Original link: click to open the link
Abstract: ETH-DS3Lab at SemEval-2018 Task 7: Effectively Combining Recurrent and Convolutional Neural Networks for Relation Classification and Extraction@theodoric008 recommends Relation Extraction This article is from DS3Lab of ETH Zurich. The article conducts very systematic experiments on entity relation extraction tasks. And won the championship in the semantic relation extraction and classification tasks of the 12th International Semantic Evaluation Competition SemEval 2018.
ETH-DS3Lab at SemEval-2018 Task 7: Effectively Combining Recurrent and Convolutional Neural Networks for Relation Classification and Extraction
@theodoric008 推荐
Relation Extraction
This article is from DS3Lab of ETH Zurich. The article has carried out a very systematic experiment on the entity relation extraction task, and won the championship in the semantic relation extraction and classification task of the 12th International Semantic Evaluation Competition SemEval 2018. This article is rigorous in thinking and worthy of careful study by domestic scholars.
Paper link
https://www.paperweekly.site/papers/1833
Personalizing Dialogue Agents: I have a dog, do you have pets too?
@yihongchen 推荐
Dialogue System
This article is the work of Facebook AI Research published at NIPS 2018. The paper trains a Profile-based chatbot on a conversation dataset called PERSONA-CHAT, which contains over 160,000 conversations.
This paper aims to address the following issues:
Chatbots lack consistent personality traits
Chatbots lack long-term memory
Chatbots often give vague responses like I don't know
Paper link
https://www.paperweekly.site/papers/1802
dataset link
https://github.com/facebookresearch/ParlAI/tree/master/parlai/tasks/personachat
DiSAN: Directional Self-Attention Network for RNN/CNN-Free Language Understanding
@zhkun 推荐
Natural Language Understanding
This article is the work of the University of Technology Sydney published in AAAI 2018. This article is another application of Self-Attention. The author proposes a new directional Attention to understand semantics more effectively.
论文链接
https://www.paperweekly.site/papers/1822
代码链接
https://github.com/shaohua0116/Group-Normalization-Tensorflow
DetNet: A Backbone network for Object Detection
@chlr1995 推荐
Object Detection
本文来自清华大学和 Face++，文章分析了使用 ImageNet 预训练网络调优检测器的缺陷，研究通过保持空间分辨率和扩大感受野，提出了一种新的为检测任务设计的骨干网络 DetNet。
实验结果表明，基于低复杂度的 DetNet59 骨干网络，在 MSCOCO 目标检测和实例分割追踪任务上都取得当前最佳的成绩。
论文链接
https://www.paperweekly.site/papers/1844
Imagine This! Scripts to Compositions to Videos
@chlr1995 推荐
Video Caption
本文以《摩登原始人》的动画片段作为训练数据，对每个片段进行详细的文本标注，最终训练得到一个可以通过给定脚本或文字描述生成动画片段的模型。
模型称为 Craft，分为布局、实体、背景，三个部分。虽然现阶段模型存在着很多问题，但是这个研究在理解文本和视频图像高层语义方面有着很大的意义。
论文链接
https://www.paperweekly.site/papers/1838
Generating Diverse and Accurate Visual Captions by Comparative Adversarial Learning
@Aidon 推荐
Image Caption
This article is from the University of Washington and Microsoft. The article proposes a GAN-based Image Caption framework. The highlights are as follows:
- It is proposed to use the comparative relevance score to measure the quality of image-text to guide the training of the model, and introduce unrelated captions in the training process;
- Using human evaluations to evaluate the accuracy of captions, the results compared with the traditional six evaluation indicators are given;
- It is proposed to evaluate the diversity of captions by comparing the variance of caption feature vectors.
Paper link
https://www.paperweekly.site/papers/1842
Simultaneously Self-Attending to All Mentions for Full-Abstract Biological Relation Extraction
@robertdlut 推荐
Self-Attention
This paper is a work of Andrew McCallum's team applying Self-Attention on the task of biomedical relation extraction. The author of this paper proposes a document-level biological relationship extraction model. The author uses Google to propose a Transformer containing Self-Attention to perform representation learning on the input text. The difference from the original transformer is that they use a CNN with a window size of 5. instead of the original FNN.
论文链接
https://www.paperweekly.site/papers/1787
代码链接
https://github.com/patverga/bran
Evaluation of Session-based Recommendation Algorithms
@Ttssxuan 推荐
Recommender System
本文系统地介绍了 Session-based Recommendation，主要针对 baseline methods, nearest-neighbor techniques, recurrent neural networks 和 (hybrid) factorization-based methods 等 4 大类算法进行介绍。
此外，本文使用 RSC15、TMALL、ZALANDO、RETAILROCKET、8TRACKS 、AOTM、30MUSIC、NOWPLAYING、CLEF 等 7 个数据集进行分析，在 Mean Reciprocal Rank (MRR)、Coverage、Popularity bias、Cold start、Scalability、Precision、Recall 等指标上进行比较。
论文链接
https://www.paperweekly.site/papers/1809
代码链接
https://www.dropbox.com/sh/7qdquluflk032ot/AACoz2Go49q1mTpXYGe0gaANa?dl=0
On the Convergence of Adam and Beyond
@chlr1995 推荐
Neural Network
本文是 ICLR 2018 最佳论文之一。在神经网络优化方法中，有很多类似 Adam、RMSprop 这一类的自适应学习率的方法，但是在实际应用中，虽然这一类方法在初期下降的很快，但是往往存在着最终收敛效果不如 SGD+Momentum 的问题。
作者发现，导致这样问题的其中一个原因是因为使用了指数滑动平均，这使得学习率在某些点会出现激增。在实验中，作者给出了一个简单的凸优化问题，结果显示 Adam 并不能收敛到最优点。
在此基础上，作者提出了一种改进方案，使得 Adam 具有长期记忆能力，来解决这个问题，同时没有增加太多的额外开销。
论文链接
https://www.paperweekly.site/papers/1841
Neural Baby Talk
@jamiechoi 推荐
Image Captioning
本文是佐治亚理工学院发表于 CVPR 2018 的工作，文章结合了 image captioning 的两种做法：以前基于 template 的生成方法（baby talk）和近年来主流的 encoder-decoder 方法（neural talk）。
论文主要做法其实跟作者以前的工作"Knowing When to Look: Adaptive Attention via A Visual Sentinel for Image Captioning"类似：在每一个 timestep，模型决定生成到底是生成 textual word（不包含视觉信息的连接词），还是生成 visual word。其中 visual word 的生成是一个自由的接口，可以与不同的 object detector 对接。
论文链接
https://www.paperweekly.site/papers/1801
代码链接
https://github.com/jiasenlu/NeuralBabyTalk
Context Encoding for Semantic Segmentation
@wanzysky 推荐
Semantic Segmentation
本文提出了一种与类别预测相关的网络结构，使得在一定程度上降低了分割任务的难度。Channel attention 和空间 attention 形成互补，Global contextual loss 增强 context 信息，同时提高了小物体的分割精度。
论文链接
https://www.paperweekly.site/papers/1814
代码链接
https://github.com/zhanghang1989/PyTorch-Encoding
Adaptive Graph Convolutional Neural Networks
@VIPSP 推荐
Convolutional Neural Network
图卷积神经网络（Graph CNN）是经典 CNN 的推广方法，可用于处理分子数据、点云和社交网络等图数据。Graph CNN 中的的滤波器大多是为固定和共享的图结构而构建的。但是，对于大多数真实数据而言，图结构的大小和连接性都是不同的。
本论文提出了一种有泛化能力且灵活的 Graph CNN，其可以使用任意图结构的数据作为输入。通过这种方式，可以在训练时为每个图数据构建一个任务驱动的自适应图（adaptive graph）。
为了有效地学习这种图，作者提出了一种距离度量学习方法。并且在九个图结构数据集上进行了大量实验，结果表明本文方法在收敛速度和预测准确度方面都有更优的表现。