的身体部位的boxes

背景物体和活动物体的特征区分是通过算和人之间的similarity来得到的
:hammer_and_wrench: Matching Visual Features to Hierarchical Semantic Topics for Image Paragraph Captioning, in IJCV 2022. [pdf] [torch]
引入Topic来做paragraph Caption的生成
动机
方法
end-to-end的方法**,即训练主题模型,又生成Caption,和以往的Two-stage不一样。

Case Study

:hammer_and_wrench: :fire: ClipCap: CLIP Prefix for Image Captioning, on arXiv (2021). [pdf] [torch]
模型非常简单,跑起来也很简单。一台1080 Ti 就可以训练了!
动机
$$ \max \theta \sum{i=1}^N \sum_{j=1}^{\ell} \log p_\theta\left(c_j^i \mid x^i, c_1^i, \ldots, c_{j-1}^i\right) $$
模型架构

Controllable Image Captioning via Prompting, in AAAI 2023. [pdf]
动机

方法极其简单

:hammer_and_wrench: Learning Distinct and Representative Modes for Image Captioning, in NIPS 2022. [pdf] [torch]


Incorporating Unlikely Negative Cues for Distinctive Image Captioning, in IJCAI 2023. [pdf]


:hammer_and_wrench: Transferable Decoding with Visual Entities for Zero-Shot Image Captioning, in ICCV 2023. [pdf]] [torch]



:hammer_and_wrench: With a Little Help from your own Past: Prototypical Memory Networks for Image Captioning, in ICCV 2023. [pdf] [torch]


:hammer_and_wrench: [多模态机器翻译] CLIPTrans: Transferring Visual Knowledge with Pre-trained Models for Multimodal Machine Translation, in ICCV 2023. [pdf] [torch]



:hammer_and_wrench: MultiCapCLIP: Auto-Encoding Prompts for Zero-Shot Multilingual Visual Captioning, in ACL 2023. [pdf] [torch]


:hammer_and_wrench: [视觉信息增强LM文本生成能力] Learning to Imagine: Visually-Augmented Natural Language Generation, in ACL 2023. [pdf] [torch]
⚠️事先声明:这篇论文实际上是做纯文本的,输入的是文本,输出的也是文本,只不过是想联想生成一些图片来增强模型文本生成的能力

:hammer_and_wrench: [多模态摘要生成] CFSum: A Coarse-to-Fine Contribution Network for Multimodal Summarization, in ACL 2023. [pdf] [torch]


:fire: :hammer_and_wrench: OxfordTVG-HIC: Can Machine Make Humorous Captions from Images?, in ICCV 2023. [pdf] [project]



:hammer_and_wrench: ALIP: Adaptive Language-Image Pre-training with Synthetic Caption, in ICCV 2023. [pdf] [torch]


:hammer_and_wrench: Noise-aware Learning from Web-crawled Image-Text Data for Image Captioning, in ICCV 2023. [pdf] [torch]


:fire: :hammer_and_wrench: Learning Descriptive Image Captioning via Semipermeable Maximum Likelihood Estimation, in NeurIPS 2023. [pdf] [torch]



:hammer_and_wrench: Exploring Diverse In-Context Configurations for Image Captioning, in NeurIPS 2023. [pdf] [torch]
探索In-context Learning如何做Image Caption,有点奇怪的论文🤔


:hammer_and_wrench: Dense and Aligned Captions (DAC) Promote Compositional Reasoning in VL Models, in NeurIPS 2023. [pdf]
看起来像是做Caption过滤的


:fire: :hammer_and_wrench: Caption Anything: Interactive Image Description with Diverse Multimodal Controls, 2023-05. [pdf] [torch]



:fire: :hammer_and_wrench: Attractive Storyteller: Stylized Visual Storytelling with Unpaired Text, in ACL 2023. [pdf]] [torch]


:hammer_and_wrench: :fire: Explore and Tell: Embodied Visual Captioning in 3D Environments, in ICCV 2023. [[project](https: //aim3-ruc.github.io/ExploreAndTell)]



Cycle-Consistency Learning for Captioning and Grounding, in AAAI 2024. [pdf]


:white_check_mark: :fire: :hammer_and_wrench: [TSN] Temporal Segment Networks: Towards Good Practices for Deep Action Recognition, in ECCV 2016. [pdf] [torch]
抽取所有帧是不现实的,TSN将其等间隔分为$K$个片段(i.e., $K=16$),在每个片段中谁寄抽取一帧作为输入
提供了非常常用的数据争强方式和一些训练时候的trick(主要包括location jittering, horizontal flipping, corner cropping, and scale jittering)
仍然利用双流的思路,让每个片段信息最后通过一个共识网络再Fusion

:white_check_mark: :fire: :hammer_and_wrench: [TRN] Temporal Relation Reasoning in Videos, in ECCV 2018. [pdf] [torch]

:white_check_mark: :fire: :hammer_and_wrench: [TSM] TSM: Temporal Shift Module for Efficient Video Understanding, in ICCV 2019. [pdf] [torch]
对某些通道shift,得到前一帧或者后一帧的特征

由于shift是有损失的,为此设计残差来进行弥补(原来的与残差的对比)

:white_check_mark: :fire: :hammer_and_wrench: [LRCN] Long-term Recurrent Convolutional Networks for Visual Recognition and Description, in CVPR 2015. [pdf] [torch]
LSTM得到每帧的时序特征关于视频特征抽取,下面讲一下
netvlad系列的结构,NextVlad就是专门针对视频帧融合来做的优化。
:fire: :hammer_and_wrench: [NetVLAD] NetVLAD: CNN architecture for weakly supervised place recognition, in CVPR 2016. [pdf] [torch (simple)]
VLAD算法(实际上就是Kmeans): $$ V(j, k)=\sum_{i=1}^{N} a_{k}\left(x_{i}\right)\left(x_{i}(j)-c_{k}(j)\right), \quad k \in K, j \in D $$
本文使用CNN模拟该VLAD算法的过程

:fire: :hammer_and_wrench: [NextVLAD] NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification, in ECCV workshop 2018. [pdf] [tensorflow]
:hammer_and_wrench: [PEFT&动作识别] AIM: Adapting Image Models for Efficient Video Action Recognition, in ICLR 2023. [pdf] [torch]
大模型时代下很好的尝试
Parameter Effectient Fine-tuning
动机

方法(很巧妙很简单)

:hammer_and_wrench: Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground, in ICCV 2023 Oral. [pdf] [torch]


我们使用一个十分经典的任务(Temporal Grounding)来看看视频的特征是如何利用的
:fire::hammer_and_wrench: [Video-NLP] Learning 2D Temporal Adjacent Networks for Moment Localization with Natural Language, in AAAI 2020. [pdf] [torch]

核心思想:
2D Temporal Feature Map Extraction所示Hadamard product)$$ \mathbf{F}=\left|\left(\mathbf{w}^{S} \cdot \mathbf{f}^{S} \cdot \mathbb{1}^{T}\right) \odot\left(\mathbf{W}^{M} \cdot \mathbf{F}^{M}\right)\right|_{F} $$
IoU进行一个scale变成监督信号$$ y_{i}= \begin{cases}0 & o_{i} \leq t_{\min } \ \frac{o_{i}-t_{\min }}{t_{\max }-t_{\min }} & t_{\min }<o_{i}<t_{\max } \ 1 & o_{i} \geq t_{\max }\end{cases} $$
$$ L o s s=\frac{1}{C} \sum_{i=1}^{C} y_{i} \log p_{i}+\left(1-y_{i}\right) \log \left(1-p_{i}\right) $$
:fire: :hammer_and_wrench: Negative Sample Matters: A Renaissance of Metric Learning for Temporal Grounding, in AAAI 2022. [pdf] [torch] [blog]
主干网络是沿用TDN
使用了metric learning的方法并且引入负样本来做Temporal Grounding的任务
IoU来标定监督信号yi,与2D-TAN一样处理得来的,记得scale一下)贡献
IoU来采样), 句子和视频对应的正负样本(负样本句子从别的视频抽取过来)Trick
DistilBERT来进行编码句子损失函数计算
2D-TDN一样的BCE_lossInfoNCE loss的设计对比损失$$ \begin{aligned} &p\left(i_{s} \mid v\right)=\frac{\exp \left(\left(\mathbf{f}{i}^{S T} \mathbf{f}^{V}-m\right) / \tau{v}\right)}{\exp \left(\left(\mathbf{f}{i}^{S T} \mathbf{f}^{V}-m\right) / \tau{v}\right)+\sum_{j \neq i}^{N_{s}} \exp \left(\mathbf{f}{j}^{S T} \mathbf{f}^{V} / \tau{v}\right)} \ &p\left(i_{v} \mid s\right)=\frac{\exp \left(\left(\mathbf{f}{i}^{V T} \mathbf{f}^{S}-m\right) / \tau{s}\right)}{\exp \left(\left(\mathbf{f}{i}^{V T} \mathbf{f}^{S}-m\right) / \tau{s}\right)+\sum_{j \neq i}^{N_{v}} \exp \left(\mathbf{f}{j}^{V T} \mathbf{f}^{S} / \tau{s}\right)} \ &L_{m m}=-\left(\sum_{i=1}^{N} \log p\left(i_{v} \mid s_{i}\right)+\sum_{i=1}^{N} \log p\left(i_{s} \mid v_{i}\right)\right) \end{aligned} $$
:fire: Event-Guided Procedure Planning from Instructional Videos with Text Supervision, in ICCV 2023. [pdf]



:fire: :hammer_and_wrench: Invariant Grounding for Video Question Answering, in CVPR 2022 Best Paper Finalist. [pdf] [torch]
这篇文章感觉是一篇很标准的
CVPR的中规中矩文章,写作用词上非常出色的
Casual)还有无关帧(补偿帧Complement)memory bank来存储所有样本 (因此要注意存储的特征维度不能太大)Video as Conditional Graph Hierarchy for Multi-Granular Question Answering,in AAAI 2022. [pdf] [torch]

:fire: :hammer_and_wrench: [交通事故QA数据集] SUTD-TraffificQA: A Question Answering Benchmark and an Effificient Network for Video Reasoning over Traffific Events, in CVPR 2021. [pdf] [project]

Cross-Modal Causal Relational Reasoning for Event-Level Visual Question Answering, in TPAMI 2022. [pdf]
这篇论文模型较为复杂,所以这里只讲诉其核心思想
动机
现有方法只关注了很简单的事件,比如说看电影,无法关注真正事件级的因果关系

语言和图像当中的干扰因素(Confounder)

方法

:hammer_and_wrench: :fire: Discovering Spatio-Temporal Rationales for Video Question Answering, in ICCV 2023. [pdf] [torch]


:fire: Redundancy-aware Transformer for Video Question Answering, in MM 2023. [pdf]
和上一篇论文动机很类似的,都是同一个作者,CVPR 22最佳论文候选那个作者

:fire: :hammer_and_wrench: Tem-adapter: Adapting Image-Text Pretraining for Video Question Answer, in ICCV 2023. [pdf] [torch]



:fire: :hammer_and_wrench: A-CQUIRED: A Dataset for Answering Counterfactual Questions In Real-Life Videos, in EMNLP 2023 [pdf] [[dataset&code](https: //github.com/PlusLabNLP/acquired)]


:fire: :hammer_and_wrench: Large Language Models are Temporal and Causal Reasoners for Video Question Answering, in EMNLP 2023. [pdf] [torch]

:fire: :hammer_and_wrench: LLCP: Learning Latent Causal Processes for Reasoning-based Video Question Answer, in ICLR 2024. [pdf]
说实话我不是很能看懂,假设前提有点太强了
详见原文



[Video Caption] VX2TEXT: End-to-End Learning of Video-Based Text Generation From Multimodal Inputs, in CVPR 2021. [pdf]
:hammer_and_wrench: :fire: [Video Caption] Robust Change Captioning, in ICCV 2019. [pdf] [torch]

:hammer_and_wrench: :fire: [Video Caption] Semantic Grouping Network for Video Captioning, in AAAI 2021. [pdf] [torch]

:hammer_and_wrench: :fire: Hierarchical Context-aware Network for Dense Video Event Captioning, in ACL 2021. [pdf] [torch]
Attention机制

graph LR
SG(Semantic-Grouping) --去掉冗余phrase--> 相似度计算
SG --attention机制 --> 对其phrase和frame --> 加入对比损失,计算没有包含negative的概率
对比损失$\mathcal{L}{c a}=\sum{(V, Y) \in \mathcal{D}} \sum_{t} \sum_{i}^{M_{t}}\left(-\log p_{c a}\left(s_{i, t}\right)\right)$, $p_{c a}\left(s_{i, t}\right)=\sum_{j=1}^{N} \alpha_{i, j, t}^{p o s}$ ($\alpha^{pos}$ 为正样本时候对齐注意力的权重)
:fire: :hammer_and_wrench: A New Comprehensive Benchmark for Semi-supervised Video Anomaly Detection and Anticipation, in CVPR 2023. [pdf] [torch]



:fire: Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model, in CVPR 2022. [pdf] [torch]
将CoOP(图像分类)论文思想放入OD当中
动机
模型方法(具体见论文)
连续型Prompt学习:


:fire: :hammer_and_wrench: Multi-Modal Classifiers for Open-Vocabulary Object Detection, in ICML 2023. [pdf] [code]


:fire: :star: Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond, in TACL 2022. [pdf] [blog]
:hammer_and_wrench: CauAIN: Causal Aware Interaction Network for Emotion Recognition in Conversations, in IJCAI 2022. [pdf] [torch]
在对话当中找情感相关的因果线索
动机:

方法
利用 ATOMIC 常识知识语料库,获得对话历史每句的 6 种因果线索

利用因果线索,建模对话历史

:hammer_and_wrench: Knowledge-Bridged Causal Interaction Network for Causal Emotion Entailment, in AAAI 2023. [pdf] [torch]
这份工作和上面👆的IJCAI 2022:CauAIN: Causal Aware Interaction Network for Emotion Recognition in Conversations那个工作idea是一样的,只不过做了两个不同的任务而已
主要动机都是找寻对话当中的因果线索

方法模型


:fire: :hammer_and_wrench: Everything Has a Cause: Leveraging Causal Inference in Legal Text Analysis, in NAACL 2021. [pdf] [torch]


:hammer_and_wrench: :fire: Deconfounded Video Moment Retrieval with Causal Intervention, in SIGIR 2021. [pdf] [torch]
动机
方法


:fire: :hammer_and_wrench: Interventional Video Grounding with Dual Contrastive Learning, in CVPR 2021. [pdf] [torch] (2023年1月仍为开源)
动机

方法

因果推理模块(通过loss来调整)


:fire: :hammer_and_wrench: Two Causal Principles for Improving Visual Dialog, in CVPR 2020. [pdf] [torch] [zhihu]
含老师团队第一篇因果推理方向的文章
一句话总结:
- 从因果图角度审视视觉对话任务,切断对话历史与答案的直接因果效应,添加混杂因子[用户偏好]构建描述真实世界的因果图。
动机
两个原则
$P_1$: H对话历史记录,不应该直接影响A

$P_2$: 用户偏好会导致回答的不一样

由于用户偏好不可观测,因此作者通过三种方式来近似采样。

构建两个原则的因果图:

具体后门调整及其概率运算公式,详见论文及其附录


:fire: :hammer_and_wrench: [因果关系] Visual Commonsense R-CNN, in CVPR 2020. [pdf] [torch] [blog]
出自MReal, 张含望老师团队的工作,非常Solid的一篇工作
- 目标是训练基于
Faster-RCNN训练一个更强的feature extractor可以捕获视觉上的常识信息。- 这篇论文实在太多细节和推理了,建议看我自己的GoodNote上的笔记!
动机
现在的模型无法学习到视觉常识(Commonsense):人和椅子 -> 人可以坐在椅子上。但在NLP中,常识的信息已经放在特征里面了

数据集的偏差会导致无法捕捉到常识信息
因果理论就是用来发现==现象背后的不变规律==的,是一种鲁棒的预测。这与常识本身不就很相似吗,我们人类也是从生活中不断总结积累这些不变的、鲁棒的经验或者因果规律,并把他们叫做常识。 比如,看见凳子知道可以坐,看见pizza知道可以吃。
Association 和 Intervention(分层)的计算 $$ \begin{gathered} P(Y \mid X)=\sum_z P(Y \mid X, z) P(z \mid X)=\frac{P(Y, X)}{P(X)} \ P(Y \mid d o(X))=\sum_z P(Y \mid X, z) P(z)=\sum_z \frac{P(Y, X, z) P(z)}{P(X, z)} \end{gathered} $$ 其中 $X, Y, z$分别代表了图片中的object label,同时这里我们用物体出现的频率来代替概率,比如 $P(Sink|Hair drier)$就是用“含有$Sink$和$Hair drier$两者的图片数”比上“只含有Hair drier的图片数”计算得到的。画出两者计算结果差异的对比图(只标明了20类):

两个Case的分析
方法(因果干预Intervention)
代理任务(无监督学习):给定RoI X的feature去预测RoI Y的类别
包括很多潜在的混杂因子,如果直接预测周围物体Y就不可避免的会被上文提到的混杂因子confounder所影响。根据我们刚刚介绍的**“do算子”的理论,解决的办法也不难,只要能找到confounder然后对他们使用backdoor理论**进行控制即可。
混杂因子是什么? 我们直接把整个数据集上的object RoI特征(Faster RCNN中来)在每个类别上取平均,当作这个类别的表示,进而构建出一个 类别数x1024 的confounder字典作为$Z$(比如MSCOCO有80类,就是 80x1024),它包含着所有可能的混杂因子。
后门调整

模型

注意:VC R-CNN的实现和原先的Faster R-CNN相比,去除了RPN网络(Region Proposal Network),不再训练网络propose边界框,而是直接将数据集ground-truth的bounding box坐标输入到其中,直接提取region的特征。而在训练完成后的feature提取阶段,相对应的,只要给定图片和bounding box坐标,都可以获得对应的VC特征。就这样,我们利用bottomup特征已有的边界框坐标提取VC特征后,将其并在先前的bottomup特征上作为新的特征。我们在传统的 Vision&Language 三大任务上挑选了经典model和SOTA model进行了测试,发现在各个任务上都取得了明显的提升,尤其是在image captioning上的提升尤其大。同时为了验证性能的提升不是由于参数增多带来的,我们还在原有特征上并上了ablative的特征(单独object特征,用correlation计算的特征),具体可以参考论文的实验部分。
:hammer_and_wrench: [指代表达] Deconfounded Visual Grounding, in AAAI 2022. [pdf] [torch] (2023.1.17未开源)
动机

方法
构建因果图

R语言信息,因为其实视觉上的bias其他一些方向论文都有所解决了由于$G$不可观测,因此需要近似采样

:fire: :hammer_and_wrench: Unbiased Scene Graph Generation from Biased Training, in CVPR 2020. [pdf] [torch] [zhihu]
动机
贡献

问题定义(有偏见的数据标注的原因)
无偏的思想(content:内因,context:外因)
人类在有偏见的大自然中生长,在拥抱好的context的同时,避免不好的context,并与content一起做出无偏见的决定。
其潜在的机制是基于因果关系的(causality-based):决策是通过追求由content引起的主要因果效应,而不是追求由context引起的副作用来做出的。然而,机器是基于可能性的(likelihood-based),会产生有偏结果。
故论文认为,无偏预测的关键是教会机器如何区分主要作用(main effect)和副作用
为了在无偏预测中追求主要作用,论文提出赋予机器反事实思维(counterfactual thinking): If i had not seen the content, would I still make the same prediction?
反事实思维:事实与反事实之间的比较,将会自然地消除context偏差的影响,因为context是两者之间唯一不变的东西。
如图,左侧图片是所谓的事实场景,也可以说是原始场景;右侧图片是反事实场景,就是**将原始场景中content(狗和冲浪板的视觉特征)去除,其他部分(如scene和object classes)保持不变,就像object的视觉特征从未出现过。**通过这两者的比较,我们可以专注于关系的主要视觉影响,同时也不丢失context。

方法


TDE(Total Direct Effect)方法没有引入任何额外的参数,也可以说没有针对模型的有偏训练进行任何改动,其使用原始SGG模型进行了两次预测,将两次预测的结果进行差值运算,最终得到无偏见的预测。所以TDE方法是模型“不可见”的,广泛适用于各种SGG模型。
:fire: :hammer_and_wrench: Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect, in NIPS 2020. [pdf] [torch] [zhihu]
动机
核心思想
需要利用原始的长尾分布来学习特征提取的原因在于,大量的尾部类别其实不足以提供足够的样本来学习鲁棒的特征表达。人类描述罕见的物体时,往往是通过和已知常见类的比较,比如会说狮鹫是有着狮子的身体,鹰的翅膀和头的生物,而不必要单独拿一堆狮鹫的图片出来,让你死记硬背住狮鹫的长相。

发现优化器的动量项时,这货不就是在训练数据时引入数据分布,从而产生shortcut的元凶么。
方法,详细的推到有点复杂,相近论文或者知乎

方法速成步骤

:fire: :hammer_and_wrench: Interventional Few-Shot Learning, in NIPS 2020. [pdf] [torch] [blog]
博客讲得很全面,具体可以见博客对本文的讲解
- 方法挺会玩的!
动机


问题建模(探索Many-shot Learning 和 Few-shot在因果关系上的不一致地方)


$$ P(Y \mid d o(X=\boldsymbol{x}))=\sum_d P(Y \mid X=\boldsymbol{x}, D=d, C=g(\mathbf{x}, d)) P(D=d) $$

:fire: :hammer_and_wrench: Introspective Distillation for Robust Question Answering, in NIPS 2021. [pdf] [torch] [zhihu]

:fire: :hammer_and_wrench: Distilling Causal Effect of Data in Class-Incremental Learning, in CVPR 2021. [pdf] [torch]
动机
对撞节点的存在使得模型对新数据会产生灾难性遗忘
过去的方法当中


思路
文章细节详见开头的博客链接
:fire: :hammer_and_wrench: Causal Attention for Vision-Language Tasks, in CVPR 2021. [pdf] [torch] [Blog]
我的想法和博客最后说的类似,似乎本文推理推得不是特别彻底,但是仍然是一篇优秀的论文
动机:解决当前VL模型当中训练数据集当中bias

思想

方法(前门调整机制做Causal Attention,此处不做展开讨论,详见论文)


结果case

:fire: :hammer_and_wrench: Counterfactual Zero-Shot and Open-Set Visual Recognition, in CVPR 2021. [pdf] [torch] [zhihu]
生成式的因果模型,博客很详细,详见知乎!
方法
样本特征和类别特征之间解耦


:fire: :hammer_and_wrench: Counterfactual VQA: A Cause-Effect Look at Language Bias, in CVPR 2021. [pdf] [torch]
动机
方法(基于ensemble的VQA模型做的一个推理框架,相当简单)


因果推理角度解释(具体见论文)

Learning Causal Effects on Hypergraphs, Best Paper of KDD 2022. [pdf]
Entropic Causal Inference: Graph Identifiability, in ICML 2022. [pdf]
新的因果推断架构,可以不借助intervention
理论性很强的一篇文章
熵因果推断:
通过寻找数据的信息-理论上最简单的结构解释,即最小熵模型,从观测数据中学习两个变量之间的因果图。
在这个工作中,首先推广了松弛假设下的因果图可辨识性结果。
然后,我们展示了第一个可识别的结果,使用熵的方法学习超过两个节点的因果图。

:hammer_and_wrench: [2023.01未开源] Disentangle and Remerge: Interventional Knowledge Distillation for Few-Shot Object Detection from A Conditional Causal Perspective, in AAAI 2023. [pdf] [torch]
本文方法上和实验上还有很多亮点,这里主要就说因果推理


Learning to Imagine: Integrating Counterfactual Thinking in Neural Discrete Reasoning, in ACL 2022. [pdf]
动机

方法

:hammer_and_wrench: :star2: Causality Inspired Representation Learning for Domain Generalization, in CVPR 2022 Oral. [pdf] [torch]
傅立叶变换结合因果推理,超级solid的工作
动机
方法
前提:傅立叶变换当中

核心思想与流程概览,具体见论文

$$ \min {\hat{g}, \hat{h}{1}, \hat{h}{2}} \mathcal{L}{c l s}^{s u p}+\mathcal{L}{c l s}^{i n f}+\tau \mathcal{L}{F a c}, \quad \min {\hat{w}} \mathcal{L}{c l s}^{s u p}-\mathcal{L}_{c l s}^{i n f}, $$
:fire: [CLIP] EI-CLIP: Entity-aware Interventional Contrastive Learning for E-commerce Cross-modal Retrieval, in CVPR 2022. [pdf]
动机


方法

[领域泛化] GCISG: Guided Causal Invariant Learning for Improved Syn-to-real Generalization, in ECCV 2022. [pdf]
个人认为这篇论文写得有点夸张了,实际上就那么回事。
动机
因果发现

方法(很简单)

:hammer_and_wrench: Causality-aware Concept Extraction based on Knowledge-guided Prompting, in ACL 2023. [pdf] [torch]
方法极其简单
动机


因果分析

方法(极其简单,加多一个分类出来的Topic)

:hammer_and_wrench: COLA: Contextualized Commonsense Causal Reasoning from the Causal Inference Perspective, in ACL 2023. [pdf] [[torch](https://github. com/HKUST- KnowComp/COLA)]
和2022年那篇ICML论文ROCK做的任务是一样的,就是纯文本当中的因果事件检测
动机

方法(理论依据详见论文)

:fire: CFL: Causally Fair Language Models Through Token-level Attribute Controlled Generation, in ACL Finding 2023. [pdf]
有点晦涩写的,不是很看得懂


模型方法


:fire: :hammer_and_wrench: Causal-Debias: Unifying Debiasing in Pretrained Language Models and Fine-tuning via Causal Invariant Learning, in ACL 2023. [pdf] [torch]
通过学习因果不变性,来缓解PLM在fine-tune时候的bias问题


$do(N=n)$求invariant loss
$$
\min \mathcal{L}_{\text {invariant }}=\mathbb{E}_n(\mathcal{R})+\operatorname{Var}_n(\mathcal{R})
$$

:fire: :hammer_and_wrench: Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference, in ACL 2023 Oral. [pdf] [torch]



[VQA] Reducing Vision-Answer Biases for Multiple-Choice VQA, in TIP 2023. [pdf]


Causal Intervention and Counterfactual Reasoning for Multi-modal Fake News Detection, in ACL 2023. [pdf]



:hammer_and_wrench: :fire: A Causal Framework to Quantify the Robustness of Mathematical Reasoning with Language Models, in ACL 2023. [pdf] [code]
非常Perfect的关于因果推理 + LLM + 数学推理的Empirical Study


:hammer_and_wrench: Random Boxes Are Open-world Object Detectors, in CVPR 2023. [pdf] [torch]





:fire: :hammer_and_wrench: Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning, in ICCV 2023. [pdf] [torch]



:fire: Variational Causal Inference Network for Explanatory Visual Question Answering, in ICCV 2023. [pdf]
本文公式推导较为复杂,具体可以见论文




:fire: A Multi-modal Debiasing Model with Dynamical Constraint for Robust Visual Question Answering, in ACL Findings 2023. [pdf]



:fire: Neuro-Symbolic Procedural Planning with Commonsense Prompting, in ICLR 2023. [pdf]
A Causal Inference Look at Unsupervised Video Anomaly Detection, in AAAI 2022. [pdf]

:hammer_and_wrench: Deconfounded Multimodal Learning for Spatio-temporal Video Grounding, in MM 2023. [pdf] [torch]



Understanding Chat Messages for Sticker Recommendation in Messaging Apps, in AAAI 2020. [pdf]

:hammer_and_wrench: :fire: Learning to Respond with Stickers: A Framework of Unifying Multi-Modality in Multi-Turn Dialog, in WWW 2020. [pdf] [torch]


SER30K: A Large-Scale Dataset for Sticker Emotion Recognition, in MM 2022 Oral. [pdf] [torch]


数据集特点(sticker emotion recognition)

方法(Baseline)


:hammer_and_wrench: :fire: Selecting Stickers in Open-Domain Dialogue through Multitask Learning, in ACL Finding 2022. [pdf] [torch]


Named Entity Recognition
当前竞赛NER任务的baseline:
:fire: Bidirectional LSTM-CRF Models for Sequence Tagging, in 2015. [pdf] [code
:hammer_and_wrench: :fire: Fast and Accurate Entity Recognition with Iterated Dilated Convolutions, in EMNLP 2017. [pdf] [tensorflow]

核心思想

Bi-LSTM-CRF快了非常多,而且精度没有下滑:hammer_and_wrench: :fire: BOND: BERT-Assisted Open-Domain Named Entity Recognition with Distant Supervision , in KDD 2020. [pdf] [torch]
远距离监督的问题
trade-off 在标注准确度和覆盖范围之间想法
第一阶段使用RoBERTa微调,适应NER任务
Early stopping 方法防止数据过拟合还有对未知数据增强泛化能力POS识别潜在实体,然后通过语料库计算最小损失确定实体
第二阶段自我学习框架 (teacher-student模型2)

两阶段的BOND框架
:hammer_and_wrench: :fire: A Boundary-aware Neural Model for Nested Named Entity Recognition , in EMNLP 2019. [pdf] [torch]
Nested NER的问题:hammer_and_wrench: :fire: Cross-Domain NER using Cross-Domain Language , in ACL 2020. [pdf] [torch]

最底下的一层是数据层,标准情况一下总共有四份语料,分别对应两个domain下的两个task(NER和语言建模)。其中Source Domain(即保证有标记数据用于NER的domain)对应之前提到的News Domain,因为论文中Source Domain使用的是新闻数据。另外如果是无监督抽取Target Domain数据则只有三份语料。
由底向上第二层是Word Embedding层,论文中的Word Embedding结合了词级别和字符级别的向量表示。即把词向量和一个词的字符序列形成的矩阵经过CNN处理后的向量concatenate起来。
第三层是双向LSTM,用于序列处理第二层的数据,生成前后向hidden state。
第四层LM和CRF,即task model层。我们可以看到第四层有三个模块,两个是用于NER的CRF模型,分别对应Source Domain和Target Domain。另一个是基于第三层BiLSTM的语言模型,NSSoftmax是指这个语言模型利用Negative Sampling Softmax的方式进行训练。
Bi-LSTM的参数是生成的,不同domain的不同task需要的LSTM的参数和$I$有关$$ \begin{equation} \theta_{\mathrm{LSTM}}^{d, t}=\mathbf{W} \otimes \mathbf{I}{d}^{D} \otimes \mathbf{I}{t}^{T}, \end{equation} $$
NER的未来
既然模型打不动了,然后我找了找 ACL2020做NER的论文,看看现在的NER还在做哪些事情,主要分几个方面
作者:王岳王院长 链接:https://zhuanlan.zhihu.com/p/166496466 来源:知乎 著作权归作者所有。商业转载请联系作者获得授权,非商业转载请注明出处。
:fire: :hammer_and_wrench: Knowledgeable Prompt-tuning: Incorporating Knowledge into Prompt Verbalizer for Text Classifification, in ACL 2022. [pdf] [torch]] [知乎博客]
将外部知识融入当prompt-tuning当中做TC任务

:fire: [层次化主题] Deep Latent Dirichlet Allocation with Topic-Layer-Adaptive Stochastic Gradient Riemannian MCMC, in ICML 2017. [pdf]
理论性非常强的论文
- 详细的理论推导请见原文。

:fire: :hammer_and_wrench: Quark: Controllable Text Generation with Reinforced [Un]learning, in NIPS 2022. [pdf] [torch]
InstructGPT很类似,但是不再训练奖励模型,而是使用prompt来取代(评判生成的是否是好句子)
:fire: :hammer_and_wrench: Large Language Models Are Human-Level Prompt Engineers, in ICLR 2023. [pdf] [torch] [zhihu]

:fire: [2023.5.6未开源] SCOTT: Self-Consistent Chain-of-Thought Distillation, in ACL 2023. [pdf]
动机

方法【teacher:GPT-neox20B student:T5 3B】

细节


:fire: Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes
方法也太简单了,其实就是当前LLM的范式,用大语言模型生成的东西来蒸馏学习小模型

:fire: Poisoning Language Models During Instruction Tuning, in ICML 2023. [pdf] [torch]
动机

:fire: :hammer_and_wrench: Outline, Then Details: Syntactically Guided Coarse-To-Fine Code Generation, in ICML 2023. [pdf] [torch]
动机
最大贡献
方法

$S3$和$S4$是重点,tokenizer的encode伪代码如下:

模型(预训练+微调)

:hammer_and_wrench: Distilling Script Knowledge from Large Language Models for Constrained Language Planning, in ACL 2023. [pdf] [torch]


:fire: :hammer_and_wrench: BadPrompt: Backdoor Attacks on Continuous Prompts, in NIPS 2022. [pdf] [torch]


:fire: :hammer_and_wrench: RL4F: Generating Natural Language Feedback with Reinforcement Learning for Repairing Model Outputs, in ACL 2023. [pdf] [torch]
太棒的工作了!

T5-large (0.77M)】
:fire: :hammer_and_wrench: Are You Copying My Model? Protecting the Copyright of Large Language Models for EaaS via Backdoor Watermark, in ACL 2023. [pdf]] [[torch](https: //github.com/yjw1029/EmbMarker)]


:hammer_and_wrench: :fire: Tree of Thoughts: Deliberate Problem Solving with Large Language Models, in 2023 05. [pdf] [torch]




:hammer_and_wrench: Plug-and-Play Knowledge Injection for Pre-trained Language Models, in ACL 2023. [pdf] [[torch](https://github.com/ THUNLP/Knowledge-Plugin)]

:hammer_and_wrench: :fire: Mixture-of-Domain-Adapters: Decoupling and Injecting Domain Knowledge to Pre-trained Language Models’ Memories, in ACL 2023. [pdf] [[torch](https://github.com/ Amano-Aki/Mixture-of-Domain-Adapters)]

:fire: :hammer_and_wrench: Grounding Language Models to Images for Multimodal Inputs and Outputs, in ICML 2023. [pdf] [torch]
单卡A6000进行训练,很友好



:hammer_and_wrench: [COT] Element-aware Summarization with Large Language Models: Expert-aligned Evaluation and Chain-of-Thought Method, in ACL 2023. [pdf] [code]
相当于重构了现在的Summarization数据集


:fire: :hammer_and_wrench: Synthetic Text Generation with Differential Privacy: A Simple and Practical Recipe, in ACL 2023. [pdf] [torch]


:fire: :hammer_and_wrench: Editing Large Language Models: Problems, Methods, and Opportunities, in 2023.05. [pdf] [torch]

:fire: :hammer_and_wrench: LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models, in ICCV 2023. [pdf] [torch]


:fire: :hammer_and_wrench: CoTDet: Affordance Knowledge Prompting for Task Driven Object Detection, in ICCV 2023. [pdf] [torch]


:hammer_and_wrench: :fire: K2: A Foundation Language Model for Geoscience Knowledge Understanding and Utilization, in WSDM 2024. [pdf] [dataset&code]

:hammer_and_wrench: :fire: Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning, in EMNLP 2023 Best Paper. [pdf] [[torch](https://github.com/lancopku/ label-words-are-anchors)] [zhihu]

✨ 标签词作为锚的信息流$H_1$ :在LLM的浅层,标签词聚集示例具体内容的信息,为深层形成语义表示。 $H_2$:在LLM的深层,模型从聚集后的标签词位置中提取信息,形成最终预测。


Q1: Analysis of different ICL formats on the final prediction like random labels, reversed labels (e.g.,True->False, False->True), and label agencies (replace labels with meaningless characters) may be helpful.
缺少对于不同的 ICL 格式对最终预测的影响的分析,如随机标签、反向标签(如真->假、假->真)和标签机构(用无意义字符替换标签),可能会有所帮助。
分析:粗略的看,本文的结论和之前的工作得到的结论“ICL中的标签正确性对于结果似乎影响不大”[1] 可能有矛盾之处。如果本文结论合理,在深层,ICL机制是进行对label words处的聚合,那在示例中标签是错误的时候,label words处聚合到的信息(或者说聚合到的文本到标签的映射关系信息)有可能是有误导性的,进而导致性能下降;但是[1]中实验表明“ICL中的标签正确性对于结果似乎影响不大”,似乎矛盾。
对此可能合理的解释是:深层聚合时,不止参考了label words处的聚合信息,还参考了待预测样本的本身信息;对于能力较强的LLM(如gpt-3.5),有一定排除前面聚合的误导信息的能力,而只参考前面的格式;对于能力较弱的(如本文用的gpt2-xl,gpt-j-6b),这种“排除误导信息”的能力较弱,因此会表现出“icl中给随机标签得到的结果和随即猜测差不多”(也就是作者在rebuttal回复的)。综上,本文所分析的icl机制也许还存在一些待补全的部分,即对于待预测样本本身文本信息的参考程度。
[1] EMNLP 2022. Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?
:hammer_and_wrench: :fire: Can Large Language Models Infer Causation from Correlation?, in ICLR 2024. [pdf]] [dataset] [torch]




:fire: :hammer_and_wrench: Knowledge Rumination for Pre-trained Language Models, in EMNLP 2023. [pdf] [torch] [zhihu]
具体见论文,是知识编辑方面的工作


:fire: :hammer_and_wrench: When Do Program-of-Thought Works for Reasoning?, in AAAI 2024. [pdf] [torch]


:fire: :hammer_and_wrench: INSTRUCTSCORE: Explainable Text Generation Evaluation with Fine-grained Feedback, in EMNLP 2023. [pdf] [torch]



:fire: :hammer_and_wrench: GENOME: GenerativE Neuro-symbOlic visual reasoning by growing and reusing ModulEs, in ICLR 2024. [pdf] [project]


:fire: :hammer_and_wrench: Learning From Mistakes Makses LLM Better Reasoner, in 2024-02. [pdf] [torch]




:fire: :hammer_and_wrench: Safe RLHF: Safe Reinforcement Learning from Human Feedback, in ICLR 2024. [pdf] [torch] [zhihu]
理论和代码可以见知乎博客,很详细

:fire: :hammer_and_wrench: Digital Socrates: Evaluating LLMs through Explanation Critiques, in ACL 2024. [[project](https://allenai.org/data/ digital-socrates)]

:fire: :hammer_and_wrench: Navigating Dataset Documentations in AI: A Large-Scale Analysis of Dataset Cards on Hugging Face, in ICLR 2024. [project]

:fire: :hammer_and_wrench: What's documented in AI? Systematic Analysis of 32K AI Model Cards, in Nature Machine Intelligence 2024. [pdf] [Code]
:hammer_and_wrench: :fire: ModelGPT: Unleashing LLM’s Capabilities for Tailored Model Generation, in 2024-02. [project]


:hammer_and_wrench: :fire: What is the Best Way for ChatGPT to Translate Poetry?, in ACL 2024. [project]

:fire: :hammer_and_wrench: INCHARACTER: Evaluating Personality Fidelity in Role-Playing Agents through Psychological Interviews, in ACL 2024. [project]


:fire: :hammer_and_wrench: Generating Summaries with Controllable Readability Levels, in EMNLP 2023. [pdf] [[torch](https://github.com/amazon-science/ controllable-readability-summarization)]


:fire: Dr.Academy: A Benchmark for Evaluating Questioning Capability in Education for Large Language Models, in ACL 2024. [pdf]


:hammer_and_wrench: OceanGPT: A Large Language Model for Ocean Science Tasks, in ACL 2024. [project]



:fire: :hammer_and_wrench: Aligning Large Language Models with Human Preferences through Representation Engineering, in ACL 2024. [pdf] [torch]


:fire: :hammer_and_wrench: LexEval: A Comprehensive Chinese Legal Benchmark for Evaluating Large Language Models, in NeurIPS 2024. [project]

:hammer_and_wrench: AST-Trans: Code Summarization with Efficient Tree-Structured Attention, in ICSE 2022. [pdf] [torch]
代码摘要的生成

方法

实验
:fire: :hammer_and_wrench: Multi-target Backdoor Attacks for Code Pre-trained Models, in ACL 2023. [pdf] [torch]
代码仓库很完善,但这篇论文写作一般般吧

:hammer_and_wrench: EGFE: End-to-end Grouping of Fragmented Elements in UI Designs with Multimodal Learning, in ICSE 2024. [pdf] [Dataset&Code]
但这个数据集好像只是组合元素,不是生成前端代码



:fire: Android in the Wild: A Large-Scale Dataset for Android Device Control, in NeurIPS 2023. [pdf] [[dataset](https://github.com/google-research/google-research/ tree/master/android_in_the_wild)]

:hammer_and_wrench: Pairwise GUI Dataset Construction Between Android Phones and Tablets, in NeurIPS 2023. [pdf] [dataset]


多模态预训练
:fire: :hammer_and_wrench: [Cross-Modal&Contrastive Learning] UNIMO: Towards Unified-Modal Understanding and Generation via Cross-Modal Contrastive Learning, in ACL(long paper) 2021. [pdf] [project from Baidu]
:hammer_and_wrench: [MultiModal] UniT: Multimodal Multitask Learning with a Unified Transformer, ICCV 2021. [pdf] [project from Fair]
:fire: :hammer_and_wrench: Leveraging Visual Knowledge in Language Tasks: An Empirical Study on Intermediate Pre-training for Cross-modal Knowledge Transfer, in ACL 2022. [pdf] [torch (2022.12.01未开源)]
本文是一篇实验性的文章,实验的方法写得不错!
什么叫做:intermediate pre-training?
- 在预训练好的模型上,补充一些数据集或者语料库继续预训练
动机

方法(以下两种方法来弥补PLM在这方面的缺陷)
text knowledge transfer: 即使用image caption进行知识迁移,设计了如下两个训练目标:
cross-modal knowledge transfer: 即使用图片和文本以及V&L的训练来进行知识迁移,设计了如下几个训练目标:
Voken Classification: voken采用token层面的text2image检索来迁移视觉知识,它假设每一个token都有一个视觉域的voken与其对应,训练目标就是在所有预先设定好的voken中将正确的voken检索出来
Masked Language Modeling with Visual Clues: 给定图片作为线索,预测mask掉的token,比MLM多了图片作为输入,目标函数是一样的
Cross-Modal Contrastive Learning (CMCL): 和CLIP一样(不同在于这里把视觉网络给冻住),是跨模态的对比学习

Cross-Modal Knowledge Distillation (CMKD): 将在MSCOCO数据集上进行对比学习的多模态模型作为teacher model,将一个语言模型作为student,在纯文本语料Wiki103上进行知识蒸馏
模型示意图

结论
:hammer_and_wrench: :fire:[Relation CLIP] RelCLIP: Adapting Language-Image Pretraining for Visual Relationship Detection via Relational Contrastive Learning, in EMNLP 2022. [pdf] [torch]
动机
CLIP可以识别对象之间的relation方法(很简单)
Commonsense Knowledge 就是Conceptual Caption数据中来

负样本收集:

:star: [MM-KG + CLIP] Contrastive Language-Image Pre-Training with Knowledge Graphs, in NIPS 2022. [pdf] [rebuttal]
动机


:hammer_and_wrench: PuMer: Pruning and Merging Tokens for Efficient Vision Language Models, in ACL 2023. [pdf] [[torch](https://github.com/ csarron/PuMer)]

:fire: Learning to Estimate Shapley Values with Vision Transformers, in ICLR 2023. [pdf] [torch]
很棒的一篇Shapley Values 结合 ViT的论文

:fire: Knowledge-Aware Prompt Tuning for Generalizable Vision-Language Models, in ICCV 2023. [pdf]
动机
传统的CLIP对于一些unseen class 泛化能力不是很好
CLIP + 外部知识 (个人觉得 CVPR的那几篇更好)


:hammer_and_wrench: [多语言CLIP] mCLIP: Multilingual CLIP via Cross-lingual Transfer, in ACL 2023. [pdf] [[torch](https://github.com/ huawei-noah/noah-research/NLP/mclip)]


:hammer_and_wrench: Rethinking Multimodal Entity and Relation Extraction from a Translation Point of View, in ACL 2023. [pdf] [torch]




:fire: :hammer_and_wrench: Preserving Modality Structure Improves Multi-Modal Learning, in ICCV 2023. [pdf] [[torch](https://github. com/Swetha5/Multi_Sinkhorn_Knopp)]

:hammer_and_wrench: [复杂的图像检索任务] A Neural Divide-and-Conquer Reasoning Framework for Image Retrieval from Linguistically Complex Text, in ACL 2023. [pdf] [torch]




:hammer_and_wrench: :fire: Bayesian Prompt Learning for Image-Language Model Generalization, in ICCV 2023. [pdf] [[torch](https://github.com/saic-fi/Bayesian- Prompt-Learning)]


:fire: :hammer_and_wrench: A Unified Objective for Novel Class Discovery, in ICCV 2021 Oral. [pdf] [torch] [blog]

:fire: :hammer_and_wrench: E2VPT: An Effective and Efficient Approach for Visual Prompt Tuning, in ICCV 2023. [pdf] [torch]

:fire: SINC: Self-Supervised In-Context Learning for Vision-Language Tasks, in ICCV 2023. [pdf]



:fire: :hammer_and_wrench: Make the U in UDA Matter: Invariant Consistency Learning for Unsupervised Domain Adaptation, in NeurIPS 2023. [pdf] [torch]


:hammer_and_wrench: ViStruct: Visual Structural Knowledge Extraction via Curriculum Guided Code-Vision Representation, in EMNLP 2023. [pdf] [[torch](https://github.com/ Yangyi-Chen/vi-struct)]



:hammer_and_wrench: :fire: DDCoT: Duty-Distinct Chain-of-Thought Prompting for Multimodal Reasoning in Language Models, in NeurIPS 2023. [project]


:fire: :hammer_and_wrench: See and Think: Embodied Agent in Virtual Environment, 2023-12. [project]


:hammer_and_wrench: [表情包] MEMECAP: A Dataset for Captioning and Interpreting Memes, in EMNLP 2023. [pdf] [torch]




:fire: :hammer_and_wrench: CLOVA: A Closed-LOop Visual Assistant with Tool Usage and Update, in 2023-12. [pdf] [torch] [知乎]

:hammer_and_wrench: :fire: [2024-01 刚放榜,未开源] ADAPTING TO DISTRIBUTION SHIFT BY VISUAL DOMAIN PROMPT GENERATION, in ICLR 2024. [pdf]


:hammer_and_wrench: DOC2PPT: Automatic Presentation Slides Generation from Scientific Documents, in AAAI 2022. [pdf] [torch]



:fire: :hammer_and_wrench: VLIS: Unimodal Language Models Guide Multimodal Language Generation, in EMNLP 2023. [pdf] [torch]


:fire: Multitask Multimodal Prompted Training for Interactive Embodied Task Completion, in EMNLP 2023. [pdf] [torch]

:fire: :hammer_and_wrench: :star2: Fine-tuning Multimodal LLMs to Follow Zero-shot Demonstrative Instructions, in ICLR 2024. [pdf] [torch]





:fire: :hammer_and_wrench: UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model, in EMNLP 2023. [pdf] [torch]] [zhihu]
依托达摩院mPLUG系列大模型的



:fire: :hammer_and_wrench: mPLUG-PaperOwl: Scientific Diagram Analysis with the Multimodal Large Language Model, in 2023-11. [pdf] [[torch](https://github.com/X-PLUG/mPLUG-DocOwl/ tree/main/PaperOwl)]


:fire: :hammer_and_wrench: Too Large; Data Reduction for Vision-Language Pre-Training, in ICCV 2023. [pdf] [code]

:hammer_and_wrench: :fire: MathVista Evaluating Math Reasoning in Visual Contexts, in ICLR 2024 Oral. [project]
动机
表现


:fire: :hammer_and_wrench: BLIVA: A Simple Multimodal LLM for Better Handling of Text-Rich Visual, in AAAI 2024. [pdf] [torch]


:fire: :hammer_and_wrench: [2024-01 未开源] Imagine That! Abstract-to-Intricate Text-to-Image Synthesis with Scene Graph Hallucination Diffusion, in NeurIPS 2023. [pdf] [torch]


:fire: :hammer_and_wrench: Generating Explanations for Embodied Action Decision from Visual Observation, in MM 2023. [pdf]


:fire: :hammer_and_wrench: InstructDoc: A Dataset for Zero-Shot Generalization of Visual Document Understanding with Instructions, in AAAI 2024. [pdf] [[torch](https://github. com/nttmdlab-nlp/InstructDoc)]



:fire: :hammer_and_wrench: Improving CLIP Training with Language Rewrites, in NeurIPS 2023. [pdf] [torch]
用ChatGPT改写CLIP文本端的Prompt

:fire: :hammer_and_wrench: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation, in 2023-12. [pdf] [torch]
动机
模型方法(单卡4090~2.5天就可以收敛)


:fire: :hammer_and_wrench: FROSTER: Frozen CLIP is A Strong Teacher for Open-Vocabulary Action Recognition, in ICLR 2024. [pdf]


:fire: :hammer_and_wrench: A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation, in ICLR 2024. [pdf]



:fire: :hammer_and_wrench: Small Language Model Meets with Reinforced Vision Vocabulary, in 2024-02. [pdf] [torch]

:fire: :hammer_and_wrench: MoE-LLaVA: Mixture of Experts for Large Vision-Language Models, in 2024-02. [pdf] [torch]


:fire: :hammer_and_wrench: Cheap and Quick: Efficient Vision-Language Instruction Tuning for Large Language Models, in NeurIPS 2023. [pdf] [project] [[INT4量
Truncated — view the full README on GitHub.
141 commits