A Deep Sequential Model for Discourse
Parsing on Multi-Party Dialogues
Zhouxing Shi, Minlie Huang
Department of Computer Science & Technology, Tsinghua University, Beijing, China
Introduction
Task definition
•
Discourse parsing on a multi-party dialogue: to identify
dependency links and the corresponding relation types
{(
j
,
, r
j
) | j 6= } from a given dialogue segmented into a
sequence of Elementary Discourse Units(EDUs)
1
,
2
, · · · ,
n
.
Motivation
•
Parse the discourse structure of a multi-party dialogue which is
different from written text.
•
Construct a discourse structure incrementally by predicting
dependency relations and building the structure jointly and
alternately.
•
Predict dependency relations with not only local information, but
also global information that encodes the EDU sequence and the
discourse structure that is already built at the current step.
•
Build the discourse structure incrementally with a structured
encoder, using the predicted links and relation types.
Figure 1: A multi-party dialogue example with its discourse struc-
ture from the STAC Corpus.
Framework
•
The overall process:
1 Compute the non-structured representations of the EDUs with
hierarchical Gated Recurrent Unit (GRU) encoders, denoted as
h
and g
NS
.
2 Make a sequential scan of the EDUs, predicting dependency
relations and constructing the discourse structure.
•
Three steps when handling
:
1 Link prediction: predict the parent node p
of EDU
with a link
predictor.
2 Relation classification: predict the relation type between p
(assume p
=
j
) and
with a relation classifier.
3 Structured representation encoding: compute the structured
representation of
with a structured representation encoder.
•
Compute the structured representation of
: apply a structured
encoder to the path from the root to
. In practice, it is computed
incrementally.
•
Speaker highlighing mechanism (SHM): compute |A| different
structured representations for each EDU such that each one
highlights a specific speaker.
g
S
,
=
0 = 0
GRU
h
(g
S
j,
, h
⊕ r
j
)
= , > 0
GRU
gen
(g
S
j,
, h
⊕ r
j
)
6= , > 0
Figure 2: An example dependency tree (left) and the structured
encoder (right).
•
Input for link prediction and relation classification:
H
,j
= h
⊕ g
NS
⊕ g
NS
j
⊕ g
S
j,
Experiments
Model Link Link & Rel
MST 68.8 50.4
ILP 68.6 52.1
Deep+MST 69.6 52.1
Deep+ILP 69.0 53.1
Deep+Greedy 69.3 51.9
Deep Sequential (shared) 72.1 54.7
Deep Sequential 73.2 55.7
Table 1: F
1
scores (%) for different models.
Model Link Link & Rel
Deep+Greedy 69.3 51.9
Deep Sequential (NS) 71.0 53.7
Deep Sequential (Random) 71.8 53.7
Deep Sequential (w/o SHM) 71.7 54.5
Deep Sequential 73.2 55.7
Table 2: F
1
scores (%) for different models.
Figure 3: Illustration of the model which consists of modules for link prediction, relation classification, and structured representation encoding.
AAAI 2019 Contact: zhouxingshichn@gmail.com, aihuang@tsinghua.edu.cn
代码深渊漫步者
- 粉丝: 21
- 资源: 320
最新资源
- 苹果手机外壳贴麦拉机组装流道（sw16可编辑+工程图+bom）全套技术资料100%好用.zip
- MATLAB 实现基于PCNN（脉冲耦合神经网络）进行时间序列预测模型的项目详细实例（含完整的程序，GUI设计和代码详解）
- MATLAB 实现基于麻雀搜索算法（SSA）进行时间序列预测模型的项目详细实例（含完整的程序，GUI设计和代码详解）
- 基于改进剪枝算法的接触熟虾图像分割技术及其质量测量应用
- 音乐爬虫python源码分享
- MATLAB 实现基于移动平均模型（MA）进行时间序列预测模型的项目详细实例（含完整的程序，GUI设计和代码详解）
- Matlab实现基于NNMF+DBO+K-Medoids的数据聚类可视化的详细项目实例（含完整的程序，GUI设计和代码详解）
- 基于SSM的球鞋交易管理平台论文+Java-HTML+球鞋交易平台+毕设-课设均可
- 计算机视觉中高效曲线检测算法的研究与应用-随机化Hough变换的新方法实现线条、圆及椭圆识别
- Matlab基于ALO-SVR蚁狮优化支持向量回归的锂离子电池剩余寿命预测的详细项目实例（含完整的程序，GUI设计和代码详解）
- Matlab实现BiTCN双向时间卷积神经网络多变量时间序列预测的详细项目实例（含完整的程序，GUI设计和代码详解）
- 医学图像中基于判别广义霍夫变换的目标定位方法研究与应用
- 华强北悦虎耳机刷机固件包
- Matlab实现RIME-BP霜冰算法优化BP神经网络多变量回归预测的详细项目实例（含完整的程序，GUI设计和代码详解）
- 桌面作图软件CAD字体库
- MATLAB 实现基于VMD（变分模态分解）进行时间序列预测模型的项目详细实例（含完整的程序，GUI设计和代码详解）
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
评论0