Dear All,
I have a question concerning the print-feature-forest command of XLE and would appreciate any help from someone who has some experience using this command.
I am working on training a probabilistic variant of the Polish LFG grammar using XLE's functionalities. When preparing the feature forests for cometc using the print-feature-forest command, I came across a problem: it seems that print-feature-forest sometimes gets the feature counts wrong. I ran the print-feature-forest command on the obj-ex.pl file attached to this e-mail, with a features file containing only one line, "fs_attrs OBJ" (so the only feature counted here is the number of OBJ's in the f-structure). obj-ex.pl is a packed parse representation produced by the Polish LFG grammar, with some attributes non-relevant to the problem, such as morphosyntactic ones, removed.
The sentence translates as "Look ahead of you and read in the beard". It contains 2 verbs and 2 adjunct prepositional phrases and has 3 possible parses; in two parses, each verb has one of the PP's attached to it; in one parse, the second PP is shared between the verbs. Since a PP which is an adjunct is represented as an f-structure headed by the preposition whose OBJ is the noun's f-structure, it seems that the first two parses contain 2 OBJ's and the third one contains 3 of them since the shared PP appears twice in the f-structure. However, the output feature-forest is:
(7 1:2 {1 (1 {2 (3 ) (4 1:2 ) } ) (2 ) } )
which corresponds to two parses with 2 OBJ's and one with 4 of them. Knowing from my previous experience that subsumption can be tricky in XLE, I replaced the two subsume statements in obj-ex.pl with equalities (in the case of this sentence, there actually is an equality), and the feature forest produced for the modified f-structure seems correct:
(7 1:1 {1 (1 {2 (3 1:1 ) (4 1:2 ) } ) (2 1:1 ) } )
Perhaps someone knows why it works like this? Am I wrong in expecting what the feature forest should look like? Or maybe it is an known issue with print-feature-forest? I would be very grateful for any advice, since data preparation for my experiments relies heavily on the print-feature-forestcommand .
Many thanks in advance, Katarzyna Krasnowska