lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "E. van Chastelet" <evanchaste...@gmail.com>
Subject Re: Spell check on a subset of an index ( 'namespace' aware spell checker)
Date Thu, 24 Nov 2011 11:13:23 GMT
Thank you Mike,

I have thought about that solution myself, but the problem with this 
approach is that the terms still need to be modified before building the 
dictionary that is feed to the spell checker.
Also, the similarity scores which are used to determine the spell 
suggestions are affected by the prefixes. So this solution is probably 
not a good idea.


On 11/23/2011 08:04 PM, Michael Sokolov wrote:
> could use simply index every term with a namespace prefix like:
>
> Q::term
>
> where Q is the namespace and term the term?
>
> Then when you do spell corrections, submit each candidate term with 
> the namespace prefix prepended
>
> -Mike
>
> On 11/23/2011 9:28 AM, E. van Chastelet wrote:
>> I currently have an idea to get it done, but it's not a nice solution.
>>
>> If we have an index Q with all documents for all namespaces, we first 
>> extract the list of all terms that appear for the field namespace in 
>> Q (this field indicates the namespace of the document).
>>
>> Then, for each namespace n in the terms list:
>>  - Get all docs from Q that match +namespace:n
>>  - Construct a temporary index from these docs
>>  - Use this temporary index to construct the dictionary, which the 
>> SpellChecker can use as input.
>>  - Call indexDictionary on SpellChecker to create spellcheck index 
>> for current namespace.
>>  - Delete temporary index
>>
>> We now have separate spell check indexes for each namespace.
>>
>> Any suggestions for a cleaner solution?
>>
>> Regards,
>> Elmer van Chastelet
>>
>>
>>
>> On 11/10/2011 01:16 PM, E. van Chastelet wrote:
>>> Hi all,
>>>
>>> In our project we like to have the ability to get search results 
>>> scoped to one 'namespace' (as we call it). This can easily be 
>>> achieved by using a filter or just an additional must-clause.
>>> For the spellchecker (and our autocompletion, which is a modified 
>>> spellchecker), the story seems different. The spell checker index is 
>>> created using a LuceneDictionary, which has a IndexReader as source. 
>>> We would like to get (spellcheck/autocomplete) suggestions that are 
>>> scoped to one namespace (i.e. field 'namespace' should have a 
>>> particular value).
>>> With a single source index containing docs for all namespaces, it 
>>> seems not possible to create a spellcheck index for each namespace 
>>> the ordinary way.
>>> Q1: Is there a way to construct a LuceneDictionary from a subset of 
>>> a single source index (all terms where namespace = %value%) ?
>>>
>>> Another, maybe better solution is to customize the spellchecker by 
>>> adding an additional namespace field to the spellchecker index. At 
>>> query-time, an additional must-clause is added, scoping the 
>>> suggestions to one (or more) namespace(s). The advantage of this is 
>>> to have a singleton spellchecker (or at least the index reader) for 
>>> all namespaces. This also means less open files by our application 
>>> (imagine if there are over 1000 namespaces).
>>> Q2: Will there be a significant penalty (say more than 50% slower) 
>>> for the additional must-clause at query time?
>>>
>>> Q3: Or can you think of a better solution for this problem? :)
>>>
>>> How we currently do it: we currently use Lucene 3.1 with Hibernate 
>>> Search and we actually already have auto completion and spell 
>>> checking scoped to one namespace. This is currently achieved by 
>>> using index sharding, so each namespace has its own index and 
>>> reader, and another for spell check and auto completion. 
>>> Unfortunately there are some downsides to this:
>>> - Our faceting engine has no good support for multiple indexes, so 
>>> faceting only works on a single namespace
>>> - Needs administration for mapping namespace identifier (String) to 
>>> index number (integer)
>>> - The number of shards (and thus name spaces) is currently 
>>> hardcoded. At this moment it is set to 100, and this means Hibernate 
>>> Search opens up 100 index readers/writers, while only n<100 are in 
>>> use. and therfore:
>>> - Much open file descriptors
>>> - Hard limit on number of namespaces
>>>
>>> Therefore it seems better to switch back to having a single index 
>>> for all namespaces.
>>>
>>> Thanks!
>>>
>>> Regards,
>>> Elmer van Chastelet
>>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
>> For additional commands, e-mail: java-user-help@lucene.apache.org
>>
>


---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message