Hi Andrew,
Thanks for the note. The work building all the models advertised in the
paper has fallen behind, but we hope to have it all resolved by the end of
the month. Hopefully that will resolve some of the problems you have
pointed out here, too. I have updated the PPDB page with a note.
I should note, though, that just like with general MT output, there is no
guarantee that the output will be grammatical or well formed. There are
also no guarantees about what kind of alternatives you'll see. So while I
think the general paraphrasing results could be better, I am not surprised
to see strange substitutions and lack of agreement.
matt
On Sat, 25 Jun 2016, Andrew Olney wrote:
I sent this to u...@joshua.incubator.apache.org and it bounced.
I'm trying to understand the behavior of Joshua English Paraphrase
(apache-joshua-ppdb-2.0-s-all-v1) on this input:
I like cheddar cheese.
The top 10 paraphrases returned are:
I felt cheddar cheese.
I prefers cheddar cheese.
I considered cheddar cheese.
I lay cheddar cheese.
I feels cheddar cheese.
I appreciates cheddar cheese.
I wants cheddar cheese.
I includes cheddar cheese.
I enjoys cheddar cheese.
I wanted cheddar cheese.
I reaffirms cheddar cheese.
I pleased cheddar cheese.
I like cheddar cheese.
Some observations
1. Only the verb is changed. It seems that cheddar cheese -> cheese without
much loss of information.
2. S/V agreement is not enforced in all cases. This is surprising.
I would like to better understand how to address 1 and 2.
Suggestions appreciated.
Best,
-Andrew
