See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221239815

Extracting Content Structure for Web Pages Based on Visual Representation

Conference Paper in Lecture Notes in Computer Science · April 2003

DOI: 10.1007/3-540-36901-5_42 · Source: DBLP

CITATIONS READS

345 1,565

4 authors, including:

Ji-Rong Wen Wei-Ying Ma

Renmin University of China Toutiao
588 PUBLICATIONS 17,883 CITATIONS 396 PUBLICATIONS 33,562 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ji-Rong Wen on 20 March 2015.

The user has requested enhancement of the downloaded file.

 Extracting Content Structure for Web Pages based on
Visual Representation

Deng Cai1*, Shipeng Yu 2*, Ji-Rong Wen* and Wei-Ying Ma*

* Microsoft Research Asia

{jrwen, wyma}@microsoft.com
1 Tsinghua University, Beijing, P.R.China

[email protected]
2 Peking University, Beijing, P.R.China

[email protected]

Abstract. A new web content structure based on visual representation is pro-

posed in this paper. Many web applications such as information retrieval, in-
formation extraction and automatic page adaptation can benefit from this struc-
ture. This paper presents an automatic top-down, tag-tree independent approach
to detect web content structure. It simulates how a user understands web layout
structure based on his visual perception. Comparing to other existing tech-
niques, our approach is independent to underlying documentation representa-
tion such as HTML and works well even when the HTML structure is far dif-
ferent from layout structure. Experiments show satisfactory results.

1 Introduction

Today the Web has become the largest information source for people. Most informa-
tion retrieval systems on the Web consider web pages as the smallest and undividable
units, but a web page as a whole may not be appropriate to represent a single seman-
tic. A web page usually contains various contents such as navigation, decoration, in-
teraction and contact information, which are not related to the topic of the web page.
Furthermore, a web page often contains multiple topics that are not necessarily rele-
vant to each other. Therefore, detecting the content structure of a web page could
potentially improve the performance of web information retrieval.
Many web applications can utilize the content structures of web pages. For exam-
ple, some researchers have been trying to use database techniques and build wrappers
for web documents [3]. If a web page can be divided into semantic related parts, wrap-
pers can be more easily matched and data can be more likely extracted. Link structure
analysis can also make good use of the content structures of web pages. Links at
different parts of a page usually act as different functions and contribute to the Page-
Rank [2] or HITS [11] differently. Recent works on topic distillation [4] and focused
crawling [5] show the usefulness of page segmentation on information analysis.
Furthermore, adaptive content delivery on small handheld devices also requires the
detection of underlying content structure of a web page to facilitate the browsing of a
large page by partitioning it into smaller units [10].
 People view a web page through a web browser and get a 2-D presentation which
has many visual cues to help distinguish different parts of the page. Generally, a web
page designer would organize the content of a web page to make it easy for reading.
Thus, semantically related content is usually grouped together and the entire page is
divided into regions for different contents using explicit or implicit visual separators
such as lines, blank areas, images, font sizes, colors, etc [16]. This motivates our work
to segment a web page into semantically related content blocks from its visual presen-
tation. If we can reconstruct the structure of a page corresponding to human visual
perception, it will better reflect the semantic structure. In this paper, we propose VIPS
(VIsion-based Page Segmentation) algorithm to extract the content structure for a web
page. The algorithm makes full use of page layout features and tries to partition the
page at the semantic level. Each node in the extracted content structure will corre-
spond to a block of coherent content in the original page.
The paper is organized as follows. Section 2 presents the related works. Section 3
defines the web page content structure based on vision. The details of the VIPS algo-
rithm are introduced in Section 4. The experimental results are reported in Section 5.
Section 6 summarizes our contributions and concludes the paper.

2 Related Works

The applications mentioned in the introduction indicate the need of techniques for
extracting the content structure of a web page. Tag tree or Document Object Model
(http://www.w3.org/DOM/) provides each web page a fine-grained structure, illustrat-
ing not only the content but also the presentation of the page. Many researchers [10,
12, 15] have considered using the tag information and dividing the page based on the
type of the tags. Useful tags include <P> (for paragraph), <TABLE> (for table),
<UL> (for list), <H1>~<H6> (for heading), etc.
Some other algorithms also consider the content or link information besides the tag
tree. Embley [8] use some heuristic rules to discover record boundaries within a page,
which assist data extraction from the web page. Chakrabarti [4] addresses the fine-
grained topic distillation and dis-aggregates hubs into regions by analyzing link struc-
ture as well as intra-page text distribution.
A Function-based Object Model (FOM) of a web page is proposed by Chen [6] for
content understanding and adaptation. Every undividable element in the tag tree is
called a basic object and can be grouped into a composite object. A function type can
be defined to each object and helps to build a hierarchical structure for the page.
However, the grouping rules and the functions are hard to define accurately, and thus
make the whole tree-constructing process very inflexible.
In [16] and [6], some visual cues are used in DOM analysis. They try to identify the
logic relationships within the web content based on visual layout information, but
these approaches still rely too much on the DOM structure. Gu [9] tries to construct a
web content structure by breaking out the DOM tree and comparing similarity among
all the basic DOM nodes. Since a normal web page may have hundreds of basic ele-
ments, the algorithm is time-consuming and inflexible, not capable to deal with a large
amount of web pages.
3 Vision-based Content Structure for Web Pages

Similar to [6], we define the basic object as the leaf node in the DOM tree that cannot
be decomposed any more. Although a DOM structure provides a hierarchy for the
basic objects in a web page, it is more for representation rather than content organiza-
tion. In this paper, we propose the vision-based content structure, where every node,
called a layout block, is a basic object or a group of basic objects. The nodes in the
content structure do not necessarily have a mapping to the nodes in the DOM tree.
Similar to the description of document representation in [14], the basic model of vi-
sion-based content structure for web pages is described as follows.
A web page Ω is specified by a triple Ω = ( O , Φ , δ ) . O = {Ω , Ω , ..., Ω } is a finite
1 2 N

set of objects or sub-web-pages. All these objects are not overlapped. Each object can
be recursively viewed as a sub-web-page and has a subsidiary content structure.
Φ = {ϕ , ϕ , ..., ϕ } is a finite set of visual separators, including horizontal separators
1 2 T

and vertical separators. Every separator has a weight indicating its visibility, and all
the separators in the same Φ have same weight. δ is the relationship of every two
blocks in O and can be expressed as: δ = O × O → Φ ∪ { NULL} . Suppose Ωi and Ωj
are two objects in Ο, δ ( Ω , Ω ) ≠ NULL indicates that Ωi and Ωj are exactly sepa-
rated by the separator δ ( Ω , Ω ) or we can say the two objects are adjacent to each
i j

i j

other, otherwise there are other objects between the two blocks Ωi and Ωj.
Since each Ωi is a sub-web-page of the original page, it has similar content structure
as Ω. Recursively, we have Ω = ( O , Φ , δ ) , O = Ω , Ω , ..., Ω
t t t t
,
t
{ 1 2 N st
}
{ }
and δ = O × O → Φ ∪ { NULL} where Ω is the tth object in
s s s s s st st st

Φ = ϕ , ϕ , ..., ϕ
t 1 2 Tst t t t t t

s st st st s s s s s
t
the sub-web-page level s, Nst and Tst are the number of objects in O and number of s

separators in Φ .
t

Fig. 1 shows an example of visual-based content structure for Yahoo! Auctions

page. It illustrates the layout structure and the vision-based content structure of the
page. In the first level, the original web page has four objects or visual blocks
VB1~VB4 and three separators ϕ ~ ϕ , as specified in Fig. 1(d). Then we can further
1 3

construct sub content structure for each sub web page. For example, VB2 has three
offspring objects and two separators. It can be further analyzed like Fig. 1(e).
For each visual block, the Degree of Coherence (DoC) is defined to measure how
coherent it is. DoC has the following properties:
• Ranges from 0 to 1;
• The greater the DoC value, the more consistent the content within the block;
• In the hierarchy tree, the DoC of the child is not smaller than its parent’s.
We can pre-define the Permitted Degree of Coherence (PDoC) to achieve different
granularities of content structure for different applications. The smaller the PDoC is,
the coarser the content structure would be. For example in Fig. 1(a), the visual block
VB2_1 may not be further partitioned with an appropriate PDoC.
The vision-based content structure is more likely to provide a semantic partitioning
of the page. Every node, especially the leaf node, is more likely to convey a semantic
meaning for building a higher semantic via the hierarchy. For instance, in Fig. 1(a) we
can say that VB2_1_1 denotes the category links of Yahoo! Shopping auctions, and
that VB2_2_1 and VB2_2_2 show details for two different comics.
 ϕ1

ϕ12 ϕ22
ϕ2
ϕ3

(a) (b)

(c)
Ο = ( VB1, VB 2, VB 3, VB 4 ) VB 2 = ( VB 2 _ 1, VB 2 _ 2, VB 2 _ 3 )

{
Φ = ϕ ,ϕ ,ϕ
1 2 3

} 2

{
Φ = ϕ ,ϕ
1

2
2

2
}
 ( VB1, VB 2 )   ϕ   (VB 2 _ 1, VB 2 _ 2 )   ϕ 
1 1

 ( VB 2, VB 3 )   ϕ   (VB 2 _ 2, VB 2 _ 3 )  =  ϕ 
2

δ
=    
2 2

δ  2

 ( VB 3, VB 4 )   ϕ     NULL 

3

else
   
 else   NULL 
(d) (e)
Fig. 1. The layout structure and vision-based content structure of an example page. (d) and (e)
show the corresponding specification of vision-based content structure.

4 The VIPS Algorithm

In the VIPS algorithm, the vision-based content structure of a page is deduced by

combining the DOM structure and the visual cues. The segmentation process is illus-
trated in Fig. 2. First, DOM structure and visual information, such as position, back-
ground color, font size, font weight, etc., are obtained from a web browser. Then,
from the root node, the visual block extraction process is started to extract visual
blocks of the current level from the DOM tree based on visual cues. Every DOM node
is checked to judge whether it forms a single block or not. If not, its children will be
processed in the same way. When all blocks of the current level are extracted, they are
put into a pool. Visual separators among these blocks are identified and the weight of
a separator is set based on properties of its neighboring blocks. After constructing the
layout hierarchy of the current level, each newly produced visual blocks is checked to
see whether or not it meets the granularity requirement. If no, this block will be further
partitioned. After all blocks are processed, the final vision-based content structure for
the web page is outputted. Below we introduce the visual block extraction, separator
detection and content structure construction phases respectively.

Fig. 2. The vision-based page segmentation algorithm

4.1 Visual Block Extraction

In this phase, we aim at finding all appropriate visual blocks contained in the current
sub-tree. In general, every node in the DOM tree can represent a visual block. How-
ever, some “huge” nodes such as <TABLE> and <P> are used only for organization
purpose and are not appropriate to represent a single visual block. In these cases, the
current node should be further divided and replaced by its children. On the other hand,
we may not extract all leaf nodes in the DOM tree due to their high volume.
At the end of this step, for each node that represents a visual block, its DoC value is
set according to its intra visual difference. This process is iterated until all appropriate
nodes are found to represent the visual blocks in the web page.
The visual block extraction algorithm DivideDomtree is illustrated in Fig. 3. Some
important cues are used to produce heuristic rules in the algorithm are:
• Tag cue: Tags such as <HR> are often used to separate different topics from vis-
ual perspective. Therefore we prefer to divide a DOM node if it contains these
tags.
• Color cue: We divide a DOM node if its background color is different from one
of its children’s.
• Text cue: If most of the children of a DOM node are Text nodes (i.e., no tags
surround them), we do not divide it.
 Algorithm DivideDomtree(pRoot, nLevel) Algorithm Dividable(pRoot, nLevel)
{ {
IF (Dividable(pRoot, nLevel) == TRUE){ IF (pRoot is the Top Block){
FOR EACH child OF pRoot { RETURN TRUE;
DivideDomtree(child, nLevel); } ELSE {
} Special routines for TABLE, TR,
} ELSE { TBODY, TD, P, UL, FORM;
Put the sub-tree (pRoot) into the Heuristic rules for general tags;
pool as a block; }
} }
}

(a) (b)
Fig. 3. The visual block extraction algorithm

• Size cue: We prefer to divide a DOM node if the standard deviation of size of its
children is larger than a threshold.
The Tag cue and Color cue are relatively straightforward to produce corresponding
heuristic rules. Below we give the details of how to use text and size cues to generate
heuristic rules for block extraction. First, several definitions are given:
1. Valid Node: a node that can be seen through the browser. The node’s width and
height are not equal to zero.
2. Block Node: the node with tag not <A>, <B>, <FONT>, <HR>, <I>, <P>,
<STRONG>, <TEXT>.
3. Text Node: the DOM node that only contains free text.
4. Virtual Text Node: The node that is not a block node and only have text node as
children.
According to the above definitions, we use the following text cue and size cue
based heuristic rules to further enhance the block extraction process:
• A node will be dropped if it has no valid child.
• If a node only has one valid child and this child is not Text node, then trace into
the child.
• If all the children of a node are Text nodes or Virtual Text nodes, then set the
DoC 1.
• If the node’s size is 3 times greater than all his children’s total size, divide it.
• If the node has Text node child or Virtual Text node child and the node’s width
or height is smaller than a threshold, set the DoC 0.8.
• Split the node which has more than two successive <BR> children. (It means
there are many space in the middle, may be two different topics)
In addition, some tags, such as <TABLE>, <TBODY>, <TR>, <TD>, <P>, <UL>
and <LI>, are very important and common in web page and are more likely to form a
content coherent sub-tree. So we define some special routines in our algorithm to
handle these tags and set higher thresholds for these tags in the above rules.
Take Fig. 1 as an example. VB1, VB2_1, VB2_2, VB2_3, VB3 and VB4 will be
extracted at the first round. The detailed process is as follows. In Fig. 4, a DOM tree
structure of VB2 is shown. In the block extraction process, when the <TABLE> node
is met, it has only one valid child <TR>. We trace into the <TR> node according to
the heuristic rules. The <TR> node has five <TD> children and only three of them are
valid. The first child’s background color is different from its parent’s background
color, so the <TR> node is split and the first <TD> node is not divided further in this
round and is put into the pool as a block. The second and fourth child of <TR> node is
not valid and will be dropped. When the third and fifth children of <TR> are consid-
ered, we use another rule to improve the efficiency – because the first <TD> is not
divided, so there will be no horizontal separators in this round projection. It is unnec-
essary to divide the third and fifth <TD> and they should be taken as a block in this
level.

Fig. 4. DOM tree structure of VB2 in the sample page

4.2 Visual Separator Detection

After all blocks are extracted, they are put into a pool for visual separator detection.
Separators are horizontal or vertical lines in a web page that visually cross with no
blocks in the pool. From a visual perspective, separators are good indicators for dis-
criminating different semantics within the page. A visual separator is represented by a
2-tuple: (Ps, Pe), where Ps is the start pixel and Pe is the end pixel. The width of the
separator is calculated by the difference between these two values.

Separator Detection. The visual separator detection algorithm is described as

follows:
1. Initialize the separator list. The list starts with only one separator (Pbe, Pee) whose
start pixel and end pixel are corresponding to the borders of the pool.
2. For every block in the pool, the relation of the block with each separator is evalu-
ated
• If the block is contained in the separator, split the separator;
• If the block crosses with the separator, update the separator’s parameters;
• If the block covers the separator, remove the separator.
3. Remove the four separators that stand at the border of the pool.

Take Fig. 5(a) as an example in which the black blocks represent the visual blocks
in the page. For simplicity we only show the process to detect the horizontal separa-
tors. At first we have only one separator that is the whole pool. As shown in Fig. 5(b),
when we put the first block into the pool, it splits the separator into S1 and S2. It is
same with the second and third block. When the fourth block is put into the pool, it
crosses the separator S2 and covers the separator S3, the parameter of S2 is updated
and S3 is removed. At the end of this process, the two separators S1 and S3 that stand
at the border of the pool are removed.
 (a) (b)
Fig. 5. A sample page and the separator detection process

Setting Weights for Separators. The separators are used to distinguish blocks with
different semantics, so the weight of a separator can be assigned based on the
information difference between its neighboring blocks. The following rules are used to
set a weight to each separator:
• The more the distance between blocks on different side of the separator, the higher
the weight.
• If a visual separator is at the same position as some tags such as <HR>, its weight is
made higher.
• If the differences of font properties such as font size and font weight are more
clearly on two sides of the separator, the weight will be increased. More over, the
weight will be increased if font size before the separator is smaller than that after
the separator.
• If background colors are different on two sides of the separator, the weight will be
increased.
• When the structures of the blocks beside the separator are very similar (e.g. both
are text), the weight of the separator will be decreased.
Take the third <TD> node in Fig. 4 as an example. The sub-page corresponding to
this node is shown in Fig. 6(b) and the DOM tree structure is shown in Fig. 6(a). We
can see that many nodes in the DOM tree are invalid in our definition and cannot be
seen on the page. They are ignored in the block extraction process. After the block
extraction phase, six blocks are put in a pool and five horizontal separators are de-
tected. Then the weights of these separators are set based on the rules we described. In
this example, the separator between Block 2 and Block 3 will get higher weight than
the separator between Block 1 and Block 2 because of the different font weights. For
the same reason, the separator between Block 4 and Block 5 will also get a high
weight. The final separators and weights are shown in Fig. 6(c), in which a thicker line
means a higher weight.

4.3 Content Structure Construction

When separators are detected and separators’ weights are set, the content structure can
be constructed accordingly. The construction process starts from the separators with
the lowest weight and the blocks beside these separators are merged to form new vir-
tual blocks. This process iterates till separators with maximum weights are met. The
DoC of each new block is also set via similar methods described in Section 4.1.
 (a) (b) (c) (d)
Fig. 6. Separators and weights among blocks

After that, each leaf node is checked whether it meets the granularity requirement.
For every node that fails, we go to the Visual Block Extraction phase again to further
construct the sub content structure within that node. If all the nodes meet the require-
ment, the iterative process is then stopped and the vision-based content structure for
the whole page is obtained. The common requirement for DoC is that DoC > PDoC, if
PDoC is pre-defined.
Take Fig. 6 as an example. In the first iteration, the first, third and fifth separators are
chosen and Block 1 and 2 are merged to form the new block VB2_2_2_1. Similar
merging is conducted for Block 3 and 4 (resulting a new block VB2_2_2_2) and
Block 5 and 6 (resulting a new block VB2_2_2_3). The new blocks VB2_2_2_1,
VB2_2_2_2 and VB2_2_2_3 are the children of VB2_2_2 and can also be viewed as
a partition of VB2_2_2. Every leaf node, such as VB2_2_2_1_1, VB2_2_2_1_2 and
VB2_2_2_2_1, will be checked to see whether it meets the granularity requirement.
After several iterations, the final vision-based content structure of the page is con-
structed.
In summary, the proposed VIPS algorithm takes advantage of visual cues to obtain
the vision-based content structure of a web page and thus successfully bridges the gap
between the DOM structure and the semantic structure. The page is partitioned based
on visual separators and structured as a hierarchy closely related to how a user would
browse the page. Content related parts could be grouped together even if they are in
different branches of the DOM tree.
VIPS is also very efficient. Since we trace down the DOM structure for visual
block extraction and do not analyze every basic DOM node, the algorithm is totally
top-down. Furthermore, the PDoC can be pre-defined, which brings significant flexi-
bility to segmentation and greatly improve the performance.
5 Experiments

We provide some performance evaluation of our proposed VIPS algorithm based on a

large collection of web pages from Yahoo. We also conduct experiments to evaluate
how the algorithm can be used to enhance information retrieval on the Web.

5.1 Performance of VIPS algorithm

We selected 140 web pages from popular sites listed in 14 main category of Yahoo
directory (http://www.yahoo.com). The web content structure detection algorithm is
run against these pages and the results are assessed by human judgments. Table 1
shows the result.

Table 1. Evaluation of the quality of page analysis

Human judgment Number of pages

Perfect 86
Satisfactory 50
Failed 4

As can be seen, 86+50=136 (97%) pages have their content structures correctly de-
tected. For those “failed” cases, one major reason is that the browser (i.e. Internet
Explorer in our experiments) provides wrong position information so that our algo-
rithm cannot get the correct content structure. Another reason lies in that several pages
use images (e.g., a very thin image that represents a line) to divide different content
blocks. Our algorithm currently does not handle this situation.

5.2 Experiments on web information retrieval

Query expansion is an efficient way to improve the performance of information re-

trieval [7]. The quality of expansion terms is heavily affected by the top-ranked
documents. Noise and multi-topics are two major negative factors for expansion term
selection in the web context. Since our VIPS algorithm can group semantically related
content into a segment, the term correlations within a segment will be much higher
than those in other parts of a web page. With improved term correlations, high-quality
expansion terms can be extracted from segments and used to improve information
retrieval performance.
We choose Okapi [13] as the retrieval system and WT10g [1] in TREC-9 and
TREC 2001 Web Tracks as the data set. WT10g contains 1.69 million pages and
amounts to about 10G. We use the 50 queries from TREC 2001 Web Track as the
query set and only the TOPIC field for retrieval, and use Okapi’s BM2500 as the
weight function and set k1 = 1.2, k3 = 1000, b = 0.75, and avdl = 61200.
An initial list of ranked web pages is obtained by using any traditional information
retrieval methods. Then we apply our page analysis algorithm with a PDoC 0.6 to the
top 80 pages and get the set of candidate segments. The most relevant (e.g. top 20)
segments from this candidate set are used to select expansion terms. These selected
terms are used to construct a new expanded query to retrieve the final results.
We compared our method with the traditional pseudo-relevance feedback algo-
rithm. Note that our method selects expansion terms from blocks while traditional
methods select expansion terms from entire web pages. The experimental result is
shown in Fig. 7.

)% 21
(
no
19
is
ic
er 17
P
eg 15
ar
ev 13
A 3 5 10 20 30 40 50 60

Number of Segments

Baseline FULLDOC VIPS

Fig. 7. Performance comparison of pseudo-relevance feedback based on two different ways of

selecting query expansion terms. VIPS is our method which selects query expansion terms from
blocks while FULLDOC represents traditional approaches where the entire pages are used for
expansion term selection.

From Fig. 7, we can see that pseudo-relevance feedback based on web page content
structure significantly outperforms traditional methods and achieves about 27% per-
formance improvement on the Web Track dataset (refer to [17] for more details). The
experiments clearly show that web page content structure is very helpful to detect and
filter out noisy and irrelevant information. Thus better expansion terms can be se-
lected to improve retrieval performance.

6 Conclusion

In this paper a new approach of extracting web content structure based on visual rep-
resentation was proposed. The produced web content structure is very helpful for
applications such as web adaptation, information retrieval and information extraction.
By identifying the logic relationship of web content based on visual layout informa-
tion, web content structure can effectively represent the semantic structure of the web
page. An automatic top-down, tag-tree independent and scalable algorithm to detect
web content structure was presented. It simulates how a user understands the layout
structure of a web page based on its visual representation. Compared with traditional
DOM based segmentation method, our scheme utilizes useful visual cues to obtain a
better partition of a page at the semantic level. It also is independent of physical reali-
zation and works well even when the physical structure is far different from visual
presentation. The experimental result has shown that our proposed scheme achieves
 very satisfactory performance. We plan to apply the scheme for adaptive content de-
livery to facilitate better web browsing on mobile devices for our future works.

References

1. Bailey, P., Craswell, N., and Hawking, D., Engineering a multi-purpose test collection for
Web retrieval experiments, Information Processing and Management, 2001.
2. Brin, S. and Page, L., The Anatomy of a Large-Scale Hypertextual Web Search Engine, In
the Seventh International World Wide Web Conference, Brisbane,Australia, 1998.
3. Buneman, P., Davidson, S., Fernandez, M., and Suciu, D., Adding Structure to Unstructured
Data, In Proceedings of the 6th International Conference on Database Theory (ICDT'97),
1997, pp. 336-350.
4. Chakrabarti, S., Integrating the Document Object Model with hyperlinks for enhanced topic
distillation and information extraction, In the 10th International World Wide Web Confer-
ence, 2001.
5. Chakrabarti, S., Punera, K., and Subramanyam, M., Accelerated focused crawling through
online relevance feedback, In Proceedings of the eleventh international conference on
World Wide Web (WWW2002), 2002, pp. 148-159.
6. Chen, J., Zhou, B., Shi, J., Zhang, H., and Wu, Q., Function-Based Object Model Towards
Website Adaptation, In the 10th International World Wide Web Conference, 2001.
7. Efthimiadis, N. E., Query Expansion, In Annual Review of Information Systems and Tech-
nology, Vol. 31, 1996, pp. 121-187.
8. Embley, D. W., Jiang, Y., and Ng, Y.-K., Record-boundary discovery in Web documents, In
Proceedings of the 1999 ACM SIGMOD international conference on Management of data,
Philadelphia PA, 1999, pp. 467-478.
9. Gu, X., Chen, J., Ma, W.-Y., and Chen, G., Visual Based Content Understanding towards
Web Adaptation, In Second International Conference on Adaptive Hypermedia and Adap-
tive Web-based Systems (AH2002), Spain, 2002, pp. 29-31.
10. Kaasinen, E., Aaltonen, M., Kolari, J., Melakoski, S., and Laakko, T., Two Approaches to
Bringing Internet Services to WAP Devices, In Proceedings of 9th International World-
Wide Web Conference, 2000, pp. 231-246.
11. Kleinberg, J., Authoritative sources in a hyperlinked environment, In Proceedings of the
9th ACM-SIAM Symposium on Discrete Algorithms, 1998, pp. 668-677.
12. Lin, S.-H. and Ho, J.-M., Discovering Informative Content Blocks from Web Documents,
In Proceedings of ACM SIGKDD'02, 2002.
13. Robertson, S. E., Overview of the okapi projects, Journal of Documentation, Vol. 53, No.
1, 1997, pp. 3-7.
14. Tang, Y. Y., Cheriet, M., Liu, J., Said, J. N., and Suen, C. Y., Document Analysis and
Recognition by Computers, Handbook of Pattern Recognition and Computer Vision, edited
by C. H. Chen, L. F. Pau, and P. S. P. Wang World Scientific Publishing Company, 1999.
15. Wong, W. and Fu, A. W., Finding Structure and Characteristics of Web Documents for
Classification, In ACM SIGMOD Workshop on Research Issues in Data Mining and
Knowledge Discovery (DMKD), Dallas, TX., USA, 2000.
16. Yang, Y. and Zhang, H., HTML Page Analysis Based on Visual Cues, In 6th International
Conference on Document Analysis and Recognition, Seattle, Washington, USA, 2001.
17. Yu, S., Cai, D., Wen, J.-R., and Ma, W.-Y., Improving Pseudo-Relevance Feedback in Web
Information Retrieval Using Web Page Segmentation, To appear in the Twelfth Interna-
tional World Wide Web Conference (WWW2003), 2003.

View publication stats