Web Search, Télécom ParisTech

Information Extraction

Fabian Suchanek, Pierre Senellart

26 June 2012

 

The purpose of this lab session is to extract structured information from a natural language text corpus.

Handing out the Assignment

The code (in a ZIP archive), as well as any other response to the assignment, must be sent by email at the end of the lab session (no extension) to pierre.senellart@telecom-paristech.fr with [INF347-IE] in the subject of the e-mail.

Data Set

Our corpus will be the Simple English Wikipedia (http://simple.wikipedia.org/), a simpler and smaller encyclopedia than the regular English Wikipedia. The whole content of the encyclopedia can be downloaded from https://dumps.wikimedia.org/. For these labs, we provide the data in the following forms:

1. Information Extraction

For all of the following, go for precision rather than recall.
  1. We first need a class Page, which has two String fields: the page title and the page content. Write a class Parser, which has a constructor that takes as argument the filename of the corpus. The class should have a method next():Page, which delivers the next page in the corpus. Remember to close the file at the end. If you want, the class Parser can implement the interfaces Closeable and Iterator<Page>. Implement this class for the preprocessed text-only corpus. Create a small sample corpus of two or three articles and test your class.
  2. Create a class Triple, which has three String fields: subject, predicate and object. Remember to implement the standard methods toString(), hashCode(), equals() and potentially compareTo(). Create an abstract class Extractor, which has a method extract(Page):Collection<Triple>. Create a simple extractor, NameExtractor, which produces triples of the form <PageTitle, hasName, "PageTitle">. Create a class InformationExtractor, which has only one method: run. This method takes as input a corpus file, a target file and a list of extractors. It iterates over all pages in the corpus, calls all extractors and writes the triples, TAB-separated, into the target file. Test this method with the NameExtractor on the small sample corpus.
  3. Write an extractor DateExtractor that uses a regular expression to find the first date mentioned in the article. Let it return a triple of the form <PageTitle, hasDate, Date>. Try the extractor with the pages of Elvis Presley and Alan Turing. If you are adventurous, try normalizing the dates you extract to the form [-]YYYY-MM-DD. Regular expressions work as follows in Java:
    Pattern pattern=Pattern.compile(EXPRESSION);
    Matcher matcher=pattern.matcher(STRING);
    while(matcher.find()) {
      // matcher.group(N) holds the N-th group of the match
      // matcher.group() holds the entire match
    }
    Use named regular expressions.

    For this extractor, as well as for all following extractors, measure the precision on 20 output pairs manually.

  4. Write an extractor TypeExtractor that extracts the type of the article entity ("Leicester is a city"). Manually exclude terms that are too abstract to be an ontological type ("member of...", "way of..."). If you are adventurous, improve your type extractor by running it on the POS-tagged corpus. For this purpose, extend the class Page and write a new TypeExtractor.
  5. Write an extractor LocationExtractor that extracts the location of a place ("Hollywood is a district in Los Angeles"). Alternatively or additionally: Write a TypeAndLocationExtractor, which first calls the TypeExtractor, checks if the article entity is a city, district, etc., and, if so, extracts the location.

2. To go further

  1. Extend your information retrieval system. Whenever a document contains an article entity X, and if you know that <X, locatedIn, Y>, add an artificial word to the index "placeInY". See whether you can find "Abdou Diouf" if you search for "president placeInWestAfrica".
  2. Write extractors that work on the full version of the simple Wikipedia. In this version, you can use the infoboxes to extract birth dates (e.g., for Elvis Presley), or population numbers (e.g., for Paris).