Information extraction from text, Week 4

The solutions should be ready for inspection by Thursday 7.3.2002 (midnight).

Remember that always, if you are in doubt what you should do, you can ask Reeta or send a message to our newsgroup!!

1. In this exercise, we study the algorithm of AutoSlog-TS. Assume, our text collection contains the following documents:

   
   text 1; relevant
   
   s: A group of terrorists
   v: attacked
   do: a post
   pp: in Nuevo Progreso.
   
   
   text 2; relevant
   
   s: The National offices
   v: were attacked
   time: today.
   s: Unidentified individuals
   v: detonated
   do: a bomb.
   s: The bomb
   v: destroyed
   do: a car.
   
   
   text 3; not relevant
   
   s: The Armed Forces units
   v: killed
   do: one rebel.
   s: They
   v: destroyed
   do: an underground hideout.
   
   
   text 4; relevant
   
   s: Unidentified individuals
   v: attacked
   do: a high tension tower.
   s: They
   v: destroyed
   do: it.
   
   
   text 5; not relevant
   
   s: The coca growers
   v: protest
   do: the destruction of their fields.
   s: The strike
   v: is supported
   pp: by the Shining Path.
   
   

   Explain the process of AutoSlog-TS using these documents and give the ranking for the extraction patterns that are generated.

   Abbreviations: s subject, v verb, do direct object, pp preposition phrase

   More information: Riloff, E. "Automatically Generating Extraction Patterns from Untagged Text" Proceedings of the Thirteenth National Conference on Artificial Intelligence (AAAI-96) , 1996, pp. 1044-1049.




2. Compare the methods for learning information extraction rules (AutoSlog, Crystal, AutoSlog-TS, Multi-level bootstrapping, ExDisco):

   • What are the prerequisites of the method: What kind of input it needs and how this input should be processed/represented?

   • Describe very briefly the basic idea of the algorithm.

   • What is the output?

   • What the human user has to do?