papers-we-love_papers-we-love/information_retrieval

Information Retrieval

Information retrieval is the activity of obtaining information resources relevant to an information need from a collection of information resources. (Says Wikipedia).

The included documents are

• 📜 Graph of Word and TW-IDF - Francois Rousseau & Michalis Vazirgiannis

  The traditional IR system stores term-specific statistics (typically a term's frequency in each document - which we call TF) in an index. Such a model ignores dependencies between terms and considers a document's terms to occur independently of each other (and is aptly called the bag-of-words model). In this paper the authors use a statistic that uses a graph representation of a document to encode dependencies between terms and replace the TF statistic with a new TW statistic based on the graph constructed and achieve significantly better results that popular existing models. This paper won a honorable mention at CIKM 2013.

• 📜 Pagerank Algorithm - Lawrence Page, Sergey Brin, Rajeev Motwani and Terry Winograd

  This paper introduces the PageRank algorithm, which forms the backbone of the present day google search engine. Pagerank operates by assessing the number of incoming and outgoing hyper links to a given web page and ranks the pages based on the link structure of a page. The authors also implemented PageRank on the backrub system (now called the Google Search Engine) in the [Anatomy of a Large-Scale Hypertextual Web Search Engine] http://infolab.stanford.edu/~backrub/google.html which assigned page ranks to every webpage in the world wide web. Google is currently the most commercially sucessful generic search engine in the world.

• 📜 Okapi System - Stephen E. Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford

  This paper introduces the now famous Okapi information retrieval framework which introduces the BM25 ranking function for ranked retrieval. It is one of the first implementations of the probabilistic retrieval frameworks in literature. BM25 is a bag of words retrieval function. The IDF(Inverse document frequency) term can be interpreted via information theory. If a query q appears in n(q) docs the probability of picking a doc randomly and it containing that term :p(q) = n(q) / D, where D is the number of documents. The information content based on shannon's noisy channel model is = -log(p(q)) = log (D / n(q)). Smoothing by adding a constant to both numberator and demoninator leads to IDF term used in BM25. BM25 has been shown to be one of the best probabilistic weighting schemes. While the paper was in postscript form, the committer has changed the format to pdf as per guidelines of papers we love via ps2pdf.

• 📜 Hits Algorithm - Jon M. Kleinberg

This paper introduces the HITS algorithm, a link analysis algorithm that rates webpages. Unlike the more famous page rank algorithm, the hits algorithm makes a distinction between webpage behavior classifies them as hubs and authorities. A page is authoratitative (in the sense the page has a large number of incoming links) or acts as a hub (a directory of sort, which can be measured by the number of outgoing link). The hits algorithm computes two scores for a page (authority and hub score) where the algorithm iteratively computes the hub score as sum of authority scores of outgoing links and authority scores as sum of hub scores of incoming links until a convergence is attained. These scores can then be used to rank documents. While this algorithm is famous in academia, its not very widely used in the industry (a variant of this algorithm was used by a company called Teoma which was acquired by AskJeeves)