文本摘要的一些研究概念
主要翻译自github,特别适合新手查看。我也是新手,看下面的翻译就知道了。
生成方式(Generation Way)
- gen-ext:提取式摘要?第一个就遇到了困难。
- gen-abs:抽象式摘要?
- gen-2stage:两个混合,压缩和混合
回归方式(Regressive Way)
regr-auto
: Autoregressive Decoder (Pointer network) 自回归解码器,指针网络regr-nonauto
: Non-autoregressive Decoder (Sequence labeling) 非自回归解码器,序列标签
任务设定(Task Settings)
task-singleDoc
: Single-document Summarization 单文本摘要task-multiDoc
: Multi-document Summarization 多文本摘要task-senCompre:
Sentence Compression 句子压缩task-sci
: Scientific Paper 科技论文task-radiologyReport
: Radiology Reports 放射科报告??这玩意怎么跑到这的?task-multimodal
: Multi-modal Summarization 多模型摘要/多模型汇总task-aspect
: Aspect-based Summarization 基于方面的摘要???task-opinion
: Opinion Summarization 可选择摘要task-review
: Review Summarization 摘要综述???task-meeting
: Meeting-based Summarization 基于会议的摘要task-conversation
: Consersation-based Summarization 基于会话的摘要task-medical
: Medical text-related Summarization 关于医学文本摘要task-covid
: COVID-19 related Summarization 关于新冠病毒的摘要task-query
: query-based Summarization 基于查询的摘要task-question
: question-based Summarization 基于问答的摘要task-video
: Video-based Summarization 基于视频的摘要task-code
: Source Code Summarization 源码摘要task-control
: Controllable Summarization 可控制的摘要task-event
: Event-based Summarization 基于事件的摘要task-longtext
: Summarization for Long Text 长文本摘要task-knowledge
: Text Summarization with External Knowledge 可提取知识文本摘要task-highlight
: Pick out important content and emphasize 选出重要内容并强调task-analysis
: Model Understanding or Interpretability 模型的可理解性或可解释性task-novel
: Novel Chapter Generation 新章节的产生,novel在这里做形容词吧task-argument
: Automatic Argument Summarization 自动参数摘要
架构-Architecture (Mechanism)
arch-rnn
: Recurrent Neural Networks (LSTM, GRU) 递归神经网络arch-cnn
: Convolutional Neural Networks (CNN) 循环神经网络arch-transformer
: Transformer 翻译器arch-graph
: Graph Neural Networks or Statistic Graph Models 图神经网络或者统计图模型arch-gnn
: Graph Neural Networks 图神经网络arch-textrank
: TextRank 不翻译arch-att
: Attention Mechanism 注意力机制arch-pointer
: Pointer Layer 在这里应该不是指针层,不是输入层,不是输出层,肯定就是隐藏层了。arch-coverage
: Coverage Mechanism 覆盖机制???
训练(Training)
train-sup
: Supervised Learning 监督学习train-unsup
: Unsupervised Learning 非监督学习train-weak
: (impliestrain-sup
): Weakly Supervised Learning 弱监督学习train-multitask
: Multi-task Learning 多任务学习train-multilingual
: Multi-lingual Learning 多语言学习train-multimodal
: Multi-modal Learning 多模型学习train-auxiliary
: Joint Training 连接学习train-transfer
: Cross-domain Learning, Transfer Learning, Domain Adaptation 跨领域学习,转移学习,领域适应train-active
: Active Learning, Boostrapping 主动学习,助人为乐?什么翻译train-adver
: Adversarial Learning 对抗学习train-template
: Template-based Summarization 基于模板的摘要train-augment
: Data Augmentation 数据参数train-curriculum
: Curriculum Learning 课程学习?train-lowresource
: Low-resource Summarization 低资源摘要train-retrieval
: Retrieval-based Summarization 基于检索的摘要train-meta
: Meta-learning 元学习
预训练模型(Pre-trained Models)
pre-word2vec
: word2vecpre-glove
: GLoVepre-bert
: BERTpre-elmo
: ELMopre-hibert
: HiBERTpre-bart
: BARTpre-pegasus
: PEGASUSpre-unilm
: UNILMpre-mass
: MASSpre-T5
: Text-to-Text Transfer Transformerpre-S2ORC
: Pretrained model on semantic scholar open research corpuspre-sciBERT
: Scientific paper based pre-trained modelpre-SPECTER
: Scientific Paper Embeddings using Citationinformed TransformERs
不可微函数的松弛/训练方法(Relaxation/Training Methods for Non-differentiable Functions)
这里应该是针对不可导不可微的一些处理方法。softmax曾经看到过。
nondif-straightthrough
: Straight-through Estimatornondif-gumbelsoftmax
: Gumbel Softmaxnondif-minrisk
: Minimum Risk Trainingnondif-reinforce
: REINFORCE
对抗方法(Adversarial Methods)
adv-gan
: Generative Adversarial Networks 生成对抗网络adv-feat
: Adversarial Feature Learning 对抗特征学习adv-examp
: Adversarial Examples 对抗样例adv-train
: Adversarial Training 对抗训练
潜在变量模型(Latent Variable Models)
latent-vae
: Variational Auto-encoder 可变自动编码器latent-topic
: Topic Model
数据集(Dataset)
data-new
: Constructing a new dataset 组件新的数据集data-annotation
: Annotation Methodology 注释方法
评价(Evaluation)
这里应该就是说你的实验出来结果,怎么评价你的文本摘要出来是符合标注还是不符合标注的,有机构有人去专门评价你的工作。
eval-human
: Human Evaluation 人类评价eval-metric-rouge
: ROUGE 一个机构eval-metric-bertscore
: BERTScoreeval-aspect-coherence
: Coherenceeval-aspect-redundancy
: Redundancy of Summaryeval-aspect-factuality
: Factualityeval-aspect-abstractness
: Abstractnesseval-referenceQuality
: Reference Qualityeval-metric-learnable
: Metrics are Learnableeval-optimize-humanJudgement
: Optimization towards human judgementeval-reference-less
: Reference-less Approach to Automatic Evaluationeval-metric-unsupervised
: Unsupervised Automatic Evaluation
Survey
survey-2020
: A survey paper in 2020