Textbook Algorithmic Aspects in Information and Management 12Th International Conference Aaim 2018 Dallas TX Usa December 3 4 2018 Proceedings Shaojie Tang Ebook All Chapter PDF

Algorithmic Aspects in Information and
Management: 12th International

Conference, AAIM 2018, Dallas, TX,
USA, December 3–4, 2018, Proceedings
Shaojie Tang
Visit to download the full and correct content document:
https://textbookfull.com/product/algorithmic-aspects-in-information-and-management-
12th-international-conference-aaim-2018-dallas-tx-usa-december-3-4-2018-proceedin
gs-shaojie-tang/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Algorithmic Aspects in Information and Management: 12th

International Conference, AAIM 2018, Dallas, TX, USA,
December 3–4, 2018, Proceedings Shaojie Tang
https://textbookfull.com/product/algorithmic-aspects-in-
information-and-management-12th-international-conference-
aaim-2018-dallas-tx-usa-december-3-4-2018-proceedings-shaojie-
tang/
Brain Informatics International Conference BI 2018

Arlington TX USA December 7 9 2018 Proceedings Shouyi
Wang
https://textbookfull.com/product/brain-informatics-international-
conference-bi-2018-arlington-tx-usa-
december-7-9-2018-proceedings-shouyi-wang/
Algorithmic Aspects in Information and Management 14th

International Conference AAIM 2020 Jinhua China August
10 12 2020 Proceedings Zhao Zhang
https://textbookfull.com/product/algorithmic-aspects-in-
information-and-management-14th-international-conference-
aaim-2020-jinhua-china-august-10-12-2020-proceedings-zhao-zhang/
Combinatorial Optimization and Applications 12th

International Conference COCOA 2018 Atlanta GA USA
December 15 17 2018 Proceedings Donghyun Kim
https://textbookfull.com/product/combinatorial-optimization-and-
applications-12th-international-conference-cocoa-2018-atlanta-ga-
usa-december-15-17-2018-proceedings-donghyun-kim/
Combinatorial Optimization and Applications 14th
International Conference COCOA 2020 Dallas TX USA
December 11 13 2020 Proceedings Weili Wu
https://textbookfull.com/product/combinatorial-optimization-and-
applications-14th-international-conference-cocoa-2020-dallas-tx-
usa-december-11-13-2020-proceedings-weili-wu/
Information Systems Security: 14th International

Conference, ICISS 2018, Bangalore, India, December
17-19, 2018, Proceedings Vinod Ganapathy
https://textbookfull.com/product/information-systems-
security-14th-international-conference-iciss-2018-bangalore-
india-december-17-19-2018-proceedings-vinod-ganapathy/
Neural Information Processing: 25th International

Conference, ICONIP 2018, Siem Reap, Cambodia, December
13–16, 2018, Proceedings, Part II Long Cheng
https://textbookfull.com/product/neural-information-
processing-25th-international-conference-iconip-2018-siem-reap-
cambodia-december-13-16-2018-proceedings-part-ii-long-cheng/

13–16, 2018, Proceedings, Part III Long Cheng
cambodia-december-13-16-2018-proceedings-part-iii-long-cheng/

13–16, 2018, Proceedings, Part V Long Cheng
cambodia-december-13-16-2018-proceedings-part-v-long-cheng/
Shaojie Tang · Ding-Zhu Du
David Woodruff · Sergiy Butenko (Eds.)
Algorithmic Aspects
LNCS 11343
in Information
and Management
12th International Conference, AAIM 2018
Dallas, TX, USA, December 3–4, 2018
Proceedings
123
Lecture Notes in Computer Science 11343
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology Madras, Chennai, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7407
Shaojie Tang Ding-Zhu Du
•
David Woodruff Sergiy Butenko (Eds.)

•
Algorithmic Aspects
in Information
and Management
12th International Conference, AAIM 2018
Dallas, TX, USA, December 3–4, 2018
Proceedings
123
Editors
Shaojie Tang David Woodruff
The University of Texas at Dallas University of California, Davis
Richardson, TX, USA Davis, CA, USA
Ding-Zhu Du Sergiy Butenko
University of Texas at Dallas Texas A&M University
Richardson, TX, USA College Station, TX, USA
ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-030-04617-0 ISBN 978-3-030-04618-7 (eBook)
https://doi.org/10.1007/978-3-030-04618-7
Library of Congress Control Number: 2018963820
LNCS Sublibrary: SL1 – Theoretical Computer Science and General Issues
© Springer Nature Switzerland AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
AAIM 2018 Co-chairs’ Message

On behalf of the Organizing Committees, it is our pleasure to welcome you to the
proceedings of the AAIM 2018, the 12th International Conference on Big Data
Computing and Communications, held in Dallas, USA, during December 3–4, 2018.
All submissions to this conference were carefully peer-reviewed, and were ranked
according to their original contribution, quality, presentation, and relevance to the
conference. Based on the review, 25 high-quality papers were accepted for presentation
and inclusion in the proceedings. We aim to attract and bring together researchers
working on algorithms, data structures, and their applications to share their latest
research results.
The organization of this conference would have been impossible without the
valuable help of many dedicated people. First of all, we would like to take this
opportunity to thank all the authors for their submissions. We also kindly thank all the
members of the Technical Program Committee who shared their valuable time and
made a great effort in providing us reviews in time. We hope you find the proceedings
interesting.
September 2018 Shaojie Tang

Ding-Zhu Du
David Woodruff
Sergiy Butenko
Organization
General Chairs
Ding-Zhu Du University of Texas at Dallas, USA
David Woodruff University of California, Davis, USA
Program Chairs
Shaojie Tang University of Texas at Dallas, USA
Sergiy Butenko Texas A&M University, USA
Technical Program Committee

Ding-Zhu Du University of Texas at Dallas, USA
Minghui Jiang Utah State University, USA
Zhao Zhang Zhejiang Normal University, China
Romeo Rizzi University of Verona, Italy
Michael Fellows University of Bergen, Norway
Jinhui Xu State University of New York at Buffalo, USA
Marek Chrobak University of California, Riverside, USA
My Thai University of Florida, USA
Guohui Lin University of Alberta, Canada
Binhai Zhu Montana State University, USA
Bhaskar Dasgupta University of Illinois at Chicago, USA
Anthony Bonato Ryerson University, USA
Chuangyin Dang City University of Hong Kong, SAR China
Kazuo Iwama Kyoto University, Japan
Binay Bhattacharya Simon Fraser University, Canada
Bo Chen University of Warwick, UK
Peng-Jun Wan Illinois Institute of Technology, USA
Leizhen Cai Chinese University of Hong Kong, SAR China
Dennis Komm ETH Zürich, Switzerland
Hans-Joachim ETH Zürich, Switzerland
Boeckenhauer
Jun Pei Hefei University of Technology, China
Dalila Fontes University of Porto, Portugal
Minming Li City University of Hong Kong, SAR China
Shaojie Tang University of Texas at Dallas, USA
Contents
Minimum Diameter k-Steiner Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Wei Ding and Ke Qiu
Factors Impacting the Label Denoising of Neural Relation Extraction . . . . . . 12

Tingting Sun, Chunhong Zhang, and Yang Ji
Makespan Minimization on Unrelated Parallel Machines with a Few Bags . . . 24

Daniel R. Page and Roberto Solis-Oba
Channel Assignment with r-Dynamic Coloring . . . . . . . . . . . . . . . . . . . . . . 36

Junlei Zhu and Yuehua Bu
Profit Maximization Problem with Coupons in Social Networks . . . . . . . . . . 49

Bin Liu, Xiao Li, Huijuan Wang, Qizhi Fang, Junyu Dong, and Weili Wu
A Bicriteria Approximation Algorithm for Minimum Submodular Cost

Partial Multi-Cover Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Yishuo Shi, Zhao Zhang, and Ding-Zhu Du
A Novel Approach to Verifying Context Free Properties of Programs . . . . . . 74

Nan Zhang, Zhenhua Duan, Cong Tian, and Hongwei Du
Determination of Dual Distances for a Kind of Perfect Mixed Codes. . . . . . . 88

Tianyi Mao
Approximation and Competitive Algorithms for Single-Minded

Selling Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Francis Y. L. Chin, Sheung-Hung Poon, Hing-Fung Ting, Dachuan Xu,
Dongxiao Yu, and Yong Zhang
An Empirical Analysis of Feasibility Checking Algorithms

for UTVPI Constraints. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
K. Subramani, Piotr Wojciechowski, Zachary Santer,
and Matthew Anderson
Quality-Aware Online Task Assignment Using Latent Topic Model . . . . . . . 124

Yang Du, Yu-E Sun, He Huang, Liusheng Huang, Hongli Xu,
and Xiaocan Wu
Calibration Scheduling with Time Slot Cost . . . . . . . . . . . . . . . . . . . . . . . . 136

Kai Wang
The k-power Domination Problem in Weighted Trees . . . . . . . . . . . . . . . . . 149

ChangJie Cheng, Changhong Lu, and Yu Zhou
VIII Contents
General Rumor Blocking: An Efficient Random Algorithm

with Martingale Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Qizhi Fang, Xin Chen, Qingqin Nong, Zongchao Zhang,
Yongchang Cao, Yan Feng, Tao Sun, Suning Gong, and Ding-Zhu Du
A Robust Power Optimization Algorithm to Balance Base Stations’ Load

in LTE-A Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Jihong Gui, Wenguo Yang, Suixiang Gao, and Zhipeng Jiang
Faster Compression of Patterns to Rectangle Rule Lists . . . . . . . . . . . . . . . . 190

Ian Albuquerque Raymundo Da Silva, Gruia Calinescu,
and Nathan De Graaf
Algorithm Designs for Dynamic Ridesharing System. . . . . . . . . . . . . . . . . . 209

Chaoli Zhang, Jiapeng Xie, Fan Wu, Xiaofeng Gao, and Guihai Chen
New LP Relaxations for Minimum Cycle/Path/Tree Cover Problems . . . . . . . 221

Wei Yu, Zhaohui Liu, and Xiaoguang Bao
Computation of Kullback-Leibler Divergence Between Labeled Stochastic

Systems with Non-identical State Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Krishnendu Ghosh
Order Preserving Barrier Coverage with Weighted Sensors on a Line . . . . . . 244

Robert Benkoczi, Daya Gaur, and Xiao Zhang
Achieving Location Truthfulness in Rebalancing Supply-Demand

Distribution for Bike Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Hongtao Lv, Fan Wu, Tie Luo, Xiaofeng Gao, and Guihai Chen
Approximation Algorithms and a Hardness Result for the Three-Machine

Proportionate Mixed Shop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Longcheng Liu, Guanqun Ni, Yong Chen, Randy Goebel, Yue Luo,
An Zhang, and Guohui Lin
A New Algorithm Design Technique for Hard Problems, Building

on Methods of Complexity Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
András Faragó
Community-Based Acceptance Probability Maximization for Target Users

on Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Ruidong Yan, Yuqing Zhu, Deying Li, and Yongcai Wang
Knowledge Graph Embedding Based on Subgraph-Aware Proximity . . . . . . . 306

Xiao Han, Chunhong Zhang, Chenchen Guo, Tingting Sun, and Yang Ji
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319

Minimum Diameter k-Steiner Forest
Wei Ding1(B) and Ke Qiu2

1
Zhejiang University of Water Resources and Electric Power, Hangzhou 310018,
Zhejiang, China
dingweicumt@163.com
2
Department of Computer Science, Brock University, St. Catharines, Canada
kqiu@brocku.ca
Abstract. Given an edge-weighted undirected graph G = (V, E, w) and

a subset T ⊆ V of p terminals, a k-Steiner forest spanning all the ter-
minals in T includes k branches, where every branch is a Steiner tree.
The diameter of a k-Steiner forest is referred to as the maximum dis-
tance between two terminals of a branch. This paper studies the mini-
mum diameter k-Steiner forest problem (MDkSFP) and establishes the
relationship between MDkSFP and the absolute Steiner k-center problem
(ASkCP). We first obtain a 2-approximation to ASkCP by a dual approx-
imation algorithm and then achieve a 2-approximation to MDkSFP. Fur-
ther, we achieve a (better) 2ρ-approximation to MDkSFP, where ρ < 1
in general, by modifying the sites of centers and re-clustering all the
terminals.
Keywords: k-Steiner forest · Diameter · Modification
1 Introduction
Given an undirected graph G = (V, E) and a subset T ⊆ V of terminals, a
Steiner tree is referred to as a connected subgraph of G that spans all the termi-
nals in T [6]. A Steiner forest is a group of disjoint connected branches spanning
T , where every branch is a Steiner tree [1,2,5,7,11–14]. They have a wide variety
of applications in communication networks, computational biology, and etc. [6].
1.1 Related Results

Let G = (V, E, w) be an edge-weighted undirected graph, where V is the set of n
vertices, E is the set of m edges, and w(·) is an edge weight function w : E → R+ .
The cost of a Steiner forest is equal to the sum of all the weights of edges used
by the Steiner forest. Given a subset T ⊆ V of p terminals and an integer,
1 ≤ k ≤ p, the k-Steiner forest problem (kSFP) asks to find a Steiner forest
of minimum cost which contains at most k branches. In [11], Ravi presented a
primal-dual approximation algorithm with a factor of 2(1 − p−k+1 1
) for kSFP.
The generalized Steiner tree problem (GSTP) (also called the Steiner
forest problem (SFP)) is another version of Steiner forest problem. Given a
c Springer Nature Switzerland AG 2018
S. Tang et al. (Eds.): AAIM 2018, LNCS 11343, pp. 1–11, 2018.
https://doi.org/10.1007/978-3-030-04618-7_1
2 W. Ding and K. Qiu
collection D = {(si , ti ) : 1 ≤ i ≤ K} of K terminal pairs, the aim of GSTP is

to find an edge subset F ⊆ E of minimum cost such that F contains an si -ti
path. Agrawal et al. [1] gave the first approximation algorithm with a factor of
2(1− p1 ), where p is the number of terminals. Refer readers to [2,13,14] for related
results in special graphs. In [7], Hajiaghayi and Jain proposed the Steiner k-
forest problem (SkFP), which aims to find a subgraph of minimum cost such
that at least k ≤ K terminal pairs of D are connected. In [12], Segev and Segev
gave an O(n2/3 log K)-approximation algorithm by using Lagrangian relaxation
technique. Very recently, Dinitz √ [5] developed the best known approximation
1 √
algorithm with a factor of n 3 (7−4 2)+ , where 13 (7 − 4 2) < 0.44772 and > 0
is an arbitrary constant.
The diameter of a tree or forest is referred to as the longest distance between
vertices (leaves) on the tree or forest. Hassin and Tamir [8] sought a minimum
diameter spanning tree problem (MDSTP) of G as well as the mini-
mum diameter k-forest problem (MDkFP). In [3], Bui et al. presented a
distributed algorithm for MDSTP, with time complexity of O(n) and message
complexity of O(mn). In [4], Ding and Qiu proposed the minimum diameter
Steiner tree (MDSTT), and gave an O(n log p)-time 2-approximation algo-
rithm and O(mn + n2 log n)-time exact algorithm.
1.2 Our Results
In the real-world problems, we are sometimes required to find a Steiner forest

spanning a subset of terminals with the aim of minimizing the maximum diame-
ter of all the branches. For instance, in a communication network with every link
having a delay, a set of clients are clustered into k groups arbitrarily and every
group of clients are interconnected by a Steiner tree, resulting in a k-Steiner
forest. The goal is to make the maximum delay between two clients as small as
possible. This problem can be modelled as the minimum diameter k-Steiner
forest problem (MDkSFP), which is formally defined in Sect. 2.1. To the best
of our knowledge, this paper is the first one to propose MDkSFP formally.
We first establish the relationship between MDkSFP and the absolute
Steiner k-center problem (ASkCP). ASkCP is NP-hard. We use the dual
approximation framework [9,10] to design a 2-approximation algorithm for
ASkCP. By clustering all the terminals into k groups with the 2-approximation
as the centers, we design an O(mp2 log m)-time 2-approximation algorithm for
MDkSFP. Further, we obtain a better approximation to ASkCP by modification.
By re-clustering all the terminals with the new approximation as the centers,
we develop an O(mp2 log m + n2 p)-time 2ρ-approximation algorithm, where ρ is
the ratio of the diameter of the latter approximation to MDkSFP over that of
the former.
Organization. The rest of this paper is organized as follows. In Sect. 2, we
define ASkCP and MDkSFP formally, and also show the relationship between
ASkCP and MDkSFP. In Sect. 3, we give a dual approximation algorithm for
ASkCP. In Sect. 4, we first design a 2-approximation algorithm for MDkSFP and
Minimum Diameter k-Steiner Forest 3
further develop a better approximation algorithm by modification. In Sect. 5, we

conclude the paper.
2 Preliminaries
2.1 Definitions and Notations
Let G = (V, E, w) be an edge-weighted undirected graph, where V =

{v1 , v2 , . . . , vn } is the set of n vertices, E is the set of m edges, and w : E → R+
is a weight function on edges. Let T = {t1 , t2 , . . . , tp } ⊆ V be a subset of p
terminals in V . For any e ∈ E, we use P(e) to denote the set of continuum
points on e. Let P be the set of all the continuum points on edges of G. So,
P = e∈E P(e). For any pair of vertices, vi , vj ∈ V , we use d(vi , vj ) to denote the
vi -vj shortest path distance (SPD). Obviously, d(·) is a metric, i.e., d(vi , vi ) = 0
and d(vi , vk ) ≤ d(vi , vj )+d(vj , vk ), ∀i, j, k. Let M = (d)n×n be the n×n distance
matrix of shortest paths in G. We assume that M is available in this paper. Note
that we also use d(x, y) to denote the x-y SPD, for any two points, x and y. A
set of cardinality of k is called a k-set.
Given a k-set, X = {x1 , x2 , . . . , xk } ⊂ P, the distance from terminal ti to X ,
d(ti , X ), is referred to as the distance from ti to the closest facility in X , for any
1 ≤ i ≤ p. So, d(ti , X ) = min1≤j≤k d(ti , xj ). The k-set, X , is called a absolute
Steiner k-center (ASkC). The maximum distance from terminals to X (ASkC) is
called the k-radius from T to X , denoted by r(T , X ). Let X ∗ = {x∗1 , x∗2 , . . . , x∗k }
denote an optimal ASkC, which minimizes the k-radius, and Cj be the subset of
terminals are assigned to x∗j (i.e., x∗j is the closest facility). We have
r(T , X ) = max d(ti , X ). (1)

1≤i≤p
and
r(T , X ∗ ) = min r(T , X ). (2)
X ⊂P
This paper first deals with the following problem.
Definition 1. Given an edge-weighted undirected graph G = (V, E, w), a sub-

set T ⊆ V of p terminals and a positive integer k ≥ 1, the goal of abso-
lute Steiner k-center problem (ASkCP) is to find an optimal k-set, X ∗ =
{x∗1 , x∗2 , . . . , x∗k } ⊂ P, to minimize the k-radius.
The discrete (vertex) Steiner k-center problem (DSkCP) is the special

case of ASkCP, where the facilities are restricted to the vertices. Similarly, we
can define the discrete Steiner k-center (DSkC). Obviously, the ordinary abso-
lute k-center problem (AkCP) is the special case of ASkCP with T = V .
Besides, DS1C and AS1C are the special cases of DSkC and ASkC with k = 1,
respectively.
Let F denote a k-Steiner forest spanning all the terminals in T , and Fj denote
k
the j-th branch (Steiner tree). So, F = j=1 Fj . Note that the k branches are
both edge-disjoint and vertex-disjoint. One is called a pseudo Steiner forest if it

has two branches having at least a common edge or vertex.
The diameter of branch Fj , diam(Fj ), is referred to as the longest distance
between leaves of Fj . As every branch Fj is a Steiner tree, its diameter is also
the longest distance between terminals. Let dj (ti1 , ti2 ) denote the distance on
branch Fj between ti1 and ti2 . The diameter of a k-Steiner forest F, diam(F),
is referred to as the maximum diameter of branches. We have
diam(Fj ) = max dj (ti1 , ti2 ). (3)

ti1 ,ti2 ∈Fj ,ti1 =ti2
and
diam(F) = max diam(Fj ). (4)
1≤j≤k
Since ASkCP and MDkSFP are trivial when p = 2, we focus on the cases
with p ≥ 3. On the other hand, the optimum is obvious (every branch contains
one terminal) when k = p. Therefore, we focus on the cases where k is (quite)
smaller than p in general.
Definition 2. Given an edge-weighted undirected graph G = (V, E, w), a subset

T ⊆ V of p terminals and a positive integer 1 ≤ k ≤ p, the minimum diame-
ter k-Steiner forest problem (MDkSFP) asks to find an optimal k-Steiner
forest, F ∗ , such that the diameter is minimized.
2.2 Fundamental Properties

Theorem 1 shows a sufficient condition of k-Steiner forest, which gives a method
of constructing a k-Steiner forest while avoiding a pseudo forest.
Theorem 1. Given a k-set with no duplicates, X = {x1 , x2 , . . . , xk } ⊂ P, we

let Tj ⊆ T be the set of terminals, which are closest to xj , for all 1 ≤ j ≤ k.
One shortest path tree (SPT) spanning Tj1 with xj1 as the origin and the
other SPT spanning Tj2 with xj2 as the origin are both edge-disjoint and vertex-
disjoint, for any j1 = j2 .
Theorem 2 is a necessary condition of minimum diameter k-Steiner forest,

which shows the relationship between MDkSFP and ASkCP. The proof of Theo-
rem 2 implies we can obtain a minimum diameter k-Steiner forest by first finding
an optimal ASkC and then computing a collection of SPT’s with the optimal
ASkC as the origins. First, we give Lemma 1.
Lemma 1. There are surely at least one facility x∗j ∗ ∈ X ∗ and at least two
terminals, ti∗1 and ti∗2 , such that d(ti∗1 , x∗j ∗ ) = d(ti∗2 , x∗j ∗ ) = r(T , X ∗ ).
Theorem 2. diam(F ∗ ) = 2 · r(T , X ∗ ).

3 Approximation to ASkCP
In this section, we study ASkCP in edge-weighted graph G = (V, E, w), and use
the dual approximation framework [9,10] to devise an approximation algorithm
with a factor of 2. Let opt be the optimal value (of k-radius), X ∗ and X A denote
an optimal solution and an algorithm solution with no duplicate, respectively.
Fig. 1. Illustration for the proof of Theorem 3.
Procedure TEST1 plays an important role, where k terminals are selected

from T as a possible solution. Let X (Δ) denote the solution generated by
TEST1 (Δ), where Δ > 0 is a given real number, and T record the unlabelled
terminals. Initially, set X (Δ) = ∅ and T = T . Every time one terminal t is
selected arbitrarily as a facility, all unlabelled terminals t ∈ T satisfying that
Procedure TEST1 (Δ):

Input: a graph G = (V, E, w) and distance matrix M ,
a set T ⊂ V of terminals, a positive real number Δ > 0;
Output: NO or YES with X (Δ).
Step 1: X (Δ) ← ∅; T ← T ;
Step 2: while |X (Δ)| < k and T = ∅ do
Select a terminal t arbitrarily from T ;
L(t ) ← {t ∈ T : d(t, t ) ≤ 2Δ};
T ← T \ L(t ); X (Δ) ← X (Δ) ∪ {t };
end
Step 3: if T = ∅ then Return NO;
else
if |X (Δ)| = k then
Let X0 be a set of k − |X (Δ)| terminals,
selected arbitrarily from T \ X (Δ);
X (Δ) ← X (Δ) ∪ X0 ;
endif Return YES and X (Δ);
endif
d(t, t ) ≤ 2Δ are labelled. This operation is repeated until |X (Δ)| = k or T = ∅.

If TEST1 ends with T = ∅, then TEST1 fails and returns NO. Otherwise, TEST1
can find a solution in the following way, and so returns YES. Since the solution is
a k-set, TEST1 needs to judge whether |X (Δ)| = k or not. If the former occurs,
then a set X0 of k − |X (Δ)| terminals are selected arbitrarily from T \ X (Δ), and
the union of X0 and X (Δ) is output as a solution. Otherwise, it returns X (Δ).
Theorem 3. Given a graph G = (V, E, w) and a subset T ⊂ V of terminals,

if there is a k-set U ⊂ V with r(T , U) ≤ Δ, then TEST1 (Δ) can find a k-set
X ⊂ V with r(T , X ) ≤ 2Δ in O(pk) time, for any Δ > 0 (see Fig. 1).
The main idea of our algorithm is to find an approximation, a k-set of T ,

to ASkCP in vertex-unweighted graphs, which is based on the fact that every
terminal can act as a facility. Based on TEST1 , we design our approximation
algorithm ASkCP-ALG for ASpCP. By Theorem 3, we know that TEST1 (opt)
returns YES as X ∗ satisfies that r(T , X ∗ ) ≤ opt. However, we do not know the
value of opt beforehand. In order to find as small value of k-radius as possible
which makes TEST1 return YES, we need to get all the possible values of k-
radius. By Lemma 1, we have to discuss all the possible combinations of the
edge containing x∗j ∗ and two terminals, ti∗1 and ti∗2 . Given an edge e = {v , v }
and two terminals, ti1 and ti2 , the corresponding two possible values of k-radius
are, see Fig. 2,
1
r = (d(ti1 , v ) + w(v , v ) + d(ti2 , v )) , (5)
2
and
1
r = (d(ti1 , v ) + w(v , v ) + d(ti2 , v )) . (6)
2

So, the number of possible values of k-radius is 2 · m p2 = mp(p − 1). By deleting
duplicates of all the possible values and then arranging the values left into an
increasing sequence, l1 < l2 < · · · < lλ , where λ ≤ mp(p − 1), Step 1 of ASkCP-
ALG can be done in O(mp2 log m) time by sorting.
Fig. 2. Two possible values of k-radius for given ti1 , ti2 and edge e = {v , v }.
Obviously, TEST1 (lλ ) returns YES. Our task is to find the smallest one, ls∗ ,
from all the possible values ls , 1 ≤ s ≤ λ which make TEST1 (ls ) return YES.
In Step 2, we find ls∗ by applying a binary search to ls , 1 ≤ s ≤ λ. So, it is
sufficient to apply TEST1 O(log λ) times. Recall that applying TEST1 once takes
at most O(pk) time. Since λ ≤ mp(p − 1) and p ≤ n ≤ m, Step 2 takes at most
O(pk log m) time. Therefore, ASkCP-ALG takes at most O(mp2 log m) time. The
details of Step 2 are presented as follows. Our search begins with l1 . If TEST1 (l1 )
outputs YES, then we claim that ls∗ = l1 and get an approximation X A = X (l1 ).
Otherwise, we find ls∗ by applying a binary search to l1 , l2 , . . . , lλ . Let LB and
UB be the lower bound and upper bound on the index s∗ , respectively, and let
M = LB+UB 2 . Initially, set LB = 1 and UB = λ. If TEST1 (lM ) outputs NO,
then M becomes a new lower bound on s∗ and UB is also its upper bound. If
TEST1 (lM ) outputs YES, then M becomes a new upper bound on s∗ and LB is
also its lower bound. Above operation is repeated until LB and UB become two
consecutive integer numbers. During the whole operation, TEST1 (lLB ) always
outputs NO while TEST1 (lUB ) always outputs YES. As a result, the final UB
is s∗ . The binary search ends with an approximation, X A = X (ls∗ ), to ASkCP
in a vertex-unweighted graph G. According to above discussions, we obtain the
following theorem.
Algorithm ASkCP-ALG:
a set T ⊂ V of terminals;
Output: a k-set X A .
Step 1: Delete duplicates of all the possible values of
k-radius, and then arrange the values left into
an increasing sequence, l1 < l2 < · · · < lλ .
Step 2: if TEST1 (l1 ) = YES then Return X (l1 );
else LB ← 1; UB ← λ;
while UB − LB = 1 do
M ← LB+UB2
;
if TEST1 (lM ) = NO then LB ← M;
else UB ← M; endif
end Return X (lUB );
endif
Theorem 4. Given a graph G = (V, E, w) with n vertices and m edges, and

a set T ⊂ V of p terminals, ASkCP-ALG is a 2-approximation algorithm for
ASkCP in G, with time complexity of O(mp2 log m).
4 Approximations to MDkSFP
In this section, we design a 2-approximation algorithm for MDkSFP, and further
develop an improved approximation algorithm with a factor of 2ρ, where ρ < 1
in general, based on the approximation to ASkCP.
4.1 A 2-Approximation Algorithm

In this subsection, we first get a better solution to the vertex-unweighted ASkCP
by modifying the algorithm solution of ASkCP-ALG, and then design an approx-
imation algorithm for MDkSFP based on the modification.
The algorithm solution, X A = {t1 , t2 , . . . , tk }, produced by ASkCP-ALG is a
set of k terminals. Let Cj , 1 ≤ j ≤ k be the set of terminals which are labelled
when tj is selected as a facility. In details, we cluster all the terminals with
t1 , t2 , . . . , tk as centers. Let CjA denote the cluster of terminals to which tj is the
closest facility. Note that CjA includes tj . The maximum distance to tj in CjA
(resp. Cj ) is denoted by r(CjA ) (resp. r(Cj )). Lemma 2 implies that the multiset
{CjA : 1 ≤ j ≤ k} is better than the multiset {Cj : 1 ≤ j ≤ k} with respect to
the value of k-radius.
Lemma 2. max1≤j≤k r(CjA ) ≤ max1≤j≤k r(Cj ).
Let FjA be the SPT (branch) spanning CjA with ti as the origin. By Theo-
rem 1 and the definition of CjA , we claim that all FjA , ∀j are both edge-disjoint
k
and vertex-disjoint. Hence, the collection of j=1 FjA forms a Steiner forest
(algorithm solution), denoted by F A . This leads to an approximation algorithm
MDkSFP-ALG for MDkSFP. Its performance analysis is shown in Theorem 5. Let
ji be the index such that d(ti , tji ) = min1≤j≤k d(ti , tj ), for any 1 ≤ i ≤ p.
Algorithm MDkSFP-ALG:
Output: a Steiner forest F A .
Step 1: Call ASkCP-ALG to get X A = {t1 , t2 , . . . , tk };
CjA ← ∅, 1 ≤ j ≤ k;
for i := 1 to p do
ji ← arg min1≤j≤k d(ti , tj );
CjAi ← CjAi ∪ {i};
end
Step 2: for j := 1 to k do
Compute a SPT (branch) FjA spanning CjA ;
end

F A ← kj=1 FjA ; Return F A ;
Theorem 5. Given an edge-weighted graph G = (V, E, w) with n vertices and

m edges, and a set T ⊂ V of p terminals, MDkSFP-ALG can compute a 2-
approximation to MDkSFP in G within O(mp2 log m) time.
4.2 A Better Approximation Algorithm
The 2-approximation algorithm MDkSFP-ALG clusters all the terminals with

the algorithm solution of ASkCP-ALG as centers. The algorithm solution is
a set of k terminals, X A = {t1 , t2 , . . . , tk }. However, the optimal solution,
X ∗ = {x∗1 , x∗2 , . . . , x∗k }, is a set of k points. In this subsection, we improve the
performance
of the solution to some k extent by modifying
k k terminals to k points.
Let CjA = pj , ∀j. So, T = j=1 CjA and p = j=1 pj . For any CjA , we use
Ding and Qiu’s algorithm [4] to compute the MDST of CjA , where CjA is the set
of terminals. By Lemma 3, we claim that the SPT spanning CjA with the AS1C,
xj , as the origin is just the MDST of CjA , while the SPT spanning CjA with the
DS1C, tj , as the origin is a 2-approximate MDST of CjA .
Lemma 3 (see [4]). Given a graph G = (V, E, w) and a subset T ⊂ V of

terminals, the MDST is the SPT spanning T with the AS1C as the origin. In
addition, the SPT spanning T with the DS1C as the origin is a 2-approximate
MDST where the factor of 2 is tight.
The center of CjA is perturbed from tj to xj , for all 1 ≤ j ≤ k, resulting
in a better solution to ASkCP, denoted by X B = {x1 , x2 , . . . , xk }. It is certain
that tj is closest to all the terminals in CjA , but it is uncertain that xj is closest
to all of them. Therefore, if the MDST (a.k.a., the SPT with xj as the origin)
spanning CjA is taken as one branch, then two different branches may have at
least a common edge. In other words, the resulting solution may be a pseudo
Steiner forest. Let FjP denote the MDST spanning CjA , and F P denote the union
of all FjP , ∀j.
In order to avoid a pseudo solution, we re-cluster all the terminals with X B
as centers. Let CjB be the cluster of terminals to which xj is the closest facility,
and let r(CjB ) denote the maximum distance to xj in CjB . The SPT spanning
CjB with xj as the origin is taken as one branch, denoted by FjB . By Theorem 1,
we conclude from the definition of CjB that all FjB , 1 ≤ j ≤ k are both edge-
disjoint and vertex-disjoint. As a consequence, all the branches FjB , ∀j form a
Steiner forest, denoted by F B . Based on above discussions, we devise an improved
approximation algorithm MDkSFP-IMP for MDkSFP. Its performance analysis
is shown in Theorem 6. For any 1 ≤ s ≤ p, we let js be the index such that
d(ts , xjs ) = min1≤j≤k d(ts , xj ).
Algorithm MDkSFP-IMP:
Output: a Steiner forest F B .
Step 1: The same as MDkSFP-ALG;
Use the Ding and Qiu’s algorithm in [4] to
compute the AS1C, xj , of CjA ;
end
CjB ← ∅, 1 ≤ j ≤ k;
for s := 1 to p do
js ← arg min1≤j≤k d(ts , xj );
CjBs ← CjBs ∪ {s};
end
Compute a SPT (branch) FjB spanning CjB ;
end

F B ← kj=1 FjB ; Return F B ;
Theorem 6. Given a graph G = (V, E, w) with n vertices and m edges, and a

set T ⊂ V of p terminals, MDkSFP-IMP produces a possible better approximation
than MDkSFP-ALG within O(mp2 log m + n2 p) time.
The proof of Theorem 6 shows diam(F B ) ≤ diam(F P ) ≤ diam(F A ).

MDkSFP-IMP can produce the exact values of diam(FjA ) and diam(FjB ), for
∀j, and further the values of diam(F A ) and diam(F B ). Let
diam(F B )
ρ= . (7)
diam(F A )
Clearly, ρ ≤ 1. Recall that diam(F A ) ≤ 2 · diam(F ∗ ). Therefore,
diam(F B ) = ρ · diam(F A ) ≤ 2ρ · diam(F ∗ ).
Corollary 1. The approximation factor of MDkSFP-IMP is 2ρ, where ρ is the

ratio of diam(F B ) over diam(F A ).
In most cases, ρ < 1. This is because ρ = 1 if and only if the center (point)
corresponding to the k-radius is at a terminal (vertex). Therefore, the approxi-
mation of MDkSFP-IMP improves that of MDkSFP-ALG by a factor of 2(1 − ρ).
5 Conclusions
This paper studied MDkSFP in undirected graphs and established the rela-
tionship between MDkSFP and ASkCP. First, we obtained an approximate
ASkC and then designed a 2-approximation algorithm for MDkSFP. Further,
we achieved a better approximate ASkC by modification, and then developed
a 2ρ-approximation algorithm for MDkSFP, where ρ < 1 in general. Our algo-

rithm can be adapted to the version where the number of the branches of Steiner
forest is at most k instead of equal to k.
The performance ratio of the approximate ASkC is one of the major factors
that influence the performance ratio of our algorithm solution of MDkSFP. We
suggest to improve the approximation to MDkSFP by achieving a better approx-
imation to ASkCP, and conjecture that ASkCP could admit a ρ-approximation
algorithm with ρ < 2.
References
1. Agrawal, A., Klein, P., Ravi, R.: When trees collide: an approximation algorithm
for the generalized Steiner problem in networks. SIAM J. Comput. 24(3), 440–456
(1995)
2. Bateni, M., Hajiaghayi, M., Marx, D.: Approximation schemes for Steiner forest
on planar graphs and graphs of bounded treewidth. J. ACM 58(5), 1–21 (2011)
3. Bui, M., Butelle, F., Lavault, C.: A distributed algorithm for constructing a mini-
mum diameter spanning tree. J. Parallel Distrib. Comput. 64(5), 571–577 (2004)
4. Ding, W., Qiu, K.: Algorithms for the minimum diameter terminal Steiner tree
problem. J. Comb. Optim. 28(4), 837–853 (2014)
5. Dinitz, M., Kortsarz, G., Nutov, Z.: Improved approximation algorithm for Steiner
k-forest with nearly uniform weights. In: 17th APPROX, pp. 115–127 (2014)
6. Du, D., Hu, X.: Steiner Tree Problems in Computer Communication Networks.
World Scientific Publishing Co., Pte. Ltd., Singapore (2008)
7. Hajiaghayi, M., Jain, K.: The prize-collecting generalized steiner tree problem via
a new approach of primal-dual schema. In: Proceedings of the 17th SODA, pp.
631–640 (2006)
8. Hassin, R., Tamir, A.: On the minimum diameter spanning tree problem. Inf.
Process. Lett. 53, 109–111 (1995)
9. Hochbaum, D.S., Shmoys, D.B.: A best possible heuristic for the k-center problem.
Math. Oper. Res. 10(2), 180–184 (1985)
10. Plesnı́k, J.: A heuristic for the p-center problem in graphs. Discrete Appl. Math.
17, 263–268 (1987)
11. Ravi, R.: A primal-dual approximation algorithm for the Steiner forest problem.
Inf. Process. Lett. 50(4), 185–189 (1994)
12. Segev, D., Segev, G.: Approximate k-Steiner forests via the Lagrangian relaxation
technique with internal preprocessing. Algorithmica 56, 529–549 (2010)
13. Winter, P.: Generalized Steiner problem in outerplanar graphs. BIT Numer. Math.
25(3), 485–496 (1985)
14. Winter, P.: Generalized Steiner problem in series-parallel networks. J. Algorithms
7(4), 549–566 (1986)
Factors Impacting the Label Denoising
of Neural Relation Extraction
Tingting Sun(B) , Chunhong Zhang, and Yang Ji
Beijing University of Posts and Telecommunications, Beijing 100876, China

{suntingting,zhangch,jiyang}@bupt.edu.cn
Abstract. The goal of relation extraction is to obtain relational facts

from plain text, which can benefit a variety of natural language process-
ing tasks. To address the challenge of automatically labeling large-scale
training data, a distant supervision strategy is introduced to relation
extraction by heuristically aligning entity pairs in plain text with the
knowledge base. Unfortunately, the method is vulnerable to the noisy
label problem due to the incompletion of the exploited knowledge base.
Existing works focus on the specific algorithms, but few works summarize
the commonalities between different methods and the influencing factors
of these denoising mechanisms. In this paper, we propose three main
factors that impact the label denoising of distantly supervised relation
extraction, including labeling assumption, prior knowledge and confi-
dence level. In order to analyze how these factors influence the denoising
effectiveness, we build a unified neural framework with word, sentence
and label denoising modules for relation extraction. Then we conduct
experiments to evaluate and compare these factors according to ten neu-
ral schemes. In addition, we discuss the typical cases of these factors and
find that influential word-level prior knowledge and partial confidence for
distantly supervised labels can significantly affect the denoising perfor-
mance. These implicational findings can provide researchers with more
insight of distantly supervised relation extraction.
Keywords: Relation extraction · Distant supervision · Label denoising
1 Introduction
Relation extraction is an active research task in information extraction [7] and
natural language processing, which aims to predict the relation between two
entities mentioned in plain text. It can benefit a variety of artificial intelligence
tasks such as question answering [1], information search [16] and knowledge base
construction [22]. However, supervised learning for relation extraction requires
a large amount of annotated data, which is costly to obtain.
Therefore, a distant supervision [12] strategy for relation extraction is pro-
posed to automatically label large-scale training data by heuristically aligning
entity pairs in plain text with the Knowledge Base (KB). However, distant
supervision tends to have the noisy labeling problem due to the incompletion
c Springer Nature Switzerland AG 2018
S. Tang et al. (Eds.): AAIM 2018, LNCS 11343, pp. 12–23, 2018.
https://doi.org/10.1007/978-3-030-04618-7_2
Factors Impacting the Label Denoising of Neural Relation Extraction 13
of the exploited KB. As shown in Fig. 1, for a relational triple born in(Donald
Trump,United States) in KB, all the sentences including two entities Donald
Trump and United States are labeled as the instances of relation born in, despite
the sentence “Donald Trump is the 45th president of the United States” fails to
express the relation. A similar noisy labeling problem exists in the entity pair
(Donald Trump,New York ), thus it is a great challenge for distantly supervised
relation extraction.
Fig. 1. A case of distantly supervised relation extraction.
To handle the noisy labeling issue, many researchers have made great efforts
to distantly supervised relation extraction. Recently, deep learning and neural
networks have achieved an excellent performance on this task. Zeng et al. [19]
selected the most likely sentence for each entity pair as a valid instance to train
the neural model. Then Lin et al. [10] selected multiple valid sentences to predict
the relation by assigning higher weight for more informative sentences. Subse-
quently, Liu et al. [11] used a soft label as the ground-truth to correct the noisy
labels, which directly achieved label denoising. Further, we also investigate some
common methods that may affect the label denoising in relation extraction such
as word-level prior knowledge [3,21] and class imbalance [6,9].
On the whole, existing works focus on specific algorithms but fail to explore
the factors impacting the label denoising of neural relation extraction. Hence, we
did a lot of research and experiments and found there are some commonalities
between different denoising methods. In summary, we propose three fundamental
and influential factors as follows:
(1) Labeling assumption. We mainly consider two labeling assumptions: one-
valid-instance assumption and multiple-valid-instance assumption. Specifi-
cally, one-valid-instance assumption [18,19] indicates that we only choose the
most effective sentence within all the sentences that mention an entity pair
to predict the relation. Multiple-valid-instance assumption [4,10] assumes
that one-valid-instance assumption may lose a lot of information from the
neglected sentences, thus we select multiple informative sentences to benefit
the relation prediction. Based on these assumptions, we can achieve sentence
denoising at different levels with respect to the distantly supervised labels.
14 T. Sun et al.
(2) Prior knowledge. We regard prior knowledge as all the information about
the target in training instances, which are objective facts and independent of
the knowledge base. Generally speaking, distant supervision mainly depends
on informative prior knowledge in plain text. For instance, as shown in Fig. 1,
in addition to entity mentions Donald Trump, United States and New York,
the contextual words born and president are more important to relation
extraction than the other words such as the and was. There are a variety
of methods to capture influential prior knowledge, including feature engi-
neering [2,13,14], neural networks [4,10,19], attention-based models [3,21].
Thus, how to effectively obtain informative prior knowledge is a challenging
factor for relation extraction.
(3) Confidence level. We define two confidence levels – complete confidence
and partial confidence in Distantly Supervised (DS) labels. With respect
to complete confidence, we treat DS labels as the gold labels of relation
extractor, ignoring the effect of false negatives and false positives. As for
partial confidence, we hope to learn a new soft label [11] as the ground-truth
by partly trusting the DS labels. At this time, we will exploit the semantic
information obtained from the correct relational facts. For example, as shown
in Fig. 1, the sentence “Donald Trump was born in New York ” will be labeled
as the instance of relation Place lived between two entities Donald Trump
and New York because the triple exists in KB, even it fails to describe the
relation. However, if we choose partial confidence in DS labels, we can correct
the noisy labels according to the similar semantic patterns (blue fonts) in
our corpora, and find the right relation Place of birth. Therefore, choosing
different confidence levels for DS labels is an important factor that impacts
the denoising work.
The main contributions of this paper are to:
(1) propose three main factors that impact the label denoising of distantly super-
vised relation extraction, including labeling assumption, prior knowledge and
confidence level.
(2) analyze these factors in a unified neural framework with word, sentence and
label denoising modules. Through the combination of different methods in
each module, we can test the denoising effectiveness of the proposed factors.
(3) conduct experiments to evaluate and compare these factors with ten neural
schemes, discuss the typical cases and discover that the important word-level
prior knowledge and partial confidence for distantly supervised labels can
remarkably improve the denoising effect.
The remainder of this paper is structured as follows. Section 2 describes our
proposed methodology in detail. Section 3 reports our experimental results and
analysis. Section 4 gives a conclusion of the whole paper.
2 Methodology
As shown in Fig. 2, in order to analyze the denoising impact of our proposed fac-
tors for relation extraction, we design a unified neural model and employ some
typical methods to model the factors. With respect to two labeling assump-
tions, we use the corresponding labeling schemes to reduce the influence of noisy
sentences for each entity pair. To evaluate the prior knowledge factor, we also
use two methods to verify the word-level denoising performance. As for differ-
ent confidence levels, we respectively apply DS label (complete confidence) and
soft label (partial confidence) as the ground-truth of relation extractor. Besides,
based on the partial confidence idea, we propose a dynamic soft-label (DySoft)
method to gradually learn the corrected labels during training. By combining
three modules – word, sentence and label denoising, we build a neural relation
extraction framework in an end-to-end training way.
Fig. 2. An overview of our denoising framework for the proposed factors.
2.1 Labeling Assumption Factor
To analyze the labeling assumption factor, we consider two typical assumptions -

one-valid-instance and multiple-valid-instance. Following these assumptions, we
will achieve sentence denoising at different levels for each entity pair. For this
reason, we utilize a sentence denoising module to model the two assumptions.
One-valid-instance assumption requires our model to only extract the most
likely sentence for each entity pair to predict the relation. We follow the work
of Zeng et al. [19] to realize this assumption, which chooses the sentence of the
highest relation probability as the valid instance.
Multiple-valid-instance assumption will select multiple informative sentences
for each entity pair as the valid candidates. For this assumption, different algo-
rithms [4,5,10] are introduced to achieve sentence-level denoising. Our model
uses the effective selective attention mechanism [10], which can learn high weight
for the valid sentences and low weight for the noisy sentences.
16 T. Sun et al.
2.2 Prior Knowledge Factor

The denoising impact of the prior knowledge factor depends on the design of
network architecture. CNN [8,20], PCNN [10,19] and LSTM [17] are common
architectures to learn sentence representation, which can obtain the discrimi-
native word sequence information of a sentence. Further, an attention mecha-
nism [3,21] provides us with the insight to achieve word-level denoising. There-
fore, we define a word denoising module to analyze the prior knowledge factor.
Concretely, the widely used PCNN in relation extraction is used to learn the
sentence representation. Besides, we introduce a word attention method to learn
the importance of each word in an instance. To obtain the relatedness between
each word in the sentence and entity pair, we compute the attention weight αi
of the i-th word in a sentence as follows:
exp(ui )
αi = (1)
k exp(uk )
ui = Ww tanh([wi , e1 , e2 ]) + bw (2)
where wi , e1 and e2 are the embeddings of the i-th word, entity pair e1 and e2
in the sentence. Ww and bw are training parameters. Here we follow the work
of Huang et al. [3] to utilize the attention mechanism.
2.3 Confidence Level Factor

With respect to different confidence levels for Distantly Supervised (DS) labels,
we employ a complete confidence method and two partial confidence methods to
test the impact of this factor. We model the factor as the label denoising module
so that we can correct DS labels directly according to the partial confidence
schemes. Firstly, for complete confidence, we use DS labels as the gold labels
of relation extraction task. While for partial confidence, we reproduce a soft-
label [11] method to learn a new label, which combines the predicted relation
score and the DS labels via a static confidence vector.
In order to resolve the problem of statically presetting the confidence vector,
(τ )
we propose a Dynamic Soft-label (DySoft) denoising method. Let ci denote
the i-th label confidence at the τ -th training step, then the soft label ŷ (τ ) at the
τ -th training step is iteratively calculated as follows:
(τ −1) (τ )
c + pi
ŷ (τ ) = arg max { i (τ −1) (τ )
} (3)
i ck + pk
k
where p(τ ) represents the relation score predicted by relation extraction model
at τ -th training step. The initial label confidence c(0) is fed with one-hot vector
y of the DS label y.
We simply learn the label confidence of the current training step by a linear
combination of previous label confidence and current relation score. The iteration
operation can model the correlation between label confidences at different train-
ing steps, and use preceding obtained soft labels to learn current new soft labels.
Hence, our proposed dynamic soft-label method remains soft-label consistency

learning and gradually achieves label denoising during training.
3 Experiments
In this section, we conduct experiments to evaluate and compare the proposed

factors. By combining different denoising modules, we implement 10 relation
extraction schemes in Table 1, including 4 reproduction and 6 original schemes.
Table 1. Our denoising schemes for the proposed factors.
Methods One One+WA Multiple Multiple+WA

Complete confidence ONE [19] ONE+WA MUL [4, 10, 17] MUL+WA
Partial confidence +Soft [11] – +Soft [11] –
+DySoft +WA+DySoft +DySoft +WA+DySoft
For each column in Table 1, one-valid-instance and multiple-valid-instance

assumptions respectively correspond to ONE and MUL methods. With respect
to the prior knowledge factor, we use PCNN as the baseline to extract informative
prior knowledge, and Word Attention (WA) module is a selective operation for
word denoising. For each row in Table 1, complete confidence uses DS labels as
the ground-truth, and partial confidence uses Soft-label method [11] and our
proposed Dynamic Soft-label method (DySoft) as the gold labels.
3.1 Dataset and Evaluation
Our proposed factors are compared and evaluated on a widely used distantly
supervised relation extraction benchmark dataset, which was developed by
Riedel et al. [13] via aligning the plain text of the New York Times corpus with
relational triples of Freebase. The dataset has 52 relation classes and a non-
relation NA class. The training set includes 522,611 sentences, 281,270 entity
pairs and 18,252 relational facts. The test set includes 172,448 sentences, 96,678
entity pairs and 1,950 relational facts. Similar to previous work [5,10,11,19],
we adopt the held-out evaluation with aggregate precision-recall curves and top
N Precision (P@N). The held-out evaluation compares the predicted relations
of entity pairs with the distantly supervised relations without requiring human
evaluation. As for parameter settings, we follow the settings of Liu et al. [11].
Specially, our proposed soft label is utilized after 3 training epochs.
3.2 Result and Analysis

As shown in Fig. 3, the neural models significantly performs better than tradi-
tional feature-based models Mintz [12], MultiR [2] and MIMLRE [15]. All the
18 T. Sun et al.
methods in Fig. 3 have compete confidence in distantly supervised labels. Obvi-

ously, adding a word attention (WA) module is beneficial to both one-valid-
instance (ONE) and multiple-valid-instance (MUL) assumptions.
Fig. 3. Performance comparison of neural methods with feature-based methods.
From Fig. 4, we present the impact of different denoising modules. For both
ONE and MUL assumptions, the Soft-label [11] method performs well when the
recall is less than 0.3, but it declines notably when the recall is greater than 0.4.
Our proposed DySoft-label method performs stable and achieves higher precision
in the MUL assumption. In particular, combining word denoising (WA) with
label denoising (Soft-label) module fails to improve the baseline. Consider that
correcting labels may have a bad impact on the attention of prior knowledge,
we get that the combination effect of WA+DySoft scheme may be not as good
as we expect.
Fig. 4. Performance comparison of different denoising schemes.
For a clear comparison, Table 2 presents the precisions of a variety of models

for different recalls R = 0.1/0.2/0.3/0.4/0.5 and Average Precision (AP) which
corresponds to the area under the precision-recall curve. Compared with the
neural baselines PCNN+ONE and PCNN+MUL which use DS labels as the
ground-truth, We can observe that adding WA or adding DySoft can achieve
the better performance than the other neural schemes. Next we will analyze the
effect of each factor in detail.
Table 2. Precisions of a variety of models for different recalls.
Method R 0.1 R 0.2 R 0.3 R 0.4 R 0.5 AP

Mintz [12] 0.399 0.286 0.168 - - 0.106
MultiR [2] 0.609 0.364 - - - 0.126
MIMLRE [15] 0.607 0.338 - - - 0.120
PCNN+ONE [19] 0.752 0.645 0.539 0.416 0.315 0.352
+WA 0.756 0.661 0.583 0.449 0.328 0.369
+Soft-label [11] 0.775 0.673 0.562 0.414 0.241 0.332
+DySoft-label 0.783 0.638 0.534 0.419 0.313 0.360
+WA+DySoft-label 0.768 0.646 0.522 0.430 0.318 0.358
PCNN+MUL [10] 0.699 0.641 0.548 0.439 0.338 0.365
+WA 0.783 0.672 0.554 0.444 0.346 0.382
+Soft-label [11] 0.777 0.682 0.553 0.390 0.225 0.331
+DySoft-label 0.762 0.668 0.584 0.481 0.358 0.370
+WA+DySoft-label 0.728 0.675 0.571 0.457 0.332 0.360
3.3 Effect of Labeling Assumption Factor

To test the effect of labeling assumption factor, we compare and evaluate two
typical assumptions – one-valid-instance assumption [19] and multiple-valid-
instance assumption [10]. Figure 4 shows that both assumptions bring an excel-
lent precision-recall performance, and multiple-valid-instance assumption per-
forms slightly better than one-valid-instance assumption due to the utilization of
more sentences. In particular, we find that the performance of both assumptions
is better than the report of Liu et al. [10] and the gap between two assumptions
is not as large as theirs. We owe the discriminative results to the CPU/GPU
hardware environment and code structure.
As multiple-valid-instance assumption has the effect on the entity pairs
which have multiple sentences, we compare and analyze the performance of two
assumptions on these entity pairs. Following the evaluation of Liu et al. [10],
we randomly select one, two and all sentences from the entity pairs which have
more than one sentence to conduct relation extraction. Table 3 presents top N
precision (P@N) of different settings for top 100/200/300 relational facts. We
discover that the whole performance of one-valid-instance and multiple-valid-
instance assumptions has 8% and 5% improvements compared with the reported
20 T. Sun et al.
Table 3. Top N precision (P@N) for relation extraction in the entity pairs with different
number of sentences.
Settings One Two All

P@N(%) 100 200 300 Mean 100 200 300 Mean 100 200 300 Mean
ONE 76.0 69.5 65.3 70.3 77.0 71.5 68.3 72.3 83.0 72.5 68.0 74.5
+WA 77.0 72.0 68.7 72.6 78.0 77.5 70.0 75.2 79.0 77.5 71.3 75.9
SATT 75.0 75.5 68.0 72.8 74.0 75.0 70.7 73.2 76.0 77.0 74.3 75.8
+WA 83.0 76.0 67.0 75.3 83.0 77.5 70.7 77.1 85.0 78.0 74.7 79.2
results of Liu et al. [10]. The average precision for different settings of multiple-
valid-instance assumption has 1% to 4% improvements compared with one-valid-
instance assumption, which is also less than the gap of 5% to 9% in previous
work [10]. Specially, the case study of labeling assumption for sentence-level
denoising has been reported [10], thus we omit it in our paper.
3.4 Effect of Prior Knowledge Factor
Since the denoising impact of the prior knowledge factor in neural networks is
usually invisible and directly reflected in model performance, we evaluate this
factor in the precision-recall curve and visualization of word attention. From
Fig. 4 and Table 2, we can find that adding a word-level denoising WA mechanism
can greatly benefit both ONE and MUL assumptions. Particularly, MUL+WA
model achieves the best AP performance compared with all the other models.
As shown in Table 2, combining WA mechanism makes the average precision
improve 3% and 4% for ONE and MUL baselines. These results demonstrate the
effectiveness of adding the word-level importance for the prior knowledge factor.
Table 4. Visualization of word attention.
Relation Instance
place of birth Roscoe Lee Browne was born on May 2, 1925, in Woodbury .
place of death His most recent book is “Murder in Amsterdam: the death of
Theo Van Gogh and the limits of tolerance.”
company “I like Eli Manning,” the NBC football analyst John
Madden said in a conference call Thursday.
founder Still , some early investors like Peter Munk , the founder and
chairman of Barrick Gold, were skeptical at first.
Table 4 presents the visualization of word attention mechanism, the bold font
and size of a word denote its importance, the italic words are entity mentions.
For example, with respect to the instance of relation place of birth, the word born
Another random document with
no related content on Scribd:
“Fear not a constraining measure!
—Yielding to the gentle spell,
Lucida! from domes of pleasure,
Or from cottage-sprinkled dell,
Comes to regions solitary,
Where the eagle builds her aery,
Above the hermit’s long-forsaken cell!
—She comes!—behold
That figure, like a ship with silver sail!
Nearer she draws; a breeze uplifts her veil;
Upon her coming wait
As pure a sunshine, and as soft a gale,
As e’er, on herbage-covering earthly mold,
Tempted the bird of Juno to unfold
His richest splendour,—when his veering gait,
And every motion of his starry train,
Seem governed by a strain
Of music, audible to him alone.”
And then we have a picture of the lady:—

“worthy of earth’s proudest throne!
Nor less, by excellence of nature, fit
Beside an unambitious hearth to sit
Domestic queen, where grandeur is unknown;
What living man could fear
The worst of fortune’s malice, wer’t thou near,
Humbling that lily stem, thy sceptre meek,
That its fair flowers may brush from off his cheek
The too, too, happy tear?
—Queen, and handmaid lowly!
Whose skill can speed the day with lively cares,
And banish melancholy
By all that mind invents, or hand prepares;
O thou, against whose lip, without its smile
And in its silence even, no heart is proof;
Whose goodness, sinking deep, would reconcile
The softest nursling of a gorgeous palace,
To the bare life beneath the hawthorn roof
Of Sherwood’s archer, or in caves of Wallace—
Who that hath seen thy beauty could content
His soul with but a glimpse of heavenly day?
Who that hath loved thee, but would lay
His strong hand on the wind, if it were bent
To take thee in thy majesty away?
Pass onward (even the glancing deer
Till we depart intrude not here;)
That mossy slope, o’er which the woodbine throws
A canopy, is smooth’d for thy repose!”
The next lady that he invokes before the astonished youth is his own
daughter—sweet Dora—the previous one was Miss Southey.
“Come, if the notes thine ear may pierce,
Come, youngest of the lovely three,
Submissive to the mighty verse
And the dear voice of harmony,
By none more deeply felt than thee!
I sang; and lo! from pastures virginal
She hastens to the haunts
Of Nature, and the lonely elements.
Air sparkles round her with a dazzling sheen;
And mark, her glowing cheek, her vesture green!
And, as if wishful to disarm
Or to repay the potent charm,
She bears the stringed lute of old romance,
That cheered the trellissed arbour’s privacy,
And soothed war-wearied knights in raftered hall.
How vivid, yet how delicate, her glee!
So tripped the muse, inventress of the dance;
So, truant in waste woods, the blithe Euphrosyne!”
But the ringlets of that head,

Why are they ungarlanded?
Why bedeck her temples less
Than the simplest shepherdess?
Is it not a brow inviting
Choicest flowers that ever breathed,
Which the myrtle would delight in,
With Idalian rose enwreathed?
But her humility is well content
With one wild floweret (call it not forlorn),—
Flower of the Winds—beneath her bosom worn—
Yet more for love than ornament.”
Then follows that beautiful description of her moral graces, already quoted
in these pages, beginning—
“Open ye thickets! let her fly,

Swift as a Thracian nymph, o’er field and height;”
the whole picture being as fine a conception, and as rich an embodyment, of

this sweet Dora,—judging from her portrait in the second volume of the
“Memoirs,” and from numerous written and spoken reports of her person
and character,—as the highest genius and the highest art combined, could
possibly have produced. And now for Miss Coleridge:—
“Last of the three, tho’ eldest born,

Reveal thyself, like pensive morn
Touched by the skylark’s earliest note,
E’er humble-gladness be afloat.
But whether in the semblance drest
Of dawn, or eve, fair vision of the west,
Come, with each anxious hope subdued
By woman’s gentle fortitude,
Each grief, thro’ meekness, settling into rest.
—Or I would hail thee when some high-wrought page
Of a closed volume, lingering in thine hand,
Has raised thy spirit to a peaceful stand
Among the glories of a happy age.”
And, behold! she is here:—

“Her brow hath opened on me—see it there,
Brightening the umbrage of her hair;
So gleams the crescent moon, that loves
To be descried thro’ shady groves.
Tenderest bloom is on her cheek;
Wish not for a richer streak;
Nor dread the depth of meditative eye;
But let thy love, upon that azure field
Of thoughtfulness and beauty, yield
Its homage offered up in purity.
What would’st thou more? In sunny glade,
Or under leaves of thickest shade,
Was such a stillness e’er diffused
Since earth grew calm while angels mused?
Softly she treads, as if her foot were loth
To crush the mountain dew-drop—soon to melt,
On the flower’s breast; as if she felt
That flowers themselves, whate’er their hue,
With all their fragrance, all their glistening,
Call to the heart for inward listening—
And tho’ for bridal wreaths and tokens true
Welcomed wisely; tho’ a growth
Which the careless shepherd sleeps on,
As fitly spring from turf the mourner weeps on—
And without wrong are cropped the marble tomb to strew.
And now the charm is over;
——“the mute phantom’s gone,

Nor will return—but droop not, favoured youth,
The apparition that before thee shone
Obeyed a summons covetous of truth.
From these wild rocks thy footsteps I will guide
To bowers in which thy fortunes may be tried,
And one of the bright three become thy happy bride.”
A fairer subject than this, for the imagination of the true painter, does
scarcely exist in poetry. The gorgeous magnificence of Miss Southey—the
wild, bird-like nature of Dora, the mystic, spiritual, meditative beauty of
Miss Coleridge. Here is material enough for the highest effort of art.
A number of poems followed this exquisite “Triad”—viz., “The Wishing
Gate,” in 1828—“The Lawn,” “Presentiments,” “The Primrose on the
Rock,” “Devotional Incitements;” these last were written between 1828 and
’32. A number of gold and silver fishes presented to the poet by Miss H. J.
Jewsbury, who subsequently died of the cholera in India,—and afterwards
removed to the pond already alluded to, under the oak in “Dora’s field,”
suggested the verses “Gold and Silver Fishes in a Vase,” and likewise
“Liberty,” and “Humanity;” “The Poet and the caged Turtle Dove” was
likewise suggested by real circumstances. Miss Jewsbury had given Dora a
pair of these beautiful birds; one of them was killed by an un-necessary cat,
and not a “harmless one;” the other survived many years, and had a habit of
cooing the moment Wordsworth began “booing” his poems, as the country
people called it.
Wordsworth gives an amusing account of a visit which he paid about this
time to “Chatsworth.” He had undertaken to ride his daughter’s pony from
Westmoreland to Cambridge, that she might have the use of it during a visit
she was about to make to her uncle at Trinity; and on his way from
Bakewell to Matlock, he turned off to see the splendid mansion of the great
Duke of Devonshire. By-and-bye a tremendous storm came on, and the poet
was drenched through to the very skin, whilst the pony, to make his rider’s
seat the more easy, went “slantwise” all the way to Derby. Notwithstanding
this, however, and the pelting of the pitiless storm, Wordsworth managed to
hold sweet and sad converse with his muse, and composed his “Lines to the
Memory of Sir George Beaumont,” who died 7th February, 1827. It is a
picture which we cannot readily forget, and shows how completely the poet
was master of himself. Sir George Beaumont and his lady were friends and
benefactors of Wordsworth—he loved them both intensely. Walking
through the grounds and gardens of Coleorton with Sir George—the
successor in the Baronetcy to his friend—and after the death of Lady
Beaumont, which took place in 1822, he comes suddenly to her ladyship’s
grotto, near the fountain, and is overwhelmed with his feelings, and the
recollection of the dead, and the happy memories which rush over his mind
in connection with this place, so that he cannot speak for tears. On his
return home he wrote the elegiac musings, already mentioned in these
memoirs, which are full of love, and the sanctity of a sweet sorrow. In the
same year (1831) were composed “The Armenian Lady’s Love,” “The
Egyptian Maid,” and “The Russian Fugitive,” poems in which all the
beauties of language are pressed, along with the simplicity which marks the
old English ballads. Lines on his portrait, painted by Pickersgill, and
preserved with sacred veneration in St. John’s College, Cambridge, were
likewise written in this year, as well as the inscription already quoted, for
the stone at Rydal.
Besides these poetical compositions, however, Wordsworth interested
himself in public affairs; and having fixed principles of political and social
economy in his own mind, regarded all public measures at variance with
them, as fatal errors, and subversive in their consequences of the highest
human concerns. In 1806, he wrote a letter to a friend, who had consulted
him respecting the education of his daughter—in which he gives some
sound and excellent advice respecting the training and development of
youthful minds. For Wordsworth had at an early period devoted his
attention to the subject of education, and had his own views respecting it—
views which were marked by the spiritual peculiarity of his mind. When he
wrote “The Excursion,” he seems to have had the highest hopes for man,
when education should become universal; and insisted that the State should
teach those to obey, from whom she exacted allegiance:—
“O for the coming of that glorious time,

When, prizing knowledge, as her noblest work
And best protection, this imperial realm
While she exacts allegiance, shall admit
An obligation on her part to teach
Them who are born to serve her and obey;
Binding herself, by statute, to secure
For all the children whom her soil maintains
The rudiments of letters, and inform
The mind with moral and religious truth,
Both understood, and practised.”
Excursion, Book ix.
He was an avowed enemy, however, at a later period—for his views

respecting the modus operandi of teaching, had undergone some change
since “The Excursion” was written—to all Infant Schools, Madras Systems,
and Bell Systems. The former he regarded as usurping the functions of
motherly duty; the latter, as dead mechanism. Speaking of the education of
girls, he says:—“I will back Shenstone’s ‘School Mistress,’ by her winter
fire, and in her summer garden seat, against all Dr. Bell’s sour-looking
teachers in petticoats. What is the use of pushing on the education of girls
so fast, and moving by the stimulus of Emulation, who, to say nothing
worse of her, is cousin-german to Envy? What are you to do with these
girls? What demand is there for the ability that they may have prematurely
acquired? Will they not be indisposed to bend to any kind of hard labour or
drudgery? And yet many of them must submit to it, or go wrong. The
mechanism of the Bell System is not required in small places; praying after
the fugleman, is not like praying at a mother’s knee. The Bellites overlook
the difference: they talk about moral discipline; but wherein does it
encourage the imaginative feelings? in short, what she practically
understands is of little amount, and too apt to become the slave of the bad
passions. I dislike display in everything; above all, in education.... The old
dame (Shenstone’s) did not affect to make theologians and logicians; but
she taught to read; and she practised the memory, often no doubt by rote,
but still the faculty was improved; something, perhaps, she explained, but
trusted the rest to parents and masters, and to the pastor of the parish. I am
sure as good daughters, as good servants, as good mothers and wives, were
brought up at that time as now, when the world is so much less humble-
minded. A hand full of employment, and a head not above it, with such
principles and habits as may be acquired without the Madras machine, are
the best security for the chastity of wives of the lower rank.”
The above extract is from a letter dated 1828, and addressed to the Rev.
Hugh Jones Rose, formerly principal of King’s College, London. It
exemplifies, in a very striking manner, the change which had come over
Wordsworth’s mind upon the subject of education, and does not strike me as
being particularly creditable to him.
On the 13th of April, 1836, Wordsworth took part in the ceremony of
laying the foundation-stone of certain new schools, about to be erected at
Bowness, Windermere, and made a speech upon the occasion; in which he
advocates a very humble kind of instruction for the working classes;
forgetting that man is to be educated because he is a man, and not neglected
because he happens to be one of the “lower orders.” I have no sympathy
with this foolish cant about educating people according to their station, and
am sorry that Wordsworth’s sanction can be quoted in its favour. I must
reserve what I have to say upon this subject, however, for my analysis of
the mind and writings of the poet.
In 1835, Wordsworth published his “Yarrow Revisited, and other
Poems.” Speaking of “Yarrow Re-visited,” he says: “In the autumn of 1831,
my daughter and I set off from Rydal, to visit Sir Walter Scott, before his
departure for Italy. This journey had been delayed by an inflammation in
my eyes, till we found that the time appointed for his leaving home would
be too near for him to receive us without considerable inconvenience.
Nevertheless we proceeded, and reached Abbotsford on Monday. I was then
scarcely able to lift up my eyes to the light. How sadly changed did I find
him from the man I had seen so healthy, gay, and hopeful, a few years
before, when he said, at the inn at Patterdale, in my presence, his daughter,
Ann, also being there, with Mr. Lockhart, my own wife and daughter, and
Mr. Quillinan: ‘I mean to live till I am eighty, and shall write as long as I
live.’ Though we had none of us the least thought of the cloud of misfortune
which was then going to break upon his head, I was startled, and almost
shocked, at that bold saying, which could scarcely be uttered by such a
man, without a momentary forgetfulness of the instability of human life.
But to return to Abbotsford. The inmates and guests we found there, were
Sir Walter, Major Scott, Anne Scott, and Mr. and Mrs. Lockhart; Mr.
Laidlaw, a very old friend of Sir Walter’s; one of Burns’s sons, an officer of
the Indian service, had left the house the day before, and had kindly
expressed his regret that he could not wait my arrival, a regret that I may
truly say was mutual. In the evening, Mr. and Mrs. Liddell sang, and Mrs.
Lockhart chaunted old ballads to her harp; and Mr. Allan, hanging over the
back of a chair, told, and acted, odd stories in a humourous way. With this
exhibition, and his daughters’ singing, Sir Walter was much amused, and,
indeed, so were we all, as far as circumstances would allow.”
On the following morning (Tuesday) Sir Walter accompanied
Wordsworth, and most of his friends, to Newark Castle, on the Yarrow, and
it was upon this occasion that the lines, “Yarrow Revisited,” were written.
On the morning of Thursday following, when the poet left Abbotsford, he
had a serious conversation with Sir Walter, who spoke with gratitude of the
happy life he had led. Sir Walter wrote also a few lines in Dora’s album,
addressed to her; and when he presented her with the book, in his study, he
said: “I should not have done a thing of this kind, but for your father’s sake;
they are probably the last verses I shall ever write.” “They shew,” says,
Wordsworth, “how much his mind was impaired; not by the strain of
thought, but by the execution—some of the lines being imperfect, and one
stanza wanting corresponding rhymes.” Poor Sir Walter!—what a spectacle
it was to see that colossal intellect tumbling into ruins.
Several poems were the result of this short tour, beside the “Yarrow
Revisited,”—such as “The Place of Burial,” “On the Sight of a Manse in the
South of Scotland,” &c., &c.—Wordsworth’s health, too, was much
improved by this tour, and a violent inflammation of the eyes—a complaint
to which he was much subject,—left him whilst walking through the
Highlands, by the side of his “open carriage,” driven by Dora!
Amongst the poems contained in the volume entitled, “Yarrow
Revisited,” were many of a political character, for they were written
between the years 1830 and 1834, when the Revolution of France, and the
Reform party in England, were agitating society to its centre. Wordsworth
now hated revolution, and reform also; was opposed to a large and
enlightened system of education; and to the admission of Dissenters to the
Universities. His plea was the old constitution of things, which could not,
he thought, be mended without being broken up and destroyed. “Since the
introduction of the Reform Bill, I have been persuaded,” he says, “that the
Constitution of England cannot be preserved. It is a question, however, of
time.” The poem entitled, “The Warning,” will give the best idea of
Wordsworth’s strong political opinions and feelings at this time. As a
contrast, however, to these narrow yet patriotic views, we turn to the
“Evening Voluntaries,” a collection of sweet poems, which were published
in the same volume as the “Yarrow.” They were written on a high part of
the coast of Cumberland, on April 7th (Easter Sunday), the author’s 63rd
birth-day, between Moresby and Whitehaven, whilst he was on a visit to his
son, who was then rector of Moresby.—Very beautiful, indeed, are these
poems, which read like twilight vespers in some old abbey’s chancel.
Wordsworth says of them—“With the exception of the eighth and ninth, this
succession of voluntaries originated in the concluding lines of the last
paragraph of this poem, [i.e. of the poem, written on the author’s birth-day,
and marked No. 2, in the “Voluntaries,” commencing, “The sun that seemed
so mildly to retire.”] With this coast I have been familiar from my earliest
childhood, and remember being struck, for the first time, by the town and
port of Whitehaven, and the white waves breaking against its quays and
piers, as the whole came into my view from the top of the high ground,
down which the road, that has since been altered, descended abruptly. My
sister, when she first heard the voice of the sea from this point, and beheld
the scene spread before her, burst into tears. Our family then lived at
Cockermouth, and this fact was often mentioned amongst us, indicating the
sensibility for which she was remarkable.”
As a specimen of the “Evening Voluntaries,” take the following:—
“Calm is the air, and loth to lose

Day’s grateful warmth, tho’ moist with falling dews.
Look for the stars, you’ll say that there are none;
Look up a second time, and one by one,
You mark them twinkling out with silvery light,
And wonder how they could elude the sight!
The birds, of late so noisy in their bowers,
Warbled awhile with faint and fainter powers,
But now are silent as the dim-seen flowers.
Nor does the village church clock’s iron tone
The time’s and season’s influence disown;
Nine beats distinctly to each other bound
In drowsy sequence—how unlike the sound
That in rough winter, oft inflicts a fear
On fireside listeners, doubting what they hear!
The shepherd, bent on rising with the sun,
Had closed his door before the day was done,
And now with thankful heart to bed doth creep,
And joins the little children in their sleep.
The bat, lured forth where trees the lane o’ershade,
Flits and reflits along the dark arcade;
The busy dor-hawk chases the white moth
With burring note, which industry and sloth
Might both be pleased with, for it suits them both.
A stream is heard—I see it not but know
By its soft music where the waters flow:
Wheels and the tread of hoofs are heard no more;
One boat there was, but it will touch the shore
With the next dipping of its slackened oar;
Faint sound that for the gayest of the gay,
Might give to serious thought a moment’s sway,
As a last token of man’s toilsome day.”
Wordsworth does not seem, during any period of his life, to have been on
intimate terms with any of his contemporaries. He preferred the flower of
the literateurs, Coleridge, Scott, Southey; and these, with the exception
perhaps of Rogers, were his chief friends. We have letters of his, however,
to much smaller fry; to Mrs. Hemans, and Miss Jewsbury, for example—
and to sundry editors of other men’s wares; but there is little or no
recognition of Byron, Shelly, Keats, Tennyson, Baily, Campbell, Moore, nor
yet of Dickens or Bulwer. His letters represent his character even better than
his poetry; they are Wordsworth in undress, without the “garland and
singing robe,” and are worthy to be studied. I like much what he says to the
Rev. Robert Montgomery, author of “The Devil and Father Luther,” and
pious Robert would do well even at this late day to think on it. Montgomery
had sent Wordsworth a copy of his poems, and in reply, the poet answers: “I
cannot conclude without one word of literary advice which I hope you will
deem my advanced age entitles me to give. Do not, my dear sir, be anxious
about any individual’s opinion concerning your writings, however highly
you may think of his genius, or rate his judgment. Be a severe critic to
yourself; and depend upon it no person’s decision upon the merit of your
works will bear comparison in point of value with your own. You must be
conscious from what feeling they have flowed, and how far they may or
may not be allowed to claim on that account, permanent respect; and above
all I would remind you, with a view to tranquillise and steady your mind,
that no man takes the trouble of surveying and pondering another’s writings
with a hundredth part of the care which an author of sense and genius will
have bestowed upon his own. Add to this reflection another, which I press
upon you, as it has supported me through life—viz.: That posterity will
settle all accounts justly, and that works which deserve to last will last; and
if undeserving this fate, the sooner they perish the better.”
In the year 1836 the sister of the poet’s wife—Miss Sarah Hutchinson,
who had resided with the family at Rydal, died, and was buried in Grasmere
church, “near the graves of two young children removed from a family to
which through life she was devoted.”
In the following year, 1837, Wordsworth, accompanied by his friend H.
C. Robinson, Esq., set off from London for Rome, returning in August. The
“Itinerary” of the travellers is contained in the “Memoirs,” along with some
memoranda by the poet; but they are not of much interest. Many fine
pieces, however, sprung as usual from the journey, as well as a goodly
number of sonnets. They originally appeared in a volume entitled “Poems,
chiefly of Early and Late Years,” in 1842. In 1839, Wordsworth received
the degree of D.C.L., from the University of Oxford, which was conferred
on him in the Sheldonian Theatre, amidst shouts of rejoicing such as had
never before been heard in that city, except upon the occasion of an
unexpected visit of the Duke of Wellington. In 1838, Wordsworth prepared
a new edition of his poems, to be published by Moxon, and continued to
live at Rydal, in his quiet and musical manner, writing poems, taking
rambles, and conducting his correspondence until 1843, when he was
appointed Poet Laureate of England, Southey having died on the 21st of
March of that year, and the appointment having been offered to Wordsworth
on the 31st of the same month. One occurrence only broke the even tenor of
the poet’s life in the interim alluded to, and this was an accident by which
he was upset from his gig, and thrown violently into a plantation. The
accident was owing to the carelessness and want of skill in the driver of a
coach, which they met on the road. No serious consequences followed,
however, and inquiries and congratulations flowed in on all sides, from the
peasant up to Queen Adelaide.
From the time of Wordsworth’s appointment as Laureate,—which it
ought to be said he at first refused, and only accepted with the
understanding that it should be an honorary office,—he wrote very little
poetry. His work, indeed, was done, his mission accomplished; and his old
days were spent in rambling over the hills, and in the quiet enjoyment of his
family, friends, fame, and fortune. Honours of a high order were
subsequently heaped upon him. In the year 1838, the University of Durham
took the initiative in conferring an academic degree on the poet; then the
grand old Mother, Oxford, followed,—and in 1846 he was put in
nomination, without his knowledge, for the office of Lord Rector of the
University of Oxford, and gained a majority of twenty-one votes, in
opposition to the premier, Lord John Russell. “The forms of election,
however,” says Wordsworth, in a letter to Sir W. Gomm, of Port Louis,
Mauritius, dated November 23, 1846, “allowed Lord John Russell to be
returned through the single vote of the sub-rector voting for his superior. To
say the truth, I am glad of this result, being too advanced in life to
undertake with comfort any considerable public duty, and it might have
seemed ungracious to have declined the office.”
On the 20th of January, 1847, Mr. William Wordsworth, the younger son
of the poet, was married at Brighton, to Fanny Eliza Graham, youngest
daughter of Reginald Graham, Esq., of Brighton, who was a native of
Cumberland; and whilst the joy of this event was still fresh in the hearts of
the Rydal household, a dread calamity awaited them in the death of Mrs.
Quillinan—the sweet Dora so often spoken of in these pages, the beloved
daughter of the poet. As previously stated, she had accompanied her
husband to Portugal for the benefit of her health,—and although the change
seemed at first to have operated favourably upon her, it was soon evident,
on her return home, that she was doomed for the silent bourne of all
travellers in this world. She died on the 9th day of July, 1847, and was
buried in Grasmere church-yard. Her death was a terrible blow to the
venerable poet, now in his eightieth year,—but he bore up patiently, with
the heart and hope of a Christian.
Three years after this sad loss, Wordsworth himself was summoned
away. On Sunday, the 10th of March, 1850, he attended at Rydal chapel for
the last time, visiting, during the day, a poor old woman, who had once
been his servant, and another person who was sick, and as the poet said,
“never complained.”
“On the afternoon of the following day, he went towards Grasmere, to
meet his two nieces, who were coming from Town End. He called at the
cottage near the White Moss Quarry, and the occupant being within, he sat
down on the stone seat of the porch, to watch the setting sun. It was a cold,
bright day. His friend and neighbour, Mr. Roughsedge, came to drink tea at
Rydal, but Mr. Wordsworth not being well, went early to bed.”
From this time he gradually grew worse; and in order to convey to him
the impressions of his physicians, Mrs. Wordsworth whispered in a soft
voice, full of deep devotion, “Dear William, you are going to Dora.” How
delicate, how affectionate, how poetical! But the poet did not hear, or did
not seem to hear; and yet, twenty-four hours after, when one of his nieces
came into the room, and gently drew aside the curtains of his bed, he caught
a glimpse of her figure, and asked, “Is that Dora?”
On the 23rd of April—the birth-day, and death-day of Shakspeare, the
great-hearted Wordsworth went back again to God.
He was buried on the 27th, in Grasmere church-yard.
Those who would know more of the poet must go to his writings; and, I
may add, that the “Memoirs” of Dr. Wordsworth are indispensable to a full
understanding both of the Poet and the Man. His letters, containing his most
private thoughts, are printed there with plentiful profuseness; and the
“Memoranda” respecting the origin of his poems are intensely interesting
and important to all students of Wordsworth. The reminiscences of various
persons who knew him, set the character of the poet before us in strong
relief. All agree in speaking of him as a most kindly, affectionate, and
hospitable man, living with the simple tastes and manners of a patriarch, in
his beautiful home. My limits prevent me from entering into an analysis of
his mind and character, as I had intended to do; I must reserve this work,
therefore, for another occasion, and will conclude with a few quotations
from the poet’s “Table-Talk,” respecting his cotemporaries.—Speaking of
Goethe, he says:—
“He does not seem to me to be a great poet in either of the classes of
poets. At the head of the first I would place Homer and Shakspeare, whose
universal minds are able to reach every variety of thought and feeling,
without bringing his own individuality before the reader. They infuse, they
breathe life into every object they approach, but you cannot find themselves.
At the head of the second class, those whom you can trace individually in
all they write, I would place Spenser and Milton. In all that Spenser writes,
you can trace the gentle, affectionate spirit of the man; in all that Milton
writes, you find the exalted, sustained being that he was. Now, in what
Goethe writes, who aims to be of the first class, the universal, you find the
man himself, the artificial man, where he should not be found; so that I
consider him a very artificial writer, aiming to be universal, and yet
constantly exposing his individuality, which his character was not of a kind
to dignify. He had not sufficiently clear moral perceptions to make him
anything but an artificial writer.
And again:—
“I have tried to read Goethe. I never could succeed. Mr.—— refers me to
his ‘Iphigenia,’ but I there recognise none of the dignified simplicity, none
of the health and vigour which the heroes and heroines of antiquity possess
in the writings of Homer. The lines of Lucretius describing the immolation
of Iphigenia are worth the whole of Goethe’s long poem. Again there is a
profligacy, an inhuman sensuality, in his works, which is utterly revolting. I
am not intimately acquainted with them generally. But I take up my ground
on the first canto of ‘Wilhelm Meister;’ and as the attorney-general of
human nature, I there indict him for wantonly outraging the sympathies of
humanity. Theologians tell us of the degraded nature of man; and they tell
us what is true. Yet man is essentially a moral agent, and there is that
immortal and unextinguishable yearning for something pure and spiritual
which will plead against these poetical sensualists as long as man remains
what he is.”
Of Scott he says:—
“As a poet, Scott cannot live, for he has never in verse written anything
addressed to the immortal part of man. In making amusing stories in verse,
he will be superseded by some newer versifier; what he writes in the way of
natural description is merely rhyming nonsense. As a prose writer, Mr.
Wordsworth admitted that Scott had touched a higher vein, because there he
had really dealt with feeling and passion. As historical novels, professing to
give the manners of a past time, he did not attach much value to those
works of Scott’s, so called, because that he held to be an attempt in which
success was impossibility. This led to some remarks on historical writing,
from which it appeared that Mr. Wordsworth has small value for anything
but contemporary history. He laments that Dr. Arnold should have spent so
much of his time and powers in gathering up, and putting into imaginary
shape, the scattered fragments of the history of Rome.”
And again:—
“He discoursed at great length on Scott’s works. His poetry he
considered of that kind which will always be in demand, and that the supply
will always meet it, suited to the age. He does not consider that it in any
way goes below the surface of things; it does not reach to any intellectual or
spiritual emotion; it is altogether superficial, and he felt it himself to be so.
His descriptions are not true to nature; they are addressed to the ear, not to
the mind. He was a master of bodily movements in his battle-scenes; but
very little productive power was exerted in popular creations.”
Moore:—
“T. Moore has great natural genius; but he is too lavish of brilliant
ornament. His poems smell of the perfumer’s and milliner’s shops. He is
not content with a ring and a bracelet, but he must have rings in the ear,
rings on the nose—rings everywhere.”
Shelley:—
“Shelley is one of the best artists of us all: I mean in workmanship of
style.”
Tennyson:—
“I saw Tennyson, when I was in London, several times. He is decidedly
the first of our living poets, and I hope will live to give the world still better
things. You will be pleased to hear that he expressed in the strongest terms
his gratitude to my writings. To this I was far from indifferent, though
persuaded that he is not much in sympathy with what I should myself most
value in my attempts—viz., the spirituality with which I have endeavoured
to invest the material universe, and the moral relations under which I have
wished to exhibit its most ordinary appearances.”
Hartley Coleridge—
He spoke of with affection. “There is a single line,” he added, “in one of
his father’s poems, which I consider explains the after life of the son. He is
speaking of his own confinement in London, and then says,—
‘But thou, my child, shalt wander like a breeze.’
“He thought highly also of some of Hartley’s sonnets.

Southey—
He said had outlived his faculties. His mind he thought had been
wrecked by long watching by the sick bed of his wife, who had lingered for
years in a very distressing state.
Coleridge—
He said the liveliest and truest image he could give of Coleridge’s talk
was that of “a mystic river, the sound or sight of whose course you caught
at intervals, which was sometimes concealed by forests, sometimes lost in
sand, and then came flashing out broad and distinct; then again took a turn
which your eye could not follow, yet you knew and felt that it was the same
river....[M] Coleridge had been spoilt as a poet by going to Germany. The
bent of his mind, at all times very much inclined to metaphysical theology,
had there been fixed in that direction.”
Lord Byron—
“Has spoken severely of my compositions. However faulty they may be,
I do not think I ever could have prevailed with myself to print such lines as
he has done, for instance—
‘I stood at Venice on the Bridge of Sighs,

A palace and a prison on each hand.’
“Some person ought to write a critical review analising Lord Byron’s
language, in order to guard others against imitating him in these respects.”
Emerson and Carlyle—
“Do you know Miss Peabody of Boston? She has just sent me, with the
highest eulogy, certain essays of Mr. Emerson. Our—— and he appear to be
what the French called esprits forts, though the French idols showed their
spirit after a somewhat different fashion. Our two present Philosophes, who
have taken a language which they suppose to be English, for their vehicle,
are, verily, par nobile fratrum, and it is a pity that the weakness of our age
has not left them exclusively to this appropriate reward—mutual
admiration. Where is the thing which now passes for philosophy at Boston
to stop?”
Such are a few random selections from the spoken opinions of the poet.
He hated innovation, hence his attack upon the two last named authors, not
made, I think, in the very best spirit. I must here leave him, however. He
will stand well upon his honours in all future generations, and must
certainly be ranked as a poet in the same category with Milton.
FINIS.
——————
J. S. Pratt, Stokesley, Yorkshire.
FOOTNOTES:
[A] These remarks do not of course apply to Cowper and Burns, to whom our
modern literature is so deeply indebted, but to their predecessors, from Pope
downwards.
[B] Memoirs of William Wordsworth, by Dr. Wordsworth, vol. 1, page 7.
[C] De Quincy, Tait’s Magazine, for 1839.
[D] It is related by De Quincy, that during Wordsworth’s early residence in the
lake country—after his return from Cambridge—his mind was so oppressed by the
gloomy aspect of his fortunes, that evening card-playing was resorted to, to divert
him from actual despondency.
[E] Dr. Wordsworth’s Memoir, page 53.
[F] Memoir, page 71-2.
[G] Chambers’ Papers for the People, article Wordsworth.
[H] Vol. 1, page 149 to 154.
[I] Memoirs, Vol. 1., page 156
[J] Dr. Wordsworth’s “Memoirs,” vol. 2, p. 94.
[K] “Memoirs,” vol. 2, p. 102-105.
[L] Memoirs, vol. ii., p. 121.
[M] This view of Coleridge is confirmed by Carlyle, in his “Life of John Sterling,”
just published.
“Coleridge sat on the brow of Highgate-hill, in those years, looking down on
London and its smoke tumult, like a sage escaped from the inanity of life’s battle;
attracting towards him the thoughts of innumerable brave souls still engaged there.
His express contributions to poetry, philosophy, or any specific province of human
literature or enlightenment, had been small and sadly intermittent; but he had,
especially among young inquiring men, a higher than literary,—a kind of prophetic,
or magician character. He was thought to hold,—he alone in England,—the key of
German and other transcendentalisms; knew the sublime secret of believing by ‘the
reason’ what ‘the understanding’ had been obliged to fling out as incredible; and
could still, after Hume and Voltaire had done their best and worst with him, profess
himself an orthodox Christian, and say and point to the Church of England, with its
singular old rubrics and surplices at Allhallowtide, Esto perpetua. * * * * He
distinguished himself to all that ever heard him as at least the most surprising talker
extant in this world,—and to some small minority, (by no means to all,) the most
excellent. The good man, he was now getting old, towards sixty perhaps,—and gave
you the idea of a life that had been full of sufferings; a life heavy-laden, half-
vanquished, still swimming painfully in seas of manifold physical and other
bewilderment. * * * * * I still recollect his ‘object’ and ‘subject,’ terms of continual
recurrence in the Kantean province; and how he sung and snuffled them into ‘om-m-
mject’ ‘sum-m-mject,’ with a kind of solemn shake or quaver, as he rolled along. No
talk, in his century or in any other, could be more surprising.
* * * * *
“He had knowledge about many things and topics,—much curious reading; but
generally all topics led him, after a pass or two, into the high seas of theosophic
philosophy, the hazy infinitude of Kantean transcendentalism, with its ‘sum-m-
mjects’ and ‘om m-mjects.’ Sad enough, for with such indolent impatience of the
claims and ignorances of others, he had not the least talent for explaining this or
anything unknown to them; and you swam and fluttered in the mistiest, wide,
unintelligible deluge of things,—for most part in a rather profitless, uncomfortable
manner. Glorious islets, too, I have seen rise out of the haze; but they were few, and
soon swallowed in the general element again. * * * * * * One right peal of concrete
laughter at some convicted flesh-and-blood absurdity, one burst of noble indignation
at some injustice or depravity rubbing elbows with us on this solid earth,—how
strange would it have been in that Kantean haze-world, and how infinitely cheering
amid its vacant air-castles, and dim-melting ghosts and shadows! None such ever
came. His life had been an abstract thinking and dreaming, idealistic one, passed
amid the ghosts of defunct bodies and of unborn ones. The mourning sing-song of
that theosophico-metaphysical monotony left on you, at last, a very dreary feeling.”

Textbook Algorithmic Aspects in Information and Management 12Th International Conference Aaim 2018 Dallas TX Usa December 3 4 2018 Proceedings Shaojie Tang Ebook All Chapter PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Textbook Algorithmic Aspects in Information and Management 12Th International Conference Aaim 2018 Dallas TX Usa December 3 4 2018 Proceedings Shaojie Tang Ebook All Chapter PDF

Uploaded by

Copyright:

Available Formats

Algorithmic Aspects in Information and

Management: 12th International

Algorithmic Aspects in Information and Management: 12th

Brain Informatics International Conference BI 2018

Algorithmic Aspects in Information and Management 14th

Combinatorial Optimization and Applications 12th

Information Systems Security: 14th International

Neural Information Processing: 25th International

Neural Information Processing: 25th International

Neural Information Processing: 25th International

David Woodruff Sergiy Butenko (Eds.)

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Library of Congress Control Number: 2018963820

LNCS Sublibrary: SL1 – Theoretical Computer Science and General Issues

© Springer Nature Switzerland AG 2018

AAIM 2018 Co-chairs’ Message

September 2018 Shaojie Tang

Technical Program Committee

Minimum Diameter k-Steiner Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Factors Impacting the Label Denoising of Neural Relation Extraction . . . . . . 12

Makespan Minimization on Unrelated Parallel Machines with a Few Bags . . . 24

Channel Assignment with r-Dynamic Coloring . . . . . . . . . . . . . . . . . . . . . . 36

Profit Maximization Problem with Coupons in Social Networks . . . . . . . . . . 49

A Bicriteria Approximation Algorithm for Minimum Submodular Cost

A Novel Approach to Verifying Context Free Properties of Programs . . . . . . 74

Determination of Dual Distances for a Kind of Perfect Mixed Codes. . . . . . . 88

Approximation and Competitive Algorithms for Single-Minded

An Empirical Analysis of Feasibility Checking Algorithms

Quality-Aware Online Task Assignment Using Latent Topic Model . . . . . . . 124

Calibration Scheduling with Time Slot Cost . . . . . . . . . . . . . . . . . . . . . . . . 136

The k-power Domination Problem in Weighted Trees . . . . . . . . . . . . . . . . . 149

General Rumor Blocking: An Efficient Random Algorithm

A Robust Power Optimization Algorithm to Balance Base Stations’ Load

Faster Compression of Patterns to Rectangle Rule Lists . . . . . . . . . . . . . . . . 190

Algorithm Designs for Dynamic Ridesharing System. . . . . . . . . . . . . . . . . . 209

New LP Relaxations for Minimum Cycle/Path/Tree Cover Problems . . . . . . . 221

Computation of Kullback-Leibler Divergence Between Labeled Stochastic

Order Preserving Barrier Coverage with Weighted Sensors on a Line . . . . . . 244

Achieving Location Truthfulness in Rebalancing Supply-Demand

Approximation Algorithms and a Hardness Result for the Three-Machine

A New Algorithm Design Technique for Hard Problems, Building

Community-Based Acceptance Probability Maximization for Target Users

Knowledge Graph Embedding Based on Subgraph-Aware Proximity . . . . . . . 306

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319

Wei Ding1(B) and Ke Qiu2

Abstract. Given an edge-weighted undirected graph G = (V, E, w) and

Keywords: k-Steiner forest · Diameter · Modification

1.1 Related Results

collection D = {(si , ti ) : 1 ≤ i ≤ K} of K terminal pairs, the aim of GSTP is

1.2 Our Results

In the real-world problems, we are sometimes required to ﬁnd a Steiner forest

further develop a better approximation algorithm by modiﬁcation. In Sect. 5, we

2.1 Definitions and Notations

Let G = (V, E, w) be an edge-weighted undirected graph, where V =

r(T , X ) = max d(ti , X ). (1)

This paper ﬁrst deals with the following problem.

Definition 1. Given an edge-weighted undirected graph G = (V, E, w), a sub-

The discrete (vertex) Steiner k-center problem (DSkCP) is the special

both edge-disjoint and vertex-disjoint. One is called a pseudo Steiner forest if it

diam(Fj ) = max dj (ti1 , ti2 ). (3)

Definition 2. Given an edge-weighted undirected graph G = (V, E, w), a subset

2.2 Fundamental Properties

Theorem 1. Given a k-set with no duplicates, X = {x1 , x2 , . . . , xk } ⊂ P, we