Information extraction from the web using a search engine Citation for published version (apa)
Yüklə 0,9 Mb. Pdf görüntüsü
|
|
p∈H
h(p,t) ≥ n, where h(p,t) is the number of search engine hits for query with pattern p combined with term t. Based on experiments with the two known restaurant chains, we set the threshold to n = 50. Evaluation We selected the 20 best scoring patterns and use them in the ontology population algorithm with the given ontology. The first task is to identify Burger King as a restaurant using the given patterns. If this instance is found, it can be used – in combination with the same patterns – to identify the countries it is located in. Using the 20 learned patterns, 53 terms were accepted as instances of restaurant (Table 4.4) 4 . The reader may recognize a number of restaurant, coffee and fast food chains, among which Burger King. Less expected are the names of geographic locations and names of famous cuisines such as ‘Chinese’ and ‘French’. The last category of false instances found that have not been filtered out, are a number of very common words (e.g. ‘It’ and ‘There’). 4 Experiment conducted in November 2005 using Google. 4.2 Identifying Burger King and its Empire 57 Chinese Bank Outback Steakhouse Denny’s Pizza Hut Kentucky Fried Chicken Subway Taco Bell Continental Holywood Wendy’s Long John Silver’s HOTEL OR This Burger King Japanese West Keg Steakhouse You BP Outback World Brazil San Francisco Leo Victoria New York These Lyons Starbucks FELIX Roy California Pizza Kitchen Marks Cities Emperor Friendly Harvest Friday Tim Hortons Vienna Montana Louis XV Greens Red Lobster Good It There That Mark Dunkin Donuts Italia French Table 4.4. Extracted instances for restaurant using Google, December 2005. In Table 4.5 the 53 extracted terms can be found, together with the acceptance scores ∑ p∈H h(p,t) as found with the Yahoo! API in June 2008. As OR is an special query operator, the hits for HOTEL OR will not reflect the actual number of occurrences for the corresponding phrases. The high scoring terms – except for HOTEL OR – correspond to large chains. Where Keg Steakhouse was accepted as an instance in 2005, no hits were found for any of the four acceptance function queries. The algorithm returned 69 instance-pairs with countries related to Burger King. On the Burger King website 5 a list of the 65 countries can be found in which the hamburger chain operates. Of these 65 countries, we identified 55. This implies that our results have a precision of 55 69 = 0.80 and recall of 55 65 = 0.85. Many of the falsely related countries – mostly in Eastern Europe – are locations where Burger King is said to have plans to expand its ‘empire’. Using post-processing, we can filter out common words (e.g. Good, It) which is likely to improve the results. The geographic locations can be recognized and filtered out using a gazetteer [Cunningham, Maynard, Bontcheva, & Tablan, 2002; Zong, Wu, Sun, Lim, & Goh, 2005]. We hence can conclude that learning patterns using only a small set of known 5 http://www.whopper.com 58 Outback Steakhouse 1748 HOTEL OR 821 Burger King 506 Starbucks 489 Red Lobster 488 Pizza Hut 463 Taco Bell 392 Subway 375 Denny’s 332 Long John Silver’s 302 Chinese 288 Wendy’s 240 This 166 Japanese 149 French 144 Outback 135 California Pizza Kitchen 111 Dunkin Donuts 104 That 103 Roy 90 These 48 Victoria 47 Kentucky Fried Chicken 46 Greens 30 Friendly 29 It 28 FELIX 25 Friday 25 Harvest 24 Louis XV 23 Marks 22 Tim Hortons 22 You 21 West 18 New York 17 Bank 12 Vienna 12 Montana 12 Good 11 Lyons 10 Continental 7 Mark 5 Italia 5 Emperor 4 Cities 3 San Francisco 2 BP 1 Leo 1 Holywood 0 Keg Steakhouse 0 World 0 Brazil 0 There 0 Hotel Or 0 Table 4.5. The 53 instances found for restaurant and their scores for the accep- tance function as found with the Yahoo! API in June 2008. instances leads to good results in this case-study. The learned patterns are quite specific, but recognizable as strings relating the instances of the two classes. The extraction of the instances of the class Restaurant is done using simple rules. A number of irrelevant terms are falsely identified, but additional filtering steps may lead to improvements. 4.3 Identifying Countries In this case-study, we focus on two tasks. First, we automatically identify hyponym patterns and compare the results with the patterns identified by Hearst [1992]. Con- trary to the previous case-study, we will use a relatively large training set of relation 4.3 Identifying Countries 59 instances to identify a set of effective patterns. The learned patterns are used in the second part of this experiment to extract instances using memory-based learning. 4.3.1 Learning Effective Hyponym Patterns We are interested whether the effective text patterns are indeed intuitive formula- tions of the given relation. As a test-case, we compute the most effective patterns for the hyponym relation using a test set with names of all countries. Taking the terms country and countries as hypernyms, we are interested which text fragments connect the names of countries with these words. Much pattern-based information extraction research (e.g. [Caraballo, 1999; Cimiano & Staab, 2004; Etzioni et al., 2005; Snow, Jurafsky, & Ng, 2006; Tjong Kim Sang & Hofmann, 2007]) is based on hyponym patterns manually identified by Hearst in [1992]. We are interested in the overlap of the automatically found hyponym patterns with the commonly used ones. This experiment was set up as follows. We again use the collected list of coun- tries (see Section 4.2). Let I Yüklə 0,9 Mb. Dostları ilə paylaş:
|