cartisan / knoex Goto Github PK

In compare_easies for a term t we find the most apprpriate synset s_x by computing the similarity scores for all of it's possible synsets s_1, ..., s_n and taking the synset with the biggest similarity.

The similarity score for asynset s_i is computed by summing up all the distances of this synset and the synsets from the flattened rest-list (containing the synsets of the othe terms).
Since the flattened rest-list contains for each term ALL of it's possible synsets the computed similarity picks out words that are closeest to all possible meanings of a term instead of finding the combination of closest synsets.

Minimalist example:
Two terms t (snake) and u (mouse).
Four synsets: s_t1 (animal), s_t2 (bad person) and s_u1 (animal), s_u2 (shy person)
We compute:
P(s_t1) = d(s_t1, s_u1) + d(s_t1, s_u2) = 1,1
P(s_t2) = d(s_t2, s_u1) + d(s_t2, s_u2) = 1,5

Thus apparently bad person is more similar to "animal and shy person" than animal is to "animal and shy person". But this doesn't help us disambiguate things. Instead we probably want to have similarity scores of synset-tuples:

P(s_t1, s_u1) = d(s_t1, s_u1) = 0,8
P(s_t1, s_u2) = d(s_t1, s_u2) = 0,4
P(s_t2, s_u1) = d(s_t2, s_u1) = 0,3
P(s_t2, s_u2) = d(s_t2, s_u2) = 0,9

Which would tell us that the most probable combination of synsets is the persons interpretation.

(@kstandvoss: Just random musing while incorporating our project into the souma study report. Please tell me what you think! :) )