问答系统

Fluent Response Generation for Conversational Question Answering 
https://www.aclweb.org/anthology/2020.acl-main.19/
seq2seq模型,----针对span 抽取模型的对nlg的忽略,生成流畅的模型,
 
训练目标是:
对问题+专家短语(简单答案),预测出一个流畅的合法句子。
思路为: 使用句法迁移模型或者指针网络生成多个句子,再使用分类器选择最优的句子。
会涉及到词向量的迁移。
20210113 整理
 
Good Question! Statistical Ranking for Question Generation 
2010
https://www.aclweb.org/anthology/N10-1086.pdf
问题生成
 
 
 
 
 
Open-Domain Question Answering
https://www.aclweb.org/anthology/2020.acl-tutorials.8.pdf
 
论文主要论述三种QA的思路:
1. 分两步模型: step 1 检索阶段,从大量的包含答案的文档中检索出top n。主要使用算法为,TF-IDF或BM25等。 step 2. 阅读模块从给定的段或文档中读取答案。一般使用自然语言的月度模型
2. 低密度检索和端到端训练:
       使用低维向量训练检索器,并在低维向量空间检索。通常会包括庞大的向量空间
3. 无训练模型
       使用大规模的预训练模型作为知识库,在推断的时候不再访问文本数据。
20210113整理
 
 
 
MIX : a Multi-task Learning Approach to Solve Open-Domain Question
Answering
https://arxiv.org/abs/2012.09766
1. 
整理20211014
 
 
Squad: 100,000+ questions for machine comprehension of text
https://arxiv.org/abs/1606.05250v3
 
 
hrase-Indexed Question Answering: A New Challenge for Scalable Document Comprehension
https://arxiv.org/pdf/1804.07726.pdf
定义Extractive question answering任务为一个对于(s,e)的分布概率。
目标是:预测目标,在给定参数下,使得记分函数分别最大的a为预测值。
1. 设计计分函数,以往的思路是,使用问题和答案做一个注意力连接。缺点在于,1. 预训练要计算问题对于每个文档,2. 不能独立的访问模型理解的文档。
2. 新的思路是把计分函数分为两部分:G(Q),H(D,a), 积分函数为G,H的內积。H独立与问题来建模文档。
语言理解能力与独立于终点任务的文本表示相关。
模型架构的主要组件为:1. 问提表示,使用LSTM+注意力,2. 文档表示 使用LSTM+注意力。3. 答案表示 a(s,e),
 
 
 
整理20210118
 
 
QANET: COMBINING LOCAL CONVOLUTION WITH GLOBAL SELF-ATTENTION FOR READING COMPREHENSION
https://openreview.net/pdf?id=B14TlG-RW
 
 
BI-DIRECTIONAL ATTENTION FLOW FOR MACHINE COMPREHENSION
https://arxiv.org/pdf/1611.01603.pdf
 
4. Attention Flow Layer.: 与传统的注意力模式不同。不是把查询向量和上下文向量统计到一个单一的向量中。
输入是上下文H,查询U,输出是向下文U的对查询相关的表示
从两个方向计算上下文:  上下文-查询, 查询到上下文
 
 
整理20210122
 
https://www.aclweb.org/anthology/2020.aacl-main.14.pdf
Exploiting WordNet Synset and Hypernym Representations for Answer Selection
 
1. wordNet在概念和语义关系上增加知识。使用同义词和上位词增加词语的表示,并统计这个两个概念对于注意力机制算出的分数。
    算法包括四部分:wordNet增强的词表示,句子编码器,Wordnet增强的注意力机制,分层的文档编码
 
Answer selection 被称为看作,从文档中选择一个句子,并进行排序的问题,
不同于基于社区的问答系统,所有的答案来自于同一篇文档,并且候选答案于问题语义相关。
 
整理20210213
 
https://www.aclweb.org/anthology/2020.aacl-main.54.pdf
Point-of-Interest Oriented Question Answering with Joint Inference of Semantic Matching and Distance Correlation
目标是:通过生成一系列的兴趣点来回答用户提出的问题。比如在,纽约附近有什么好玩的?回答可能是一些列的兴趣点,比如一系列的餐厅,饭店等。
 
 
https://www.aclweb.org/anthology/2020.aacl-main.70.pdf
Answering Product-related Questions with Heterogeneous Information
产品评论相关问答,
自然文本和属性值---优势在于一个集成的框架处理不同的信息源。
 
https://www.aclweb.org/anthology/2020.iwdp-1.8.pdf
Bridging Question Answering and Discourse: The case of Multi-Sentence Questions
多句问答
 
 
https://www.aclweb.org/anthology/2020.acl-main.85.pdf
Contextualized Sparse Representations for Real-Time Open-Domain Question Answering
通过检索的方式,在以前的检索的算法精确度低,通过上下文的表示来提高算法的精度。
 
 
https://arxiv.org/pdf/1906.05807v2.pdf
Real-Time Open-Domain Question Answering with Dense-Sparse Phrase Index  【2019】
端到端的检索方式。
 
https://arxiv.org/pdf/1704.00051.pdf 
Reading Wikipedia to Answer Open-Domain Questions 【2017】
稀疏向量表示
 
 
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.90.pdf
Learning to Identify Follow-Up Questions in Conversational Question Answering
后续问题的解决
 
https://www.aclweb.org/anthology/2020.acl-main.91.pdf
Query Graph Generation for Answering Multi-hop Complex Questions from Knowledge Bases
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.132.pdf
Bridging Anaphora Resolution as Question Answering
回指,一个名称,在前面段落中,代表什么?
 
 
https://www.aclweb.org/anthology/2020.acl-main.247.pdf
Span Selection Pre-training for Question Answering
bert用于阅读理解,不是生成答案,而是在span 抽取方式。
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.411.pdf
DeFormer: Decomposing Pre-trained Transformers for Faster Question Answering
对bert模型的改进,是的更加的快,并且该变了他的目标loss
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.412.pdf
Improving Multi-hop Question Answering over Knowledge Graphs using Knowledge Base Embeddings
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.413.pdf
Template-Based Question Generation from Retrieved Sentences for Improved Unsupervised Question Answering
无监督学习,给定一段文字,抽取一个问答对,是的答案是给定的一段里面。
 
 
https://www.aclweb.org/anthology/2020.acl-main.498.pdf
Crossing Variational Autoencoders for Answer Retrieval
用深度学习对语义进行解析,-----todo
 
 
https://www.aclweb.org/anthology/2020.acl-main.454.pdf
FEQA: A Question Answering Evaluation Framework for Faithfulness Assessment in Abstractive Summarization
一种新的评估方式
 
https://www.aclweb.org/anthology/2020.acl-main.499.pdf
Logic-Guided Data Augmentation and Regularization for Consistent Question Answering
很多问题,需要定性、定量或逻辑比较在两个实体或事件之间。
 
 
https://www.aclweb.org/anthology/2020.acl-main.501.pdf
Probabilistic Assumptions Matter: Improved Models for Distantly-Supervised Document-Level Question Answering
远程监督+span抽取方式
 
 
https://www.aclweb.org/anthology/2020.acl-main.503.pdf
Selective Question Answering under Domain Shift
在没有正确答案的死后,如何弃权,以提高准确性
 
 
https://www.aclweb.org/anthology/2020.acl-main.504.pdf
The Cascade Transformer: an Application for Efficient Answer Sentence Selection
 
https://www.aclweb.org/anthology/2020.acl-main.505.pdf
Transformers to Learn Hierarchical Contexts in Multiparty Dialogue for Span-based Question Answering
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.508.pdf
WinoWhy: A Deep Diagnosis of Essential Commonsense Knowledge for Answering Winograd Schema Challenge
 
 
https://www.aclweb.org/anthology/2020.acl-main.600.pdf
Harvesting and Refining Question-Answer Pairs for Unsupervised QA
对无标注问答对的处理
 
 
https://www.aclweb.org/anthology/2020.acl-main.602.pdf
R4C: A Benchmark for Evaluating RC Systems to Get the Right Answer for the Right Reason
一个评估方式
 
 
 
https://www.aclweb.org/anthology/2020.acl-main.642.pdf
Aligned Dual Channel Graph Convolutional Network for Visual Question Answering
给定一个图片做问答
 
https://www.aclweb.org/anthology/2020.acl-main.643.pdf
Multimodal Neural Graph Memory Networks for Visual Question Answering
给图片做问答
 
https://www.aclweb.org/anthology/2020.acl-main.772.pdf
QUASE: Question-Answer Driven Sentence Encoding
qa是否对其他的任务有效?
 
https://www.aclweb.org/anthology/2020.acl-demos.5.pdf
Talk to Papers: Bringing Neural Question Answering to Academic Search
从一堆论文中精确的找出答案,
 
 
https://www.aclweb.org/anthology/2020.bionlp-1.12.pdf
Entity-Enriched Neural Models for Clinical Question Answering
在医学问答中,对主任务加入逻辑形式的训练目标,并组成逻辑增强模型
 
 
https://slideslive.com/38929544/bidirectional-answertoanswer-coattention-for-short-answer-grading-using-deep-learning
视频
 
 
 
 
https://github.com/deepset-ai/COVID-QA
 
 
https://www.aclweb.org/anthology/2020.ecnlp-1.5.pdf
Using Large Pretrained Language Models for Answering User Queries from Product Specifications
对产品规格的问答
 
https://www.aclweb.org/anthology/2020.ecnlp-1.11.pdf
SimsterQ: A Similarity based Clustering Approach to Opinion Question Answering
相似的基于观点的问答
 
https://www.aclweb.org/anthology/2020.nli-1.1.pdf
Answering Complex Questions by Combining Information from Curated and Extracted Knowledge Bases
策划回答复杂问题
 
 
https://www.aclweb.org/anthology/2020.tacl-1.30.pdf
TYDI QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages
 
https://www.aclweb.org/anthology/2020.emnlp-main.10.pdf
Learning to Explain: Datasets and Models for Identifying Valid Reasoning Chains in Multihop Question-Answering
为什么答案是正确的?
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.11.pdf
Self-Supervised Knowledge Triplet Learning for Zero-Shot Question Answering
所有问答系统是推广为出现问题。标注昂贵,而且还可能引起非计划内的偏差。一种启发式的方法。
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.12.pdf
More Bang for Your Buck: Natural Perturbation for Robust Question Answering
标注很昂贵,提出人工选择一些种子,并在相反的一些方向做出一些扰动,来够着语料,
 
https://www.aclweb.org/anthology/2020.emnlp-main.49.pdf
Event Extraction by Answering (Almost) Natural Questions
在时间抽取过程中,实体识别是一个重要的步骤,因为可能导致错误传递,
把他看成问答方式,按照端到端的方式进行提取。
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.63.pdf
MUTANT: A Training Paradigm for Out-of-Distribution Generalization in Visual Question Answering
视觉问答方面
 
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.84.pdf
Look at the First Sentence: Position Bias in Question Answering
在一般的qa模型里面,抽取式问答,预测位置。位置便宜较大的情况下,如何考虑,测试偏移的问题
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.85.pdf
ProtoQA: A Question Answering Dataset for Prototypical Common-Sense Reasoning
问答数据库
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.87.pdf
Unsupervised Adaptation of Question Answering Systems via Generative Self-training
 
https://www.aclweb.org/anthology/2020.emnlp-main.99.pdf
Scalable Multi-Hop Relational Reasoning for Knowledge-Aware Question Answering
推理方面的,
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.170.pdf
If Beam Search is the Answer, What was the Question?
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.188.pdf
产品相关的问答,如何取得一个准确的问答,
 
 
https://www.aclweb.org/anthology/2020.emnlp-main.189.pdf
Context-Aware Answer Extraction in Question Answering
基于上下文的问答抽取
 
https://www.aclweb.org/anthology/2020.emnlp-main.190.pdf
What do Models Learn from Question Answering Datasets?
 
https://www.aclweb.org/anthology/2020.emnlp-main.244.pdf
Don’t Read Too Much Into It: Adaptive Computation for Open-Domain Question Answering
开发领域的问答读写与机器阅读的区别在于,答案是从文档还是从段落中读取。
 
3、在ODQA中调整计算。对小答案集合中预料在每个transform层中建立Tower,以减少计算。记录被处理的tower和最后的值,
 
 
https://www.aclweb.org/anthology/2020.spnlp-1.12.pdf
Deeply Embedded Knowledge Representation & Reasoning For Natural Language Question Answering: A Practitioner’s Perspective
知识的表示和推理很难有一个好的鲁棒性。目前有一些追求通用的表示,但是也与现实有一定的差距。
目标是,不依赖于解析建立一个知识推理和表示模型。使用神经网络代替解析。 目标神经网络和一个可解释的逻辑形式。
 
对知识库的描述:如果存在问题和知识被表示为R,则存在算法A,能够计算出答案。目前没有一个端到端的模型能保存精度对其进行表示。
提出一个知识表示和推理方案,并使得他在神经网络中模拟运行。
有三种设计方案: 1. 如果编码一个符号到要给表示到向量R里面,使得特定的过程能把他解释为原始符号。2. 构造一个神经网络作为对向量R的一个反馈。 3,使用算法进行计算。
 
 A KR Solution:是一个解析转化器:把自然语言输入和一些规则,然后得到给定的答案。
 
数据的表示和推断,
----表示的形式,
----推断:
 
 
https://www.aclweb.org/anthology/2020.splu-1.4.pdf
They Are Not All Alike: Answering Different Spatial Questions Requires Different Grounding Strategies
 
 
 
https://www.aclweb.org/anthology/2020.scai-1.3.pdf
Multi-Task Learning using Dynamic Task Weighting for Conversational Question Answering
多任务的组合方式
 
 
https://www.aclweb.org/anthology/P13-2125.pdf
Learning Semantic Textual Similarity with Structural Representations
 
https://www.aclweb.org/anthology/2020.nlpcovid19-2.33.pdf
Using the Poly-encoder for a COVID-19 Question Answering System
 
 
https://openreview.net/pdf?id=SkxgnnNFvH
Poly-encoders: architectures and pre-training strategies for fast and accurate multi-sentence scoring
两种典型的fine-ture结构,双向编码和交叉编码。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

你可能感兴趣的:(问答,自然语言处理)