lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Israel Tsadok" <itsa...@gmail.com>
Subject Re: Any way to ignore repeated terms in TF calculation?
Date Thu, 15 Jan 2009 09:37:51 GMT
Hi Umesh,

>   I am trying to put the problem more concisely.
> 1. Fields where term frequency is very very relevant. E.g.
>   Body:
>   Example:
>        if TF of badger in Body of doc 1  >   TF of badger in Body of doc 2
>   doc 1 scores higher.
>
> 2. Fields where term frequency is irrevalent
>   Page_Title:
>   Example:
>        TF of badger in PageTitle doesn't affect the score.
>

This is not quite what I was talking about. I was talking about documents
with a single field. I want the text "Badgers are mammals. Badgers are cute"
to score higher than the text "Badger Badger" for the term query
"text:badger".
Ideally, what I want is to add another factor to the scoring at index time,
a "sparsity factor" which should cancel out the term frequency as the
average distance between terms nears 1.
i.e. if the score formula is:

score(q,d) = coord(q,d) x queryNorm(q) x sigma(t in q) of ( tf(t in d) x
idf(t)^2 x t.getBoost() x norm(t,d) )

I want to make it:

score(q,d) = coord(q,d) x queryNorm(q) x sigma(t in q) of ( tf(t in d) x
idf(t)^2 x t.getBoost() x norm(t,d) x sparsity(t in d) )
where
sparsity(t in d) = 1 / (1 + ( tf(t in d) - 1) / (1 + e ^ (avg_d(t in d) -
7)))
where
avg_d(t in d) = average distance between terms t in document d

Sorry about the weird math, I just mean (as I said above) that the sparsity
factor should cancel out the tf completely if avg_d<=1 and become 1 as avg_d
gets larger.

I looked at Similarity.computeNorm(), which may make it possible for me to
add this inside the normalization value, but I'm not sure if that's really
possible, plus the method is not available yet in 2.4.

Having unloaded all that off my chest, I have to say that I really like your
proposal, and it might solve 90% of my problems without resorting to my
overreaching redesign of Lucene core...


If that is the case:
> then one solution is
> 1. Build the query programmatically.
> 2. Form Normal Queries on FieldType 1 ( e.g. Body)
> 3. Form ConstantScore variation of queries on FieldType 2 (e.g. Page_Title,
> ConstantScoreTermQuery)
>
> There is no need to change anything at index time.
>

OK, so I really like this. The only problem is that it's not going to be
easy to build the query programmatically, since currently I'm using
QueryParser with a little help from MultiFieldQueryParser and
PerFieldAnalyzerWrapper.

I think that the best course of action would be to subclass
MultiFieldQueryParser so that for Body fields it will behave normally, but
for Page_Title fields it will emit a ConstantScoreQuery wrapping the
original field query.

Can you think of an easier way to do this?
Thanks,
Israel

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message