Mailing-List: contact lucene-dev-help@jakarta.apache.org; run by ezmlm
Precedence: bulk
Reply-To: "Lucene Developers List" <lucene-dev@jakarta.apache.org>
Received-SPF: pass (hermes.apache.org: local policy)
From: Paul Elschot <paul.elschot@xs4all.nl>
To: lucene-dev@jakarta.apache.org
Subject: Re: idf^2
Date: Sun, 17 Oct 2004 20:50:24 +0200
User-Agent: KMail/1.5.4
References: 
 <E3381E0825F1954D953A4E347017BB1C0258339B@reh001-1.REX001.ExchangeByRegister.com>
In-Reply-To: 
 <E3381E0825F1954D953A4E347017BB1C0258339B@reh001-1.REX001.ExchangeByRegister.com>
MIME-Version: 1.0
Content-Type: text/plain;
  charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
Content-Disposition: inline
Message-Id: <200410172050.24372.paul.elschot@xs4all.nl>

Chuck,

On Sunday 17 October 2004 18:30, Chuck Williams wrote:
> Paul,
>
> I'm glad to hear you are doing this -- it sounds like a great project.
> After sending the message below, I was thinking about the possibility of
> a flexible scoring architecture that would allow application developers
> to easily customize scoring to their needs.  One simple thing I'm
> missing in the current Similarity mechanism is a parameter that
> identifies the field for which the various factors are being computed.
> For various reasons, I'd like to use different computations for
> different fields in my application.

For that you can use:

new TermQuery(new Term(fieldName, termText)) {
  public Similarity getSimilarity(Searcher) {return yourSimilarity;}
}

>
> Probabilistic scoring models generally use some kind of training or
> learning mechanism to induce optimal coefficients and additive factors
> for combining a variety of score components (including tf, idf, field
> length and others -- e.g., OKAPI uses a term related to but different
> from idf:  the ratio of number of docs not containing a term to the
> number of docs that do contain the term).  It would be great to have a
> scoring mechanism that defined an extensible library of standard score
> components and provided the flexibility to combine them arbitrarily, and

Lucene comes a long way in allowing arbitrary combinations of Similarity
implementations.

> to use training to induce an optimal combining function.  The
> probabilistic models have a couple significant advantages:
>   1.  An application developer can optimize relevance by using
> training/learning techniques.

If one has a relevant training set with a good set of real world queries...

>   2.  Scores are true probabilities that reflect the likelihood a result
> is relevant.  This allows absolute comparability across searches and
> various thresholding, segregating and other UI devices to communicate
> the quality of results to the user.

Lucene is already quite close to true probabilities: apart from the various
weights, a TermQuery scores by the square root of the term frequence in
document divided by the square root of the document field length, ie.
the score is proportional to the square root of the term  density.

> I would love to see a flexible scoring mechanism to go along with a more

Lucene's Weight and Scorer also allow to override getSimilarity(),
adding to its flexibility.

> structured query mechanism in Lucene, and would be interested in helping

The current query parser already does quite a bit, however it does not
allow truncation and distance to be combined (unless this was added
recently.)

> to create this.  Within this mechanism, the default scoring formula
> should be one of the best empirically tested and tuned variations, e.g.,
> Salton's final SMART formula or Robertson's OKAPI.

The difference between Lucene's scoring and Okapi is actually quite little.
Lucene as a coordination factor, which Okapi doesn't have.
Okapi treats the document length somewhat differently, but in the end
the net effect is mostly a saturation on term density, like the
square root of Lucene. (Which looks much like an inverse power norm, btw.)
Lucene mostly lacks the possibility of synonyms, ie. adding the densities
before applying the saturation.

> Re. combining idf with structured query languages, I believe a number of
> people do this.  Let me describe what we did at InQuira briefly.  (I use
> the past tense here only because I've left the company.)  The user query
> language was either traditional keyword (with standard phrase, +/-, etc.
> operators) or a natural language question.  We used a front-end rule
> system to rewrite the user query into a structured back-end query
> language.  Those rules contain many heuristics, especially relative to
> the natural language queries (e.g., "when" translates into a date
> search, "how much" into a quantity or currency search, etc.).  The
> back-end query language has the normal Boolean, distance, etc.,
> operations, along with some less common ones (e.g., we had an ontology,
> indexed concepts, and could do concept searches).  Parts of the scoring
> were fixed while other parts were flexible:  e.g., there were three
> macro-components that could be weighted per-application or even
> per-query.  The macro-components are excerpt-relevance (how likely a
> sentence of paragraph is to precisely answer the question), document

Sentences and paragraphs are not directly supported in Lucene. Quite a while
ago Doug mentioned some mechanism to extend an index with levels,
which could probably be used for this, but I have not heard of it anymore.
OTOH one can easily define extra fields for them with different term positions
than normal, so it's not too hard to implement now.

> relevance (tfidf is a major factor here, along with page rank, etc.),
> and recency.  Re. per-query weighting, consider that current events
> oriented queries generally emphasize recency, topical queries emphasize
> document relevance and authority (tfidf and page rank), and specific
> questions emphasize excerpt relevance.
>
> Through empirical testing we found that the optimal relationship between
> the macro-score components was one of tie-breaking.  E.g., tfidf was a
> good tie-breaker among results that appeared to have equally good
> excerpts.

Would you have some background on what constitutes an excerpt here?

> I don't think Lucene should go in the question-answering direction, but
> do think it should support a range of useful scoring components that can
> be tuned and combined to yield great relevancy.  E.g., page rank is an

The recently added term vectors can be used for standard relevance feedback
mechanisms. I have not looked into these in much detail yet.
Personally I'm a bit skeptic about parameters tuned for some training set,
mostly because they are not easy to explain to end users.

> extremely valuable factor in web search, and in any domain where this is
> a rich hyper-link structure.  Having the flexibility in Lucene to easily
> incorporate this into scoring for domains where it is available and
> useful would be great.

Nutch has implemented a page rank mechanism using Lucene.

>
> If you are interested in creating a flexible scoring architecture to go
> along with your new query mechanism and I could help, please let me
> know.

I would be interested in implementing (substantial parts of) the Inquery
query language on top of  Lucene. Inquery is the best query
language specification that is publicly available, AFAIK.
However, if you know a better query language...

Earlier this year I posted some code on what I called the Surround
query language on top of Lucene. It is a first attempt to combine
proximity and truncation using the then newly introduced span queries:
http://www.mail-archive.com/lucene-dev@jakarta.apache.org/msg05504.html
It also has explicit operators for everything, but no adaptations to the
scoring mechanism of Lucene.
With some hindsight it's a nice proof of concept for a structured query
language on top of Lucene.

Regards,
Paul Elschot

>
> Chuck
>
> > -----Original Message-----
> > From: Paul Elschot [mailto:paul.elschot@xs4all.nl]
> > Sent: Sunday, October 17, 2004 3:15 AM
> > To: Lucene Developers List
> > Subject: Re: idf^2
> >
> > Chuck,
> >
> > I'm working on a more structured query language with
> > the usual explicit boolean and distance operators.
> >
> > For the scoring I'm still in mostly the design stage.
> > I'm considering to drop the idf altogether for several reasons.
> >
> > One reason is the issue of very low frequency terms (typo's in full
> > text of articles and pdf's for example) that get into queries via
>
> prefix
>
> > terms and fuzzy terms. This is also occasionaly reported  in the IR
> > literature
> > but I don't have a reference handy.
> >
> > Another reason is that the users of a structured query language
> > have their own ideas about the importance of terms. They take
> > the effort to relate their terms with the operators, and the
> > concept of idf doesn't seem to fit in there.
> >
> > Also, I'm not aware of research results on the combination of
> > idf, structured query languages and full text search. I have the
> > impression that idf is good for text like queries on abstracts,
> > and that the document length and document term
> > frequency are more important than idf for searching in full text.
> >
> > Finally, in my case, typically a classification based limitation is
> > used that reduces the searched documents to about 0.5% - 1.5%
> > of the total number of documents available, leaving the idf less
> > accurate.
> >
> > Any comments?
> >
> > ...
> >
> > >However, Lucene is a very good
> > >search engine and it seems right that it would have a "best of class"
> > >scoring formula out of the box.
> >
> > I fully agree. It's also a very good environment for
> > designing/implementing
> > another query language, although that by itself turns out to be
> > harder than I anticipated...
> >
> >
> > Regards,
> > Paul Elschot
> >
> > On Sunday 17 October 2004 02:22, Chuck Williams wrote:
> > > Doug Cutting wrote:
> > > > If someone can demonstrate that an alternate formulation produces
> > > > superior results for most applications, then we should of course
> > >
> > > change
> > >
> > > > the default implementation.  But just noting that there's a factor
> > >
> > > which
> > >
> > > > is equal to idf^2 in each element of the sum does not do this.
> > >
> > > I researched the idf^2 issue further and believe that empirical
>
> studies
>
> > > have consistently concluded that one idf factor should be dropped.
> > > Salton, the originator of the IR vector space model, decided to drop
>
> the
>
> > > idf term on documents in order to avoid the squaring.  I hear he did
> > > this after studying recall and precision for many variations of his
> > > formula.  Here is a quote from his Trec-3 paper, which references
>
> the
>
> > > same comment in his Trec-1 and Trec-2 papers.  Note the final
>
> sentence:
> > ...
> >
> > >    To allow a meaningful final retrieval similarity, it is
>
> convenient to
>
> > > use
> > >    a length normalization factor as part of the term weighting
>
> formula.
>
> > > A
> > >    high-quality term weighting formula for wik, the weight of term
>
> Tk
>
> > >    in query Qi is
> > >
> > >      wik= [ log(fik) + 1.0) * log(N/nk) ] /
> > >           sqrt(sum(j=1, t) [(log(fij) + 1.0) * log(N/nj)]^2)
>
> (1)
>
> > >    where fik is the occurrence frequency of Tk in Qi, N is the
> > > collection
> > >    size, and nk is the number of documents with term Tk assigned.
>
> The
>
> > > factor
> > >    log(N/nk) is an inverse collection frequency ("idf") factor which
> > >    decreases as terms are used widely in a collection, and the
> > > denominator
> > >    in expression (1) is used for weight normalization. This
>
> particular
>
> > > form
> > >    will be called "ltc" weighting within this paper.
> > >
> > >    The weights assigned to terms in documents are much the same. In
> > > practice,
> > >    for both effectiveness and efficiency reasons the idf factor in
>
> the
>
> > >    documents is dropped.[2, 1]"
> > >
> > > Similarly, another successful scoring formula that has been
>
> extensively
>
> > > tuned through empirical studies, OKAPI, uses idf linearly and not
> > > quadratically.  I checked with some friends that are more expert in
>
> this
>
> > > area than me, Edwin Cooper the founder and Chief Scientist at
>
> InQuira,
>
> > > and his father Bill Cooper, a pioneer in IR and professor emeritus
>
> at
>
> > > Berkeley.  Both of them report that squaring the idf term seems
> > > "strange" and is not consistent with the best known scoring
>
> formulas.
>
> > > At InQuira, we did extensive empirical tests for relevance in many
> > > different domains.  Our situation was different as the product
>
> retrieves
>
> > > specific passages so our model was much more complex (extracting
> > > sentences or paragraphs from a large corpus that specifically answer
>
> a
>
> > > natural language question -- e.g. try a question like "what are roth
>
> vs
>
> > > regular iras" in the search box at www.bankofamerica.com).  However,
>
> we
>
> > > did include document relevance factors including tfidf and did not
> > > square the idf.  None of our testing indicated that would have been
>
> an
>
> > > improvement, although we did not explicitly try it.
> >
> > ...
> >
> > > Chuck
> > >
> > > > -----Original Message-----
> > > > From: Doug Cutting [mailto:cutting@apache.org]
> > > > Sent: Wednesday, October 13, 2004 9:25 AM
> > > > To: Lucene Developers List
> > > > Subject: Re: Contribution: better multi-field searching
> > > >
> > > > Paul Elschot wrote:
> > > > >>Did you see my IDF question at the bottom of the original note?
>
> I'm
>
> > > > >>really curious why the square of IDF is used for Term and Phrase
> > > > >>queries, rather than just IDF.  It seems like it might be a bug?
> > > > >
> > > > > I missed that.
> > > > > It has been discussed recently, but I don't remember the
>
> outcome,
>
> > > > > perhaps some else?
> > > >
> > > > This has indeed been discussed before.
> > > >
> > > > Lucene computes a dot-product of a query vector and each document
> > > > vector.  Weights in both vectors are normalized tf*idf, i.e.,
> > > > (tf*idf)/length.  The dot product of vectors d and q is:
> > > >
> > > >    score(d,q) =  sum over t of ( weight(t,q) * weight(t,d) )
> > > >
> > > > Given this formulation, and the use of tf*idf weights, each
>
> component
>
> > > of
> > >
> > > > the sum has an idf^2 factor.  That's just the way it works with
>
> dot
>
> > > > products of tf*idf/length vectors.  It's not a bug.  If folks
>
> don't
>
> > > like
> > >
> > > > it they can simply override Similarity.idf() to return
>
> sqrt(super()).
>
> > > > If someone can demonstrate that an alternate formulation produces
> > > > superior results for most applications, then we should of course
> > >
> > > change
> > >
> > > > the default implementation.  But just noting that there's a factor
> > >
> > > which
> > >
> > > > is equal to idf^2 in each element of the sum does not do this.
> > > >
> > > > Doug
> >
> > ...
> >
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: lucene-dev-unsubscribe@jakarta.apache.org
> > For additional commands, e-mail: lucene-dev-help@jakarta.apache.org
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: lucene-dev-unsubscribe@jakarta.apache.org
> For additional commands, e-mail: lucene-dev-help@jakarta.apache.org


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