This paper (first reference) is the result of a class project I was part of
almost two years ago for CSCI 5417 Information Retrieval Systems. It builds
on a class project I did in CSCI 5832 Natural Language Processing and which
I presented at Wikimania '07. The project was very late as we didn't send
the final paper in until the day before new years. This technical report was
never really announced that I recall so I thought it would be interesting to
look briefly at the results. The goal of this paper was to break articles
down into surface features and latent features and then use those to study
the rating system being used, predict article quality and rank results in a
search engine. We used the [[random forests]] classifier which allowed us to
analyze the contribution of each feature to performance by looking directly
at the weights that were assigned. While the surface analysis was performed
on the whole english wikipedia, the latent analysis was performed on the
simple english wikipedia (it is more expensive to compute). = Surface
features = * Readability measures are the single best predictor of quality
that I have found, as defined by the Wikipedia Editorial Team (WET). The
[[Automated Readability Index]], [[Gunning Fog Index]] and [[Flesch-Kincaid
Grade Level]] were the strongest predictors, followed by length of article
html, number of paragraphs, [[Flesh Reading Ease]], [[Smog Grading]], number
of internal links, [[Laesbarhedsindex Readability Formula]], number of words
and number of references. Weakly predictive were number of to be's, number
of sentences, [[Coleman-Liau Index]], number of templates, PageRank, number
of external links, number of relative links. Not predictive (overall - see
the end of section 2 for the per-rating score breakdown): Number of h2 or
h3's, number of conjunctions, number of images*, average word length, number
of h4's, number of prepositions, number of pronouns, number of interlanguage
links, average syllables per word, number of nominalizations, article age
(based on page id), proportion of questions, average sentence length. :*
Number of images was actually by far the single strongest predictor of any
class, but only for Featured articles. Because it was so good at picking out
featured articles and somewhat good at picking out A and G articles the
classifier was confused in so many cases that the overall contribution of
this feature to classification performance is zero. :* Number of external
links is strongly predictive of Featured articles. :* The B class is highly
distinctive. It has a strong "signature," with high predictive value
assigned to many features. The Featured class is also very distinctive. F, B
and S (Stop/Stub) contain the most information.
:* A is the least distinct class, not being very different from F or G. =
Latent features = The algorithm used for latent analysis, which is an
analysis of the occurence of words in every document with respect to the
link structure of the encyclopedia ("concepts"), is [[Latent Dirichlet
Allocation]]. This part of the analysis was done by CS PhD student Praful
Mangalath. An example of what can be done with the result of this analysis
is that you provide a word (a search query) such as "hippie". You can then
look at the weight of every article for the word hippie. You can pick the
article with the largest weight, and then look at its link network. You can
pick out the articles that this article links to and/or which link to this
article that are also weighted strongly for the word hippie, while also
contributing maximally to this articles "hippieness". We tried this query in
our system (LDA), Google (
site:en.wikipedia.org hippie), and the Simple
English Wikipedia's Lucene search engine. The breakdown of articles occuring
in the top ten search results for this word for those engines is: * LDA
only: [[Acid rock]], [[Aldeburgh Festival]], [[Anne Murray]], [[Carl
Radle]], [[Harry Nilsson]], [[Jack Kerouac]], [[Phil Spector]], [[Plastic
Ono Band]], [[Rock and Roll]], [[Salvador Allende]], [[Smothers brothers]],
[[Stanley Kubrick]]. * Google only: [[Glam Rock]], [[South Park]]. * Simple
only: [[African Americans]], [[Charles Manson]], [[Counterculture]], [[Drug
use]], [[Flower Power]], [[Nuclear weapons]], [[Phish]], [[Sexual
liberation]], [[Summer of Love]] * LDA & Google & Simple: [[Hippie]],
[[Human Be-in]], [[Students for a democratic society]], [[Woodstock
festival]] * LDA & Google: [[Psychedelic Pop]] * Google & Simple: [[Lysergic
acid diethylamide]], [[Summer of Love]] ( See the paper for the articles
produced for the keywords philosophy and economics ) = Discussion /
Conclusion = * The results of the latent analysis are totally up to your
perception. But what is interesting is that the LDA features predict the WET
ratings of quality just as well as the surface level features. Both feature
sets (surface and latent) both pull out all almost of the information that
the rating system bears. * The rating system devised by the WET is not
distinctive. You can best tell the difference between, grouped together,
Featured, A and Good articles vs B articles. Featured, A and Good articles
are also quite distinctive (Figure 1). Note that in this study we didn't
look at Start's and Stubs, but in earlier paper we did. :* This is
interesting when compared to this recent entry on the YouTube blog. "Five
Stars Dominate Ratings"
http://youtube-global.blogspot.com/2009/09/five-stars-dominate-ratings.html…
I think a sane, well researched (with actual subjects) rating system
is
well within the purview of the Usability Initiative. Helping people find and
create good content is what Wikipedia is all about. Having a solid rating
system allows you to reorganized the user interface, the Wikipedia
namespace, and the main namespace around good content and bad content as
needed. If you don't have a solid, information bearing rating system you
don't know what good content really is (really bad content is easy to spot).
:* My Wikimania talk was all about gathering data from people about articles
and using that to train machines to automatically pick out good content. You
ask people questions along dimensions that make sense to people, and give
the machine access to other surface features (such as a statistical measure
of readability, or length) and latent features (such as can be derived from
document word occurence and encyclopedia link structure). I referenced page
262 of Zen and the Art of Motorcycle Maintenance to give an example of the
kind of qualitative features I would ask people. It really depends on what
features end up bearing information, to be tested in "the lab". Each word is
an example dimension of quality: We have "*unity, vividness, authority,
economy, sensitivity, clarity, emphasis, flow, suspense, brilliance,
precision, proportion, depth and so on.*" You then use surface and latent
features to predict these values for all articles. You can also say, when a
person rates this article as high on the x scale, they also mean that it has
has this much of these surface and these latent features.
= References =
- DeHoust, C., Mangalath, P., Mingus., B. (2008). *Improving search in
Wikipedia through quality and concept discovery*. Technical Report.
PDF<http://grey.colorado.edu/mediawiki/sites/mingus/images/6/68/DeHoustM…
- Rassbach, L., Mingus., B, Blackford, T. (2007). *Exploring the
feasibility of automatically rating online article quality*. Technical
Report.
PDF<http://grey.colorado.edu/mediawiki/sites/mingus/images/d/d3/Rassbach…