2020-02-16 10:11:00
Author | Tencent AI Lab Zebian | Jia Wei
AAAI is one of the top international conferences in the field of artificial intelligence. This year's AAAI 2020 is the 34th, held in New York on February 7 to 12.
This year, Tencent AI Lab a total of 31 papers were selected for the third participants, covering natural language processing, computer vision, machine learning, robotics. Today the paper interpreted as: "Relation Extraction Exploiting Full Dependency Forests".
This paper led by Tencent AI Lab, and Ohio State University and other institutions in collaboration. The article first presents the data structures are completely dependent forests. It is possible to significantly mitigate error propagation caused by the use of 1-best dependency syntax tree, and can be further adjusted according to the terminal of the syntax analyzer task.
Papers link: https: //freesunshine0316.github.io/files/RE_parser_joint__AAAI_2020_.pdf
Many natural language processing (NLP) tasks require knowledge of syntactic analysis, especially the dependency parsing a direct relationship between the word and the word modeling. At present work is to obtain 1-best dependency tree in the preprocessing stage, and then again these dependency tree used as an additional input for training and test terminals task model.
Such defects have to do
Error (1) parsing task will be propagated to the terminal;
(2) parsing model can not be adjusted in accordance with the terminal adaptation task.
Recent work [1] introduces dependencies forest, 1-best compared dependency tree, forest dependency can more (n-best) dependency tree information into a compact configuration in FIG compression, without significantly increasing the complexity of structure more knowledge into the premise degrees. But this is only alleviate the first drawback; and either 1-best dependency tree or forest dependent are discrete structures, to adapt to the adjustment of syntactic analysis model, the difficulty will increase a lot.
FIG 1 is completely dependent exemplary Forest
为了缓解以上的问题，作者提出了“完全依存森林”。如图1所示，一个完全依存森林被定义成一个3D的向量，包含了全部可能的词与词的依存关系，区别于已有的1-best句法树和句法森林，完全依存森林
1）包括了全部的句法分析的信息，这样终端任务模型会根据任务信号从完全依存森林中学习获取对它有用的信息；
2）提供了可微分的连接，使得句法分析模型很容易根据终端任务的信号进行适应调整。
作者在关系抽取任务上对完全依存森林进行了验证，该任务的目标是识别出所有有关系的实体对并确定它们的关系。
图 2 基于数据依存型 CNN 的模型
如图2所示，为了更好地表示完全依存森林，作者提出一种基于数据依存型 CNN（DDCNN）的模型，左边是基于Deep biaffine [2]的依存分析器，它直接输出依存森林给右边DDCNN模型。右边的DDCNN模型考虑了一个句子会有多对目标提及对（target mention pair）（如图2中红色和蓝色标记的）的情况，通过将目标提及对输入给右上部分的核生成网络以动态地产生CNN网络的核参数，可用来计算最终分类器（图2中间上部分）的输入。
为了验证完全依存森林的有效性，作者在几个标准的（包括新闻和生物领域）关系抽取数据集上进行了验证。结果显示在生物领域数据上的提升比较明显，在BioCreative VI CPR和SemEval-2010 task 8数据的结果如表 1和表2所示。
表 1 在 BioCreative VI CPR测试集上的结果
表 2 在 SemEval-2010 task 8测试集上的结果
表3模拟了资源受限（小规模树库训练依存分析器）的情况下，给出了在依存分析器精度进一步下降的情况下各个模型的表现情况，从中可以看出该文章提出的框架在只有1K树库的情况下依然表现优异。
表 3 树库资源受限（依存分析精度下降）的情况下，在 BioCreative VI CPR的测试集上结果
[1] Leveraging Dependency Forest for Neural Medical Relation Extraction. Linfeng Song, Yue Zhang, Daniel Gildea, Mo Yu, Zhiguo Wang and Jinsong Su. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).
[2] Deep Biaffine Attention for Neural Dependency Parsing. Timothy Dozat and Christopher D. Manning. In ICLR 2017.
AAAI 2020 报道：
新型冠状病毒疫情下，AAAI2020 还去开会吗？
美国拒绝入境，AAAI2020现场参会告吹，论文如何分享？
AAAI 最佳论文公布
AAAI 2020正式开幕，37%录用论文来自中国，连续三年制霸榜首
一文全览，知识图谱@AAAI 2020
Hinton AAAI2020 演讲全文：这次终于把胶囊网络做对了
AAAI 2020 论文集：AAAI 2020 论文解读会 @ 望京（附PPT下载）
AAAI 2020上的NLP有哪些研究风向？
微软 6 篇精选 AAAI 2020 论文
京东数科 6 篇精选AAAI 2020 论文
AAAI 2020 论文解读系列：
01. [中科院自动化所] 通过识别和翻译交互打造更优的语音翻译模型
02. [中科院自动化所] 全新视角，探究「目标检测」与「实例分割」的互惠关系
03. [北理工] 新角度看双线性池化，冗余、突发性问题本质源于哪里？
04. [复旦大学] 利用场景图针对图像序列进行故事生成
05. [腾讯 AI Lab] 2100场王者荣耀，1v1胜率99.8%，腾讯绝悟 AI 技术解读
06. [复旦大学] 多任务学习，如何设计一个更好的参数共享机制？
07. [清华大学] 话到嘴边却忘了？这个模型能帮你 | 多通道反向词典模型
08. [北航等] DualVD：一种视觉对话新框架
09. [清华大学] 借助BabelNet构建多语言义原知识库
10. [微软亚研] 沟壑易填：端到端语音翻译中预训练和微调的衔接方法
11. [微软亚研] 时间可以是二维的吗？基于二维时间图的视频内容片段检测
12. [清华大学] 用于少次关系学习的神经网络雪球机制
13. [中科院自动化所] 通过解纠缠模型探测语义和语法的大脑表征机制
14. [中科院自动化所] 多模态基准指导的生成式多模态自动文摘
15. [南京大学] 利用多头注意力机制生成多样性翻译
16. [UCSB 王威廉组] 零样本学习，来扩充知识图谱（视频解读）
17. [上海交大] 基于图像查询的视频检索，代码已开源！
18. [奥卢大学] 基于 NAS 的 GCN 网络设计（视频解读）
19. [中科大] 智能教育系统中的神经认知诊断，从数据中学习交互函数
20. [北京大学] 图卷积中的多阶段自监督学习算法
21. [清华大学] 全新模型，对话生成更流畅、更具个性化（视频解读，附PPT）
22. [华南理工] 面向文本识别的去耦注意力网络
23. [自动化所] 基于对抗视觉特征残差的零样本学习方法
24. [计算所] 引入评估模块，提升机器翻译流畅度和忠实度（已开源）
25. [北大&上交大] 姿态辅助下的多相机协作实现主动目标追踪
26. [快手] 重新审视图像美学评估 & 寻找精彩片段聚焦点
27. [计算所&微信AI] 改进训练目标，提升非自回归模型翻译质量（已开源）
28. [中科院&云从科技]：双视图分类，利用多个弱标签提高分类性能
29. [Sun Yat-sen] based tree strategy of gradual reinforcement learning
30. [Northeastern University] (video interpretation) based on neural machine translation union representation
31. [calculated] The evaluation module is introduced, and the hoisting machine translation fidelity fluency (Reading Video)
32. [Tsinghua University] snowball mechanisms for neural network (video interpretation) less time learning relationship
33. [calculated] non autoregressive machine translation (interpretation video) based on the training of the target n-tuple
34. [Tsinghua University] judicial examination data set (video interpretation, with download link)
35. [Automation] semantic segmentation weak supervision (Reading Video)
36. [Automation] sparse binary neural network, does not require any skill, to obtain accuracy SOTA (Reading Video)
37. [Branch & Ali] from the boundary to the text - A method of detecting a text of arbitrary shape
38. [cloud from the National Chiao Tung University & Technology] DCMN + model: crack "reading comprehension" problems, won the world's leading performance
39. [Automation & micro-channel AI] dual-channel multi-step reasoning model, better visual dialogue to solve the problems generated
40. [ETH Zurich] reverse R? Weaken the salient features of the upgrade is to bring fine-grained classification
41. [USTC] RiskOracle: a space-time fine-grained traffic accident prediction method
45. [Shang] New weak supervision object detection framework
46. [Fourth Pattern] NAS optimization algorithm 10 times faster!
47. [Oxford] use attention mechanism to achieve optimal positioning of the camera (video interpretation)
48. [Tianjin University] DIoU and CIoU: IoU correct in target detection Open
49. [Ningxia University] people-oriented large-scale video unlabeled face alignment method
50. [Shang Dynasty] KPNet, the pursuit of lightweight face detector (Video Interpretation)
