This paper aims to focus on the design of algorithms and techniques for an effective set expansion. A tool that finds and extracts candidate sets of tuples from the World…
This paper aims to focus on the design of algorithms and techniques for an effective set expansion. A tool that finds and extracts candidate sets of tuples from the World Wide Web was designed and implemented. For instance, when a given user provides <Indonesia, Jakarta, Indonesian Rupiah>, <China, Beijing, Yuan Renminbi>, <Canada, Ottawa, Canadian Dollar> as seeds, our system returns tuples composed of countries with their corresponding capital cities and currency names constructed from content extracted from Web pages retrieved.
The seeds are used to query a search engine and to retrieve relevant Web pages. The seeds are also used to infer wrappers from the retrieved pages. The wrappers, in turn, are used to extract candidates. The Web pages, wrappers, seeds and candidates, as well as their relationships, are vertices and edges of a heterogeneous graph. Several options for ranking candidates from PageRank to truth finding algorithms were evaluated and compared. Remarkably, all vertices are ranked, thus providing an integrated approach to not only answer direct set expansion questions but also find the most relevant pages to expand a given set of seeds.
The experimental results show that leveraging the truth finding algorithm can indeed improve the level of confidence in the extracted candidates and the sources.
Current approaches on set expansion mostly support sets of atomic data expansion. This idea can be extended to the sets of tuples and extract relation instances from the Web given a handful set of tuple seeds. A truth finding algorithm is also incorporated into the approach and it is shown that it can improve the confidence level in the ranking of both candidates and sources in set of tuples expansion.
Querying search engines with the keyword “jaguars” returns results as diverse as web sites about cars, computer games, attack planes, American football, and animals. More…
Querying search engines with the keyword “jaguars” returns results as diverse as web sites about cars, computer games, attack planes, American football, and animals. More and more search engines offer options to organize query results by categories or, given a document, to return a list of links to topically related documents. While information retrieval traditionally defines similarity of documents in terms of contents, it seems natural to expect that the very structure of the Web carries important information about the topical similarity of documents. Here we study the role of a matrix constructed from weighted co‐citations (documents referenced by the same document), weighted couplings (documents referencing the same document), incoming, and outgoing links for the clustering of documents on the Web. We present and discuss three methods of clustering based on this matrix construction using three clustering algorithms, K‐means, Markov and Maximum Spanning Tree, respectively. Our main contribution is a clustering technique based on the Maximum Spanning Tree technique and an evaluation of its effectiveness comparatively to the two most robust alternatives: K‐means and Markov clustering.