Crescenzi et al, 2001

Citation

V. Crescenzi, G. Mecca, and P. Merialdo. ROAD RUNNER: Towards automatic data extraction from large web sites. In Proc. of the 2001 Intl. Conf. on Very Large Data Bases, pages 109–118, 2001.

Online version

[[1]]

Summary

This paper introduces a novel technique for automatic wrapper generation by comparing HTML pages and building a wrapper based on the similarity between web pages. This technique can be applied on websites that contain large amount of data (i.e. data-intensive). They also have assumed that the webpages of the given website have fairly similar structure. The main advantages of this technique are:

- This technique does not require any interaction with user during the process of wrapper generation. This extends the applicability of their technique to automatically learn wrappers for input website without getting any supervision from human.

- The technique doesn't have any prior knowledge about the structure of the input web pages.

Given two web pages as the input of the system, this technique compares the content of these two web pages and generate a wrapper by comparing similarities and dissimilarities of these two pages. They have developed a matching technique to extract wrapper from the input webpages. The matching algorithm works on two objects at the same time: (1) a list of tokens and (2) a wrapper. It initially considers one of the input webpages as a wrapper and then iteratively refines the wrapper by processing new web pages. When it processes the new web pages it may finds a mismatch between the structure of the web page and the current wrapper. In these cases it tries to generalize the wrapper to solve the mismatch.

Given two webpages the algorithm first convert the HTML codes to XHTML. It then selects on of the webpages randomly, for example page 1, and build a wrapper based on it. The wrapper is then refined based on the content of the second webpage by solving mismatches between wrapper and content of the second webpage. There are two kinds of mismatches:

- String mismatches: mismatches between different strings that occur at the same position.

- Tag mismatches: mismatches between different tags in the wrapper and the new webpage.

In most of the cases mismatches can be solved by the followings ways:

- String mismatches may refer to different values of a database.

- Tag mismatches (discovering optionals): If a mismatch happens for two tags then it means that either we have an optional field or an iterator. In the case of optional we have either in the new webpage or in the wrapper a piece of HTML code that is not present on the other side. In this case we first determine the boundary of the optional field and then generalize the wrapper to cover this optional field.

- Another reason for tag mismatches is different cardinalities in the the number of items in the webpage (e.g. number of book titles). To solve the mismatches, we need to identify all these repeated patterns. They have described the detail of their technique to handle this case in their paper.

We mentioned in the above an example of solving mismatches. In the paper they have discussed more complex examples in details.

This technique is tested on several well known data-intensive web sites. For each web site they have downloaded 10-20 pages with similar structure. These web pages are given to the program to generate a wrapper. The results show that this technique has been able to extract dataset of 8 websites (among 10 different tested websites).