http:^^www.cs.washington.edu^research^projects^softbots^papers^metacrawler^www4^html^overview.html

来自「This data set contains WWW-pages collect」· HTML 代码 · 共 1,261 行 · 第 1/4 页

<tr>
<th align=left>WebCrawler  </td><td align=center>70.2          </td><td align=center> 17.6 / 25          </td>
</tr>
<tr>
<th align=left>Galaxy      </td><td align=center>56.9          </td><td align=center> 28.4 / 50          </td>
</tr>
<tr>
<th align=left>InfoSeek    </td><td align=center>43.5          </td><td align=center>  4.4 / 10          </td>
</tr>
<tr>
<th align=left>Yahoo       </td><td align=center>11.1          </td><td align=center> 11.1 / 100         </td>
</tr>
</table>

<blockquote>
The first column shows the percentage of the maximum hits allowed that
each service returned. Each percentage was calculated by dividing the
average hits 
returned by the maximum allowed for that service, as shown in the
second column. This percentage is a measure of how many hits a
service will provide given a pre-set maximum. The MetaCrawler used
different maximum values for services, as some had internal maximum
values, and others would either accept only certain maximums or none
at all. 
</blockquote>
<hr>
<hr>
<BR><STRONG>Table 6:</STRONG> <A NAME=ccuniquehits></A>Unique References by Service<BR>
<A NAME=547></A>

<table>
<tr>
<th align=left></th><th align=center>% of Unique Hits Returned          </th>
</tr>
<tr>
<th align=left>Galaxy      </th><td align=center> 99.6            </td>
</tr>
<tr>
<th align=left>Yahoo       </th><td align=center> 92.8            </td>
</tr>
<tr>
<th align=left>Lycos       </th><td align=center> 90.6            </td>
</tr>
<tr>
<th align=left>WebCrawler  </th><td align=center> 90.3            </td>
</tr>
<tr>
<th align=left>Open Text   </th><td align=center> 88.8            </td>
</tr>
<tr>
<th align=left>InfoSeek    </th><td align=center> 79.5            </td>
</tr>
</table>

<blockquote>
This table shows the percentage of references each service
returned that were unique to that service. 
</blockquote>
<hr><p>
<H4><A NAME=SECTION00032200000000000000> Relevance</A></H4>
To measure relevance, two metrics are used. The first is which
service is returning the most references that people follow. This is
shown by Table <!WA37><!WA37><!WA37><!WA37><A HREF="#followedtable">2</A>. The second metric is what percent
of references returned by each service are people following. 
Table <!WA38><!WA38><!WA38><!WA38><A HREF="#ccrelevance">7</A> summarizes these
calculations. It shows that nearly 6% of all references returned by
InfoSeek were followed. Lycos, Open Text, and
WebCrawler follow, each having about 2.5% of their hits
followed.
<P>
This data has two caveats. The first is that the relevant
information for people may be the list of references itself. For example,
people who wish to see how many links there are to their home page may
search on their own name just to calculate this number. The second
caveat is that these numbers may be skewed by the number of hits
returned by each service. Thus, while InfoSeek has nearly 6% of its
results followed, only a total of 13,045 references returned by
InfoSeek were followed, compared with the 24913 references followed
that were contributed by Lycos, which on average had only 2.5% of its
19 hits examined.
<p>
<hr>
<BR><STRONG>Table 7:</STRONG> <A NAME=ccrelevance></A>Relevance of Returned Hits by Service<BR>
<A NAME=548></A>

<table>

<tr>
<th align=left></th><th align=center> % of Hits Returned that are Followed </th><th align=center> Total Hits Followed </th>
</tr>
<tr>
<th align=left>InfoSeek    </th><td align=center> 5.89                </td><td align=center> 13,045      </td>
</tr>
<tr>
<th align=left>Lycos       </th><td align=center> 2.56                </td><td align=center> 24,913      </td>
</tr>
<tr>
<th align=left>Open Text    </th><td align=center> 2.51                </td><td align=center>  4,025      </td>
</tr>
<tr>
<th align=left>WebCrawler  </th><td align=center> 2.42                </td><td align=center> 21,631      </td>
</tr>
<tr>
<th align=left>Yahoo       </th><td align=center> 1.33                </td><td align=center>  7,503      </td>
</tr>
<tr>
<th align=left>Galaxy      </th><td align=center> 0.83                </td><td align=center> 12,022      </td>
</tr>
</table>

<blockquote>
This table shows the percentage of followed hits for each
service. References returned by multiple services
are counted multiple times. Column 2 shows the actual number of
references followed for each service. These numbers are out of
50,878 queries, except Open Text which is out of 19,951 queries. 
</blockquote>
<hr><p>
<H4><A NAME=SECTION00032300000000000000> Performance</A></H4>
Finally, we measure each service's response time. Table
<!WA39><!WA39><!WA39><!WA39><A HREF="#ccperformance">8</A> summarizes our findings. It is not surprising,
although disappointing, to find that average times vary from just
under 10 seconds to just under 20. The percent of time the services timed out
is under 5% for all services except Open Text, which is the newest
and presumably still going through some growing pains. One explanation
for the length of times taken by these services is that the majority
of requests are during peak hours. Thus, results are naturally
skewed towards the times when the services are most loaded. Times
during non-peak hours are much lower.
<p>
<hr>
<BR><STRONG>Table 8:</STRONG> <A NAME=ccperformance></A>Performance of Services<BR>
<P><A NAME=549></A>

<table>
<tr>
<th align=left>            </th><th align=center>  Ave. Time (sec)  </th><th align=center> % Timed Out </th>
</tr>
<tr>
<th align=left>WebCrawler  </th><td align=center>   9.64            </td><td align=center>   2.30       </td>
</tr>
<tr>
<th align=left>InfoSeek    </th><td align=center>  12.30            </td><td align=center>   3.01       </td>
</tr>
<tr>
<th align=left>Open Text    </th><td align=center>  16.26            </td><td align=center>  14.13       </td>
</tr>
<tr>
<th align=left>Yahoo       </th><td align=center>  18.32            </td><td align=center>   2.28       </td>
</tr>
<tr>
<th align=left>Lycos       </th><td align=center>  18.99            </td><td align=center>   4.87       </td>
</tr>
<tr>
<th align=left>Galaxy      </th><td align=center>  19.52            </td><td align=center>   3.10       </td>
</tr>
</table>

<blockquote>
This table shows the average time in seconds taken by each service to fulfill a
query. The second column gives the percent of time that the service
would time out, or fail to return any hits under one minute. 
</blockquote>
<hr><p>

<A NAME=relatedwork></A>
<H2><A NAME=SECTION00040000000000000000> Related Work</A></H2>
<P>
Unifying several databases under one interface is far from novel. Many
companies, such as PLS[<!WA40><!WA40><!WA40><!WA40><A HREF="#wwwpls">21</A>],
Lexis-Nexis[<!WA41><!WA41><!WA41><!WA41><A HREF="#wwwlexisnexis">14</A>], and Verity[<!WA42><!WA42><!WA42><!WA42><A HREF="#wwwverity">30</A>]
have invested several 
years and substantial capital creating systems that can handle and
integrate heterogeneous databases. Likewise, with the emergence of
many Internet search services, there have been many different efforts
to create single interfaces to the sundry databases. Perhaps the most
widely distributed is CUSI[<!WA43><!WA43><!WA43><!WA43><A HREF="#wwwcusi">18</A>], the Configurable Unified
Search Index, which is a large form which allows users to select one
service at a time and query that service. There are also several other
services much like CUSI, such as the All-in-One Search
Page[<!WA44><!WA44><!WA44><!WA44><A HREF="#wwwallinone">2</A>], or 
W3 Search Engine list[<!WA45><!WA45><!WA45><!WA45><A HREF="#wwww3enginelist">19</A>]. Unfortunately, while the
user has many services on these lists to choose from, there is no
parallelism or 
collation. The user must submit queries to each service individually,
although this task is made easier by having form interfaces to the
various services on one page.
<P>
The Harvest system[<!WA46><!WA46><!WA46><!WA46><A HREF="#harvesttr">6</A>] has many similarities to the
MetaCrawler; however, rather than using existing databases as they are
and post-processing the information returned, Harvest uses ``Gatherers''
to index information and ``Brokers'' to provide different interfaces
to extract this information. However, while Harvest may have many
different interfaces to many different internal services, it is still a
search service like Lycos and WebCrawler, instead of a meta-service like
MetaCrawler.
<P>
There are also other parallel Web search services. Sun Microsystems
supports a very primitive service[<!WA47><!WA47><!WA47><!WA47><A HREF="#wwwsunmultithreaded">27</A>], and IBM
has recently announced infoMarket[<!WA48><!WA48><!WA48><!WA48><A HREF="#wwwinfoMarket">11</A>] which, rather
than integrating similar services with different coverage,
integrates quite different services, such as
DejaNews[<!WA49><!WA49><!WA49><!WA49><A HREF="#wwwdejanews">26</A>], a USENET news search service, 
McKinley[<!WA50><!WA50><!WA50><!WA50><A HREF="#wwwmckinley">29</A>], a clone of Yahoo with some editorial
ratings on various pages, in addition to Open Text and Yahoo.
<P>
The closest work to the MetaCrawler is SavvySearch[<!WA51><!WA51><!WA51><!WA51><A HREF="#wwwsavvy">3</A>], an
independently created multi-threaded search service released in May
1995. SavvySearch has a larger repertoire of search services, although
some are not WWW resource services, such as Roget's
Thesaurus. SavvySearch's main focus is categorizing users' queries,
and sending them to the most appropriate subset of its known
services[<!WA52><!WA52><!WA52><!WA52><A HREF="#dreilingersavvy">4</A>].  

<P>
Like the MetaCrawler, the Internet Softbot[<!WA53><!WA53><!WA53><!WA53><A HREF="#etzioniuicacm">8</A>] is a
meta-service that leverages existing services and collates their results.  
The Softbot enables a human user to state
what he or she wants accomplished.  The Softbot attempts to
disambiguate the request and to dynamically determine how and where
to satisfy it, utilizing a wide range of Internet services.  Unlike
the MetaCrawler, which 
focuses on indices and keyword queries, the Softbot accesses
structured services such as Netfind and databases such as Inspec.  The
Softbot explicitly represents the semantics of the 
services, enabling it to chain together multiple services in response
to a user request.  The Softbot utilizes automatic planning technology
to dynamically generate the appropriate sequence of service accesses.
While the MetaCrawler and the Softbot rely on radically different
technologies, the vision driving both systems is the same.  Both
seek to provide an expressive and integrated interface to Internet
services.
<P>
<A NAME=futurework></A>
<H2><A NAME=SECTION00050000000000000000> Future Work</A></H2>
<P>
We are investigating how the MetaCrawler will scale to use new
services. Of particular importance is how to collate results returned
from different types of Internet indices, such as USENET news and Web
pages. Also important is determining useful methods for
interacting with specialized databases, such as the Internet Movie
Database[<!WA54><!WA54><!WA54><!WA54><A HREF="#wwwmoviedatabase">28</A>]. If the information requested is
obviously located on 
some special purpose databases, than it does not make sense to query
each and every service. We are
investigating methods that will enable the MetaCrawler to determine
which services will return relevant data based solely on the query text
and other data provided by the user.
<P>
<A NAME=futuredesign></A>
<H3><A NAME=SECTION00051000000000000000> Future Design</A></H3>
<P>
The existing MetaCrawler prototype can cause a substantial network load when
it attempts to verify a large number of pages. While one query
by itself is no problem, multiple queries occurring at the same time
can cause the system and network to bog down. However, with the
emergence of universally portable Internet-friendly languages, 
such as Java[<!WA55><!WA55><!WA55><!WA55><A HREF="#javawhitepaper">10</A>] or Magic
Cap[<!WA56><!WA56><!WA56><!WA56><A HREF="#genmagicguidetocodewarrior">25</A>], load problems can be lessened
by having users' machines take on the workload needed to perform their
individual query, as discussed in Section
<!WA57><!WA57><!WA57><!WA57><A HREF="#clientserverdesign">2.2</A>. The JavaCrawler, a prototype next
generation 
MetaCrawler written in Java, supports most of the  
features already present in the MetaCrawler. However, instead of users
running queries on one central service, each user has a local copy of
the JavaCrawler and uses that copy to directly send queries to
services. The load caused by verification will be taken by the user's
machine, rather than the central server. This has
the added benefit of inserting downloaded pages into the local cache, making
retrieval of those pages nearly instantaneous. The JavaCrawler is loaded
automatically from the MetaCrawler home page when visited with a
Java-compatible browser.
<P>
<H3><A NAME=SECTION00052000000000000000> Impact on Search Service Providers</A></H3>
<P>
We anticipate that a wide range of
meta-services like the MetaCrawler will emerge over the next few
years. However, it is 
uncertain what the relationship between these meta-services and search
service providers will be. We envision that this relationship will hinge on
what form the ``information economy'' used by service providers takes.
We discuss two different models.
<P>
<H4><A NAME=SECTION00052100000000000000> Charge-per-Access</A></H4>
<P>
In the charge-per-access model, service providers benefit from
any access to their database. 
InfoSeek has already taken this model with their commercial
service. InfoSeek is financially rewarded regardless of who or what
sends a query to their commercial database. Many other databases, both
on and off the Web, also use this model.
<P>
The MetaCrawler fits in well with this model.
Since service providers benefit from any access, the added
exposure generated by the MetaCrawler is to their advantage. Further,
this model creates an implicit sanity check on the claims this paper
makes on the use of its features. In order for the MetaCrawler, or any
meta-service, to survive in such an economy, it must charge more per
transaction than the underlying services, as the MetaCrawler will in
turn have to pay each service for its information. Thus, users must
be willing to pay the premium for the service. By voting with their
pocketbook, they can determine if those features are truly desirable.
<P>
<H4><A NAME=SECTION00052200000000000000> Advertising</A></H4>
<P>
In the advertising model, service providers benefit from sponsors who
in turn gain benefit from exposure provided by the service.
Nearly all major search services that do not charge users directly have
adopted this model, as have many other unrelated services which are
heavily accessed.
<P>
Under this model, the providers' relationship with the
MetaCrawler can become problematic as the MetaCrawler filters away
superfluous information such as advertisements. One promising method
to ensure profitable co-existence