[lkb] generation ranking in LKB

[Stephan Oepen]

any number will activate n-best selective unpacking, so you can choose  
how many realizations you want returned (at most).  as for the  
specific model, it's the ‘combined’ model from the EMNLP paper;  
however, just now i'm not sure the MEM file has been updated for the  
change in ERG rule identifiers.  in principle, that model should be  
retrained, i guess.  we have been a little slack about this in recent  
ERG releases ...

cheers, oe



On 10. juli 2010, at 11.44, Xuchen Yao <xuchen at coli.uni-saarland.de>  
wrote:

> Hi Stephan,
>
> Thanks for the reply! I wasn't taking advantage of this ranking  
> mechanism until your response.
>
> I've seen the big "bnc.blm" file at your prompt. May I further know  
> the actual ranking used is based on language models or MaxEnt? I  
> want to have an impression of how well it should work according to  
> the result of your original paper.
>
> BTW, The number "42" in
>
> (generate-from-mrs
>            (mrs::extract-mrs (first *parse-record*))
>            :nanalyses 42)
>
> was a bit confusing at first but after digging into the code, I saw  
> you wanted to "have some cake and eat some", so I guess I should  
> always use "42" here.;-)
>
> Xuchen
>
> On 07/10/2010 11:15 AM, Stephan Oepen wrote:
>> hi xuchen,
>>
>> i am assuming you are thinking of generation using the ERG here?  the
>> following lines in the `script' file enable realization ranking:
>>
>>   #+(and :tsdb :logon)
>>   (tsdb::read-model (lkb-pathname (parent-directory) "jhpstg.g.mem"))
>>   #+(and :tsdb :logon)
>>   (setf *unpacking-scoring-hook* #'tsdb::mem-score-configuration)
>>
>> the reason this is only supported in the LOGON tree is its dependency
>> on the CMU LM toolkit and the actual language model, which were  
>> deemed
>> too space-consuming to be included in the general LKB SVN.
>>
>> so, assuming you are actually in the LOGON tree, you should be able  
>> to
>> use realization ranking.  it will be turned on by default when  
>> running
>> in [incr tsdb()], i.e. batch generating.  for interactive use, i  
>> think
>> you would have to pass the optional :nanalyses argument.  e.g.
>>
>>   LKB(22): (do-parse-tty "kim gave the big angry dog a bone.")
>>   720
>>   375
>>   -1
>>   30937
>>   10
>>   LKB(23): (generate-from-mrs
>>             (mrs::extract-mrs (first *parse-record*))
>>             :nanalyses 42)
>>   ("Kim gave the angry big dog a bone."
>>    "Kim gave the big angry dog a bone."
>>    "Kim gave a bone to the angry big dog."
>>    "Kim gave a bone to the big angry dog.")
>>   86312
>>   2780
>>   728
>>   3008
>>   800
>>   337
>>   635
>>   LKB(24): (loop for edge in *gen-record* collect (edge-score edge))
>>   (2.20243 1.8947849 1.7799873 1.4181345)
>>
>> too bad the adjective ordering ends up unnatural in this example, but
>> the model appears to be doing something (useful).
>>
>>                                                         best  -  oe
>>
>> ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 
>> +++++++++++
>> +++ Universitetet i Oslo (IFI); Boks 1080 Blindern; 0316 Oslo; (+47) 2284 0125
>> +++    --- oe at ifi.uio.no; stephan at oepen.net; http://www.emmtee.net/oe/ 
>>  ---
>> ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 
>> +++++++++++
>>