Q:  Look up "sewer (noun)", and "plane (noun)".  What do you think the homonyms are?  

'sewer':  {Senses 1,3}, {Sense 2}.  The "conduit" and "one who sews" meanings are clearly distinct.  Sense 3 is derived from 1 (a negative evaluation).

'plane': {Sense 1}, {Senses 2-5} 

Q: What training data for the senses of "interest" could we get from Table 4 and Japanese-English parallel corpora?

   English sentences containing "interest" that are aligned with Japanese sentences containing:

       "rishi" are training sentences for "interest":sense 1
       "rioku" are training sentences for "interest":sense 1
       "riken" are training sentences for "interest":sense 2
       "kanshin" are training sentences for "interest":sense 3,4
       "kyomi" are training sentences for "interest":sense 3,4
       "rieki" are training sentences for "interest":sense 5
   

Q: Now consider "I ate Spaghetti with meatballs."  
   Give two syntax trees for this sentence.
   For each of your trees, give the semantic roles for "ate" (you can call them "AGENT", "THEME" and "INSTRUMENT")
   Give "reasonable" selectional restrictions for the INSTRUMENT of ate 

TREE 1:
(S (NP (pronoun I))
   (VP (V ate)
       (NP (NP (Nominal Spaghetti))
           (PP (Prep with)
               (NP (Nominal meatballs))))))

"ate" in this analysis has semantic roles:
AGENT: (NP (pronoun I)) 

THEME: (NP (NP (Nominal Spaghetti))
           (PP (Prep with)
               (NP (Nominal meatballs))))

There is no INSTRUMENT in this analysis.

TREE 2:
(S (NP (pronoun I))
   (VP (V ate)
       (NP (Nominal Spaghetti))
       (PP (Prep with)
           (NP (Nominal meatballs)))))

"ate" in this analysis has semantic roles:
AGENT: (NP (pronoun I)) 

THEME: (NP (Nominal Spaghetti))

INSTRUMENT: (NP (Nominal meatballs))

== One way to define and check for selectional restrictions is the following:

   Suppose you have the annotated data from the CoNLL 2005 shared task (your project)

   Find all sentences for which "eat" is a target word (i.e., for which the semantic roles are given) which has an "INSTRUMENT" semantic role.  Well, our data doesn't have a role called "INSTRUMENT". Suppose the annotators of our data used the "AM-ADV" role for "fork" in sentences such as "I ate my dinner with a fork".

  Extract the constituents that fill the AM-ADV roles in these sentences (call them Cs).

  Look up heads of the Cs in WordNet, and record their "inherited hypernym" paths.

  Look up "meatball" in WordNet, and look at its "inherited hypernym".  You will find that its "inherited hypernym" does not have much in common with those of the Cs.   Thus, it is likely that "meatball" does not meet the selection restriction.

Since "meatball" is a poor match for the selectional restriction, we reject the second syntax/semantic role analysis.