Evaluate the data available in Web pages is necessary to allow the suggestion of adequate content to a specific public, for such, an idea is to develop an automatic detection mechanism of the quality present in Web pages content. Due to its quickly and not standardized growth, the content freely available in the Web has reached very large proportions, becoming extremely difficult to automatic manage this large mass of data. One of the main reason for this complexity is the not appliance of standards for its construction, which should lead to a standardized access. This article describes the development of models and techniques to locate, standardize and extract the content in web pages associated with health issues. After that, the objective is to provide an appropriate content to evaluate the quality of a web page according to specific metrics.
One of the biggest challenges in Web systems is to automatically deal with the large amount of data available in the Web as a large distributed system. As a consequence of its rapid not standardized growing, the Web reached a huge proportion that became extremely difficult to be efficiently controlled, being the automatic management of data very complex.
The automatic web-data management is especially desired when the subject is the quality evaluation of a given web page for its subsequent recommendation to a specific public. In the case of health information content it is necessary a careful check on the validity of the information before an effective recommendation. We know that check all health information on the Web to determine if is reliable it‘s very difficult, but we can evaluate a number of indicators present in a web page to try to guarantee a minimum of quality. Examples of indicators, or metrics, could be: sponsors, if exists and who are; the subject of the page; the ways of contact with the responsible by the web page; the web page readability; if the content is continuously updated (freshness).
To achieve this objective there are organizations that grant
certificates of quality in form of quality-stamps to ensure the
quality of the information presented on a website. That is the case of
Web Mèdica Acreditada1 (WMA), an organization of the
Medical Association of Barcelona that offers ―a quality programme
addressed to medical websites. Through a voluntarily
certification process, websites that follow that programme meet a set of
quality criteria, making possible a trustworthy virtual
community on the Internet for general public, patients and health
professionals‖ [1]. Other examples of organizations which grant
quality seals to websites are the Internet Content Rating Association
(ICRA) [
To enable automatic retrieval of data from web pages about health, some models and techniques are presented here to support: (i) location of health web-data (ii) standardization and (iii) (ii) extraction of this data, based on pre-established criteria. Conscious of the complexity of dealing to the entire Web, the location, standardization and extraction models presented here are applied to specific search engines.
The remainder of this article is structured as follows. Section 2 explores some related works. Section 3 describes the general vision of this work, showing our starting point and the relevant data in estimating the quality of a web page. Section 4 is associated with the task of automatic location of web page about health, explaining its general model. Section 5 present the task of automatic extraction data from the localized web pages pre
senting problems related to the lack of standardization. Section 6 shows a case study developed on pages related with Alzheimer‘s disease. Finally section 7 presents the conclusions and future works.
An example of project related to the analysis of quality content
on health web pages is the QUATRO Project (The Quality
Assurance and Content Description Project) and its successor,
QUATRO+ [
Related to information extraction, the Project AQUA (Assisting
Quality Assessment) [
The general model presented in this work intends to cover critical points considered to achieve the final goal: evaluate the quality of a Web page about health. In a general view it is necessary to accurately automate the following steps:
Localization – answer to the question: given an item X
stored in some dynamic set of nodes in the system, how to
find it? [
Standardization – relates to existing standards in presentation of data in web pages which may indicate ways to achieve automatic access (automatic extraction) of this data.
Extraction – once the web page is located, how to obtain
and properly manage its data to be evaluated?
Quality – is the analysis of collected data where metrics
are applied to define the level of quality present in the
evaluated web page, according to a specific user profile.
As we mentioned in section 1, in this work only the phases of
Location, Standardization and Extraction are treated. The
starting point in developing models was the research described in
[
As we can see in Figure 1, information about the web page is obviously necessary to define its quality. Information about author is very important because if the content was written or revised by a specialist in the subject, possibly the quality will be enhanced. Is either important define if the web page is sponsored by some organization and if is a recognized organization, for example, a governmental organization, university or industry.
The E-R (Entity-Relationship) model of the Figure 1 describes the main classes for determining the page quality: (i) the page itself, with data related with its title, language, authorship, references and dates of creation and update; (ii) the web page author, with data related with contact and expertise in the subject covered by the web page; (iii) the organization that sponsor the Web page (if exists); (iv) the links of another pages or sites in the Web pointing to this (inLinks); and (v) which are the web links from this page to other web pages (outLinks).
In the development of the techniques to collect data on the model of Figure 1, was noted the difficulty in obtain information about the date of creation and update of web pages. This is very important in determining the quality of a web page, because expresses the content freshness. We observed that this happen in general or by the absence of this kind of information or because is presented in a non machine readable format, like in an image. Thus, the set of data in which was possible to automatic extract information in this work is related to the site language and its inLinks and outLinks. About the authorship, it was possible to determine the author expertise and contact (e-mail address). The following sections describe the models and techniques developed to perform the automatic location and extraction of web pages to obtain the set of data described on the model presented in the Figure 1.
One of the main points in the system‘s operation was the
automatic location of health web pages to be evaluated and
extracted. The results achieved must be returned to the evaluation
process in order of relevance, i.e. according to how they satisfy
the search criteria [
Given the estimated size of the Web, find the more appropriate page to the user‘s intention is not a trivial task. Its necessary a sophisticate algorithm, combined with massive computational power, to accomplish this task. For that reasons, in this work was chose to use search engines specifically oriented for healthpages retrieval. In this category of search engines the search is performed on a database of previously indexed and constantly updated pages. Some of the specialized search engines are SearchMedica4 and MedStory5, which are specific to the medical field. In the next sections is present the model and prototype developed to locate health web pages, but not focusing on specific search engines – as this choice is application-dependent. The specific configuration of this model to a practical application is presented in section 6, with a Case Study about Alzheimer‘s disease.
The model presented in the Figure 2 is based on the interaction
among the application, a database and the Internet. It is
application dependent because it needs the existing pattern for
presenting the results of the search engine used in the application, i.e.,
the internal standard to display the results to the user.
Initially, the data about the search engine (pattern related) is
required by the system. Then, the criterion to be used by the
search engine is recovered from the database and for each
criteria is stored the first URLs returned as answer by this search
engine. For more details about this model see [
ing of returned results as well as for its storage in the database
[
After recovering the URLs from the search engines, obtained at the location step, the system starts the next task, which extracts the content that will be used to evaluate the quality of the web page. As this task was idealized to occur in an automatic way a search is performed for specific attributes that are relevant to determine the page quality, those attributes can be found in the Figure 1. The initial supposition was that it should be possible to find a standardization way that allows encompassing the maximum portion of data existent in web pages in order to make the data extraction easier and faster. However, as presented in a more detailed way in the 5.1 subsection, it was not possible to use this approach because until today there are no standardization models for naming the tags or to place data in pages. Therefore, specific strategies were proposed, trying to attend the necessary requirements to the search of each data existent in the ER model (Figure 1), as: site language; authorship data; inLinks and outLinks.
The next sections present each one of those strategies, starting with troubles and adopted strategies for standardization of web pages content and structure.
The considered standardization has the objective to define the
ways in which the content of a web page could be structured.
That research is applied later as an information source to the
a higher score, and consequently, appear in a better position in
the ranking. In this work, inLinks are recovered as follows [
Through the expression cited in the previous step, the
duplicated links that are hosted in the same domain of the
page are removed;
When counting the number of outLinks, the duplicates are
removed and the links that point to the same domain are
counted as only one entry.
extraction phase. The structure and content of search engines, as
well as web pages, is extremely dynamic. Consequently, to
overcome this constantly changes the standardization model adopted
here apply terms defined in the WorldNet [
There are others interesting approaches that can be applied as a
standardization strategy. An example is the application of
ontologies [
In this work, the recognition of HTML tags in web pages had
generated special attention. Techniques for that are applied in
areas as Information Storage and Retrieval to identify the
relevant topic of a page, since that definition is very important to the
information filtering phase [
These step objectives find out the language in which the text of
the web page was written. In this project the system informs the
URL of the health web page that must to be analyzed. The first
step refers to search for meta tags that indicate the language of
the web page as defined in the Dublin Core pattern: meta tag
language (1)[
<meta name="language" content=”english”/>
(1)
If the system didn't found the meta tag, is necessary to make a
deeper analysis on the content of the web page to verify that. In
this case, some programs that analyses the text in order to detect
the language can be used. Examples of analyzers used in the
application are: Fuzzums6, Applied Language7 e Google
Language Detection8. After obtaining the page content subsets of
text, as well as the keyword used in this search, are extracted
from the content. In the most part of the cases tested the analysis
made by two of those programs was enough to detect the web
page language. When the analysis made by two programs
disagreed, one extra program was used to reach a conclusion [
Supposing the website X, inLinks are all the links existing in the Web that point to X, outLinks are the links of the website X that point to sites that are hosted in other domains. The inLinks and outLinks of each website should be stored for posterior analysis. InLinks are used for some search engines in order to build a ranking of the results. Websites that have many inLinks receive 6 http://www.fuzzums.nl/~joost/talenknobbel/index.php 7 http://www.appliedlanguage.com 8 http://www.google.com/uds/samples/language/detect.html
Data about the website author are important for determine the
quality of the presented content. If the author is considered a
specialist on the subject, the web page will have more credibility
then the others, written by people not recognized as such.
However, this kind of information is not easily determined in a
manual way and in case of automatic treatment of data it became
even more complex. There are some techniques that could be
applied to help in analyzing data about the author, as is the case
of h-index [
The models described in the previous sections were developed as prototypes to evaluate the proposed approach. During the tests the topic used for search was "Alzheimer Disease". The first step was focused in the automatic location of data, the considered tasks were:
Study of the existent search engines, as well as analysis of each one of them in order to identify which ones were appropriate to search data about health; Definition of the criteria that will be used for the searches related to the topic "Alzheimer Disease".
The section 6.1 below presents details of the search engines. About the criteria, without a real population of users to develop this case study, was chosen to use common criteria from the topic. A more detailed description about the applied criteria is presented in the subsection 6.2.
The queries have two kinds of results: texts that are appropriated
to general public; technical texts with more detailed information
of the topic. During the search engine analysis, three of them
have been choose. That was necessary for automating the
location of web pages, because the developed tool should be
prepared for a specific pattern of the engines. Among the analyzed
ones, the following were adopted [
Medstory: Search in the Web and sorts the results based on several keywords, indicating which ones occur more frequently. These keywords, about healthcare, belong to a predefined static list, grouped by categories, as: drugs, symptoms, procedures and so on. This list predefined by the website allows refining the results presented to the user. After the search, several complementary categories of data related to the topic are presented to the user. These categories consist of pre-processed information, stored in a specific database.
SearchMedica: Allows setting up the search scope in which the search engine will search: the whole Web or sites defined by the user. Also, the search can be made over a determined health area (e.g. cardiology, geriatrics, dermatology, etc.). The results page has an area in which several related information are presented.
Google: Allows search in the whole Web, so, can find
since information for laymen until news or scientific work.
However, there are no guarantees about the reliability of
the presented data in the resulting websites, unlike what
happens in the Medstory, which search in a recognized
source internally indexed. At the same time, searches that
encompass a bigger number of pages can find more
relevant results for the user. In these cases, algorithms as
PageRank™ [
As described in the introduction of this case study, the search
criteria applied were chosen according to the topic "Alzheimer
Disease", so, they are application dependent. Thus, an analysis
should be done in order to choose good criteria for more
coherent results. This issue refers to business analysis [
These search criteria allows the automatic location of websites by the execution of searches over the used search engines (subsection 6.1), as well as store the 10 first URLs, associated with each search engine. After the location and storing the URLs, the system can go further - data extraction.
This step intends to extract the following attributes: page
language, authorship data, inLinks and outLinks. Each one of those
attributes was described during the previous sections. In order to
develop the proposed solution as a prototype some tools were
used to support retrieve data in the websites, they are: Web
Harvest9 e XPather [
The Web-Harvest provide an API (Application Program Interface) that allows: consult Web servers; recover the HTML page; convert it in XHTML(Extensible Hypertext Markup Language) file and apply some manipulation techniques of XML (Extensible Markup Language) documents, as XPath, XQuery and XSLT (Extensible Stylesheet Language Transformations) to extract the desired information; The XPather allows search for patterns in websites, being used in a semi-automatic manner.
This article described the development of models, techniques and prototypes aiming the automatic retrieval of content presented in web pages with health subject. After the automatic location, standardization and extraction of page‘s data the objective is to deliver normalized data to perform evaluations on the quality of the health associated web page. The process of developing the model for automatic retrieval of data anticipated a series of challenges and steps to overcome and the process was directed to work with these problems. For location were applied well known search engines as they use quite efficient algorithms for Web search. The search for patterns for presenting data in web pages has led to the conclusion that despite the already existence of standards such as Dublin Core, in practice, these patterns are not strictly applied by the vast majority of pages, or is partially applied. Thus, the need of some kind of standardization for the development of these pages is clearly needed. As a future work, it would be necessary to end the process of evaluating the quality of a health web page making the task of quality evaluation itself automatic.
This work was partially supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico - CNPq, Brazil, Edital Universal - MCT/CNPQ - 14/2010. We would also like to thank the group of students of the discipline CMP112, PPGC/UFRGS, that developed the initial work on this subject.