1. 任务类型
1.1 生成式摘要(重写)和抽取式摘要(句子压缩任务)
1.1.1 生成式摘要abstractive summarization
本节内容参考了以下论文的文献综述部分:
序列生成(文本生成NLG)问题,一般使用seq2seq (S2S) 架构(encoder-decoder架构)。
sentence-fusion和重写(paraphrasing)
rephrasing and introducing new concepts/words(语出Friendly Topic Assistant for Transformer Based Abstractive Summarization)
基于结构的方法:
- 基于树的方法:tree linearization
- 基于模板的方法:
Generating single and multi-document summaries with gistexter
sArAmsha-A Kannada abstractive summarizer
- 基于实体的方法
- Lead and Body Phrase Method(lead指开头。总之是找一些重要短语然后做一些操作的方法,具体的其实我也没看懂,可以参考这篇博客:Towards Automatic Summarization. Part 2. Abstractive Methods. | by Sciforce | Sciforce | Medium)
- Rule Based Method
- 基于语义的方法
多模态语义模型
基于information item的方法
基于语义图的方法
常见问题及针对该问题提出的解决方案:
- 文本重复
PGN(Get to the point: Summarization with pointer-generator networks.)中提出的coverage机制就是用来解决这一问题的(虽然我觉得实验上好像文本重复问题还是非常严重)
- 事实不一致问题
- 衡量原文与摘要的事实一致性:
The Factual Inconsistency Problem in Abstractive Text Summarization: A Survey
Assessing The Factual Accuracy of Generated Text
Multi-Fact Correction in Abstractive Text Summarization
Evaluating the Factual Consistency of Abstractive Text Summarization
Asking and Answering Questions to Evaluate the Factual Consistency of Summaries
FEQA: A Question Answering Evaluation Framework for Faithfulness Assessment in Abstractive Summarization
Falsesum: Generating Document-level NLI Examples for Recognizing Factual Inconsistency in Summarization
QAFactEval: Improved QA-Based Factual Consistency Evaluation for Summarization
Investigating Crowdsourcing Protocols for Evaluating the Factual Consistency of Summaries
- 直接解决事实不一致问题:
Joint Parsing and Generation for Abstractive Summarization
Masked Summarization to Generate Factually Inconsistent Summaries for Improved Factual Consistency Checking
- 文本不连贯(fluent或coherent)
- 原文太长,难以直接输入模型(Transformer模型的quadradic复杂度)
- 抽取+生成范式:证明这种范式比直接生成的效果更好:Bottom-Up Abstractive Summarization, Improving neural abstractive document summarization with explicit information selection modeling
- 切分数据范式
- 改进模型
典型的使用seq2seq+attention范式做生成式摘要的论文:
- A Neural Attention Model for Abstractive Sentence Summarization
- Abstractive Sentence Summarization with Attentive Recurrent Neural Networks
- Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond
- Get To The Point: Summarization with Pointer-Generator Networks
- Abstractive Document Summarization with a Graph-Based Attentional Neural Model
- 感觉没之前几篇那么典型:Query Focused Abstractive Summarization: Incorporating Query Relevance, Multi-Document Coverage, and Summary Length Constraints into seq2seq Models
- A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents
- Structure-Infused Copy Mechanisms for Abstractive Summarization
- 2019年综述:Abstractive summarization: An overview of the state of the art
1.1.2 抽取式摘要extractive summarization
本节内容参考了以下论文的文献综述部分:12
缺点:在话题切换时缺乏连贯性。
- Term Frequency-Inverse Document Frequency Method
- Cluster Based Method:聚类出各主题,文档表示方法为单词的TF-IDF得分,High frequency term represents the theme of a cluster,基于句子与簇中心的关系选择摘要句
- Text Summarization with Neural Network
- Text Summarization with Fuzzy Logic
- Graph based Method
- Latent Semantic Analysis Method: LSA
- Machine Learning approach
- Query based summarization、
常见范式:做句子的二分类任务(该句是否属于摘要),将预测为“属于”的句子拼起来,组成摘要。
identify and then concatenate the most representative sentences as a summary(语出Friendly Topic Assistant for Transformer Based Abstractive Summarization)
模型分成3层来做表示学习(单词→句子→文档),使用attention等机制提高表示能力。
- 用基于图的表征来捕获显著textual units:TF-IDF similarity(Lexrank: Graph-based lexical centrality as salience in text summarization.) ;discourse relation(Textrank: Bringing order into text.);document-sentence two-layer relations(An exploration of document impact on graph-based multi-document summarization.);multi-modal (Graph-based multi-modality learning for topic-focused multidocument summarization.) 和 query information (Mutually reinforced manifold-ranking based relevance propagation model for query-focused multi-document summarization. )
- 使用GNN方法捕获文档间关系:Graph-based neural multi-document summarization.(构建discourse图并用GCN表示textual units); Hierarchical transformers for multi-document summarization.(用entity linking technique捕获句子间的全局依赖,用基于图的神经网络模型对句子进行排序)
使用深度学习方法做抽取式摘要的经典论文:
- SummaRuNNer: A Recurrent Neural Network Based Sequence Model for Extractive Summarization of Documents
- Extractive Summarization using Deep Learning
- Neural Extractive Summarization with Side Information
- Ranking Sentences for Extractive Summarization with Reinforcement Learning
- Fine-tune BERT for Extractive Summarization
- Extractive Summarization of Long Documents by Combining Global and Local Context
- Extractive Summarization as Text Matching
1.2 单文档摘要和多文档摘要
1.2.1 单文档摘要single-document summarization
主题论文总结4:单文档摘要(以罗列为主)(持续更新ing…)
1.2.2 多文档摘要multi-document summarization
本节内容参考了以下论文的文献综述部分:2
看了几篇MDS的论文感觉无非就是一种长文本摘要啊……有的论文就是单纯把多篇文档拼在一起,用[END]token作间隔。(A Multi-Document Coverage Reward for RELAXed Multi-Document Summarization)
输入的多文档可能是冗余的,甚至含有自相矛盾的内容(A common theory of information fusion from multiple text sources step one: cross-document structure.)
迁移单文档摘要的模型到多文档摘要上,以回避缺乏小规模数据集的问题:
Generating wikipedia by summarizing long sequences.:定义Wikipedia生成问题,并提出WikiSum数据集。
Towards a neural network approach to abstractive multi-document summarization.
Multi-news: A large-scale multi-document summarization dataset and abstractive hierarchical model. :提出MultiNews数据集,在抽取过程后应用seq2seq模型生成摘要。
Leveraging graph to improve abstractive multi-document summarization.:用显式图表征建模文档间关系,结合预训练语言模型处理长文本。
1.3 重要研究方向
- 长文本摘要
- 结构化文本摘要:主题论文总结1:structured text summarization(持续更新ing…)_诸神缄默不语的博客-CSDN博客
- 对话/会议摘要:主题论文总结2:会议/对话摘要任务(持续更新ing…)_诸神缄默不语的博客-CSDN博客
- 维基百科生成:主题论文总结3:维基百科生成任务(持续更新ing…)_诸神缄默不语的博客-CSDN博客
- 科技文献(论文)摘要:主题论文总结5:科技文献(论文)摘要
2. 抽取式摘要
2.1 无监督方法
重要模型:
LEAD-3算法
TextRank算法
2.2 有监督方法
重要模型:
Fine-tune BERT for Extractive Summarization
BertSum算法(官方源代码:nlpyang/BertSum: Code for paper Fine-tune BERT for Extractive Summarization;热心网友写的可以直接用中文数据作为输入的版本:425776024/bertsum-chinese: chinese bertsum ; bertsum 抽取式模型中文版本;给出案例数据、全代码注释;下载即可训练、预测、学习)
3. 生成式摘要
3.1 抽取+生成
3.1.1 解耦的
对范式的介绍:
Abstractive multi-document summarization via phrase selection and merging.:分成两个阶段:第一步,通过无监督的方法或语言学知识来抽取原文中的关键文本元素(key textual elements)。第二步,用语言学规则或文本生成方法来rewrite或paraphrase抽取出来的元素,生成原文的准确摘要。(转引自LCSTS: A Large Scale Chinese Short Text Summarization Dataset)
重要模型:SPACES模型(苏剑林的介绍博文:SPACES:“抽取-生成”式长文本摘要(法研杯总结) - 科学空间|Scientific Spaces;官方源代码:bojone/SPACES: 端到端的长本文摘要模型(法研杯2020司法摘要赛道);热心网友写的PyTorch版复现(不完全复现):eryihaha/SPACES-Pytorch: 苏神SPACE pytorch版本复现)
3.1.2 端到端(end2end)的
3.2 纯生成式方法
3.2.1 基础seq2seq模型
Transformers版,参考PyTorch官方教程:Language Modeling with nn.Transformer and TorchText — PyTorch Tutorials 1.11.0+cu102 documentation
其他参考资料:LCSTS: A Large Scale Chinese Short Text Summarization Dataset:seq2seq (RNN) 没有代码
3.2.2 Pointer-Generator模型
Get to the point: Summarization with pointer-generator networks.
4. 评估指标
本节内容参考了:3
以下评估指标往往也用于翻译、QA等其他文本生成(NLG)任务。
常用术语:
模型生成的句子、预测结果——candidate
真实摘要、标签——reference、ground-truth
precision:candidate中匹配reference的内容占candidate比例
recall:candidate中匹配reference的内容占reference比例
示例:
Reference: I work on machine learning. Candidate A: I work. Candidate B: He works on machine learning.
在这个例子中,用unigram4衡量匹配:A就比B的precision更高(A的匹配内容I work占candidate 100%,B的on machine learning占60%),但B的recall更高(60% VS 40%)。
4.1 人工评估指标
文本的流畅程度、对原文的忠实程度、对原文重要内容的包含程度、语句的简洁程度等
4.2 ROUGE (Recall Oriented Understudy for Gisting Evaluation)
出处:ROUGE: A Package for Automatic Evaluation of Summaries
感觉没有2004年之后的文本摘要论文不使用这个指标的,如果看到有的话我会专门来这里提一嘴的。
分类:ROUGE-N(常用其中的ROUGE-1和ROUGE-2), ROUGE-L,ROUGE-W,ROUGE-S(后两种不常用)
原版论文中ROUGE主要关注recall值,但事实上在用的时候可以用precision、recall和F值。
4.2.1 计算指标
每种rouge值原本都是计算recall的,和前面第4节开头介绍的precision和recall值计算方法差不多,主要区别在于这个匹配文本的单位的选择:
4.2.2 对rouge指标的更深入研究和改进
A Graph-theoretic Summary Evaluation for ROUGE
4.3 BLEU (Bilingual Evaluation Understudy)
更常用于翻译领域。
出处:Bleu: a Method for Automatic Evaluation of Machine Translation
precision用modified n-gram precision估计,recall用best match length估计。
Modified n-gram precision:
n-gram precision是candidate中与reference匹配的n-grams占candidates的比例。但仅用这一指标会出现问题。
举例来说:
Reference: I work on machine learning. Candidate 1: He works on machine learning. Candidate 2: He works on on machine machine learning learning.
candidate 1的unigram precision有60%(3/5),candidate 2的有75%(6/8),但显然candidate 1比2更好。
为了解决这种问题,我们提出了“modified” n-gram precision,仅按照reference中匹配文本的出现次数来计算candidate中的出现次数。这样candidate中的on、machine和learning就各自只计算一次,candidate 2的unigram precision就变成了37.5%(3/8)。
对多个candidate的n-gram precision,求几何平均(因为precision随n呈几何增长,因此対数平均能更好地代表所有数值):
Best match length:
recall的问题在于可能存在多个reference texts,故难以衡量candidate对整体reference的sensitivity6。显然长的candidate会包含更多匹配文本,但我们也已经保证了candidate不会无限长,因为这样的precision可能很低。因此,我们可以从惩罚candidate的简洁性(文本短)入手来设计recall指标:
在modified n-gram precision中添加一个multiplicative factor B P :
其中 c 是candidates总长度,r 是reference有效长度(如reference长度平均值),随着candidate长度(c)下降,B P 也随之减少,起到了惩罚短句的作用。
4.4 Perplexity
4.5 METEOR (Metric for Evaluation for Translation with Explicit Ordering)
出处:METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments
也是常用于翻译领域。
这个指标声称是跟人工评估指标关联性更高。
BLEU的问题在于 B P BPBP 值所用的长度是平均值,因此单句得分不清晰。而METEOR调整了precision和recall的计算方式,用基于mapping unigrams的weighted F-score和penalty function for incorrect word order来代替。
4.6 Bertscore
使用该指标的论文:Rewards with Negative Examples for Reinforced Topic-Focused Abstractive Summarization
4.7 其他指标
Revisiting Automatic Evaluation of Extractive Summarization Task: Can We Do Better than ROUGE?
Benchmarking Answer Verification Methods for Question Answering-Based Summarization Evaluation Metrics
SARI
InfoLM: A New Metric to Evaluate Summarization & Data2Text Generation
SPICE
Play the Shannon Game With Language Models: A Human-Free Approach to Summary Evaluation
Reference-free Summarization Evaluation via Semantic Correlation and Compression Ratio:还没有放出来