close, E in Eq. (1) is assigned a large value. When the query
and the nonrelevant cluster are not close, the value of E is
set close to the conventional value of 0.5.
Thus, parameters D and E in Eq. (1) are set as follows,
based on the maximum inner product of the normalized
query vector and the center vectors of the relevant and
nonrelevant clusters [8]:
Here cG
r
and cG
n
are the center vectors of G
r
and G
n
,
respectively. The function is assumed continuous at the
thresholds p
k
,
r
a and p
k
,
n
b. The constants c
1
D
, c
2
D
,
c
1
E
, and
c
2
E
are determined by setting
a
, b, D
p
k
,
r
0
and
E
p
k
,
n
1
to suitable positive values. In the experiment
assuming that the users information needs change dynami-
cally, two cases are examined: the case in which parameters
D and E in Eq. (1) have fixed values of 2 and 0.5, respec-
tively, and the case in which the values of
a
,
b
, D
p
k
,
r
0
and
E
p
k
,
n
1
are empirically determined and the values
of
D
and
E are adjusted by Eqs. (2) and (3). Comparing the
results of these two cases, we verified that the retrieval
effectiveness was greatly improved by adjusting the pa-
rameters [11].
2.2. Incremental query expansion based on
document clustering
It has been shown in the application of relevance
feedback that performance measures such as the retrieval
effectiveness are improved by displaying the process to the
user and making the query expandable by the user [12]. In
one proposed interface, each time the user makes a judg-
ment that a retrieved document is relevant or nonrelevant,
the candidate query is incrementally updated, and the user
manipulates the query expansion based on the candidate
[13].
In AIQEC proposed by the authors [7, 8], each time
a cluster created by clustering retrieval results is selected
by the user as containing relevant (or nonrelevant) informa-
tion, the query used in the retrieval is successively expanded
or refined by relevance feedback with the adaptive parame-
ters as described in Section 2.1. The result is then presented
to the user. The query used in the retrieval is called the
context query, and the query which is automatically modi-
fied is called the suggested query. The user is allowed to
delete unnecessary words from the context query and the
suggested query. The results are called the user-modified
context query and the user-modified suggested query, re-
spectively.
The suggested query, the user-modified context
query, and the user-modified suggested query can be used
for a re-search, thus providing feedback not from the indi-
vidual document, but from the document cluster, which
should yield an unbiased query. The adaptive parameters
make it possible to obtain a query matched to the users
information needs, even if the needs are vague or change
dynamically. By applying AIQEC to the CBB technique
described in Section 3, the query is incrementally modified
through interaction between the system and the user during
browsing of the clustered retrieval results [7, 8].
3. Cluster-Based Browsing
CBB is one of the component techniques in the
proposed method, as shown in Fig. 1. A technique for
browsing retrieval results based on the cluster hypothesis
that relevant documents are more similar to each other than
to nonrelevant documents has been proposed [46]. First,
the results of document search are clustered on the basis of
the distances between individual documents. Then the user
selects relevant clusters, and the set of documents contained
in these clusters is again clustered by the system. By iterat-
ing such interactions several times, the system helps to
narrow the set of documents to those appropriate to the
users interests.
CBB uses a seed-based nonhierarchical clustering
algorithm. The procedure must contain three layers:
(1) finding k initial seed points,
(2) assigning each document to a seed point, and
(3) refining the partition so constructed.
This paper uses the fractionation algorithm [4] of
Cutting and colleagues to find the initial seed points in the
seed-based nonhierarchical clustering algorithm. The frac-
tionation algorithm is designed for interactive processing
and emphasizes speed rather than clustering accuracy. It
assumes the availability of an algorithm that can achieve
highly accurate clustering even if the execution speed is
(2)
(4)
(5)
(6)
(7)
(3)
(8)
(9)
66
剩余10页未读，继续阅读
weixin_38705873
- 粉丝: 7
- 资源: 926
我的内容管理
展开
最新资源
- 解决Eclipse配置与导入Java工程常见问题
- 真空发生器：工作原理与抽吸性能分析
- 爱立信RBS6201开站流程详解
- 电脑开机声音解析：故障诊断指南
- JAVA实现贪吃蛇游戏
- 模糊神经网络实现与自学习能力探索
- PID型模糊神经网络控制器设计与学习算法
- 模糊神经网络在自适应PID控制器中的应用
- C++实现的学生成绩管理系统设计
- 802.1D STP 实现与优化：二层交换机中的生成树协议
- 解决Windows无法完成SD卡格式化的九种方法
- 软件测试方法：Beta与Alpha测试详解
- 软件测试周期详解：从需求分析到维护测试
- CMMI模型详解：软件企业能力提升的关键
- 移动Web开发框架选择：jQueryMobile、jQTouch、SenchaTouch对比
- Java程序设计试题与复习指南
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
