Hi Jim,
Determining the intent of a particular search is indeed very difficult, and
is not really feasible to even attempt it at the scale needed for machine
learning (unless you have an immense budget like some for-profit search
engine companies).
For our machine learning training data, we use click models suggested by
academic research. These models allow us to score the results for a given
query based on which results users actually clicked on (and didn't click
on). The results aren't perfect, but they are good, and they can be
automatically generated for millions of training examples taken from real
user queries and clicks.
These scores serve as a proxy for user intent, without needing to actually
understand it. As an example, if 35% of people click on the first result
for a particular query, and 60% on the second result, the click scores
would indicate that the order should be swapped, even without knowing the
intent of the query or the content of the results.
Swapping the top two results isn't really a big win, but the hope is that
by identifying features of the query (e.g., number of words), of the
articles (e.g., popularity), and of the relationship between them (e.g.,
number of words in common between the query and the article title) we will
learn something that is more generally true. If we do, then we may move a
result for a different query from, say, position 8 (where few people ever
click) to position 3 (where there is at least a chance of a click).
Iterating the whole process will allow us to detect that the result newly
in position 3 is actually a really popular result so we should adjust the
model to boost it even more, or that it's not that great and we should
adjust the model to put something better in the #3 slot. Of course, all of
the "adjusting" of the model happens automatically during training.
Through this iterative process of modeling, training, evaluation, and
deployment, we are attempting to take into account the relationship between
the user's intent and the search results—inferred from the user's
behavior—to improve the search results.
Cheers,
—Trey
Trey Jones
Software Engineer, Discovery
Wikimedia Foundation
On Fri, Jun 16, 2017 at 10:26 AM, James Salsman <jsalsman(a)gmail.com> wrote:
Hi Trey,
Thanks for your very detailed reply. I have a followup question.
How do you determine search intents? For example, if you see someone
searching for "rents" how do you know whether they are looking for
economic or property rents when evaluating the quality of the search
results? If you're training machine learning models from "5, 50, or
500," example you need to have labels on each of those examples
indicating whether the results are good or not.
Do you interview searchers after the fact? Ask people to search and
record the terms they search on? What kind of infrastructure do you
have to make sure you're getting correct intents robust enough to
score the example results? Maybe surveys occurring on some small
fraction of results asking users to describe in greater detail exactly
what they were trying to find?
Best regards,
Jim
On Thu, Jun 15, 2017 at 10:40 PM, Deborah Tankersley
<dtankersley(a)wikimedia.org> wrote:
James Salsman wrote:
How will the Foundation's approach to machine learning of search
results ranking guard against overfitting?
Overfitting, for those who aren't familiar with the term, describes the
situation where a machine learning model inappropriately learns very
specific details about its training set that don't generalize to the real
world. From the point of view of training, the model seems to be getting
better and better, while real-world performance is actually decreasing.
As
a somewhat silly example, a model could learn
that queries that have
exactly 38 words in them are 100% about baseball—because there is only
one
example of a query in the training set that is 38
words long, and it is
about baseball. For more on overfitting, see Wikipedia.[1]
We employ the usual safeguards against overfitting. Certain parameters
that
control how a specific type of model is built can
discourage overfitting.
For example, not allowing a decision inside the model to be made on too
little data—so rather than 1 or 2 examples to base a decision on, the
model
can be told it needs to see 5, or 50, or 500.
We also have separate training and testing data sets. So we build a model
on one set of data, then evaluate the model on another set. The estimate
of
model performance from the training set will
always be at least a bit
optimistic, but the testing set—which is large enough to be
representative
and which does not overlap with the training
set—gives a more realistic
estimate. We choose the model that performs the best on the testing set.
Overfitted models will do worse on the testing set, and we won't use
them.
We have other methods of validating our models as well.
We have a set of machines and software that we collectively call
Relevance
Forge (a.k.a. RelForge) that we can use to run
large sets of queries
against different versions of the same index. We can compare the before
and
after results, both automatically and manually.
RelForge lets us easily
gauge the *impact* of a change. For example, a 1% net improvement could
come from making 1% of queries a bit better, or from making 49% a bit
worse
and 50% a bit better. So, we can easily see
whether 1% or 99% of results
change. If we see a 2% improvement but a 99% impact, something weird is
happening, and we'd investigate more deeply.
We also have many definitions of "results change" that we can evaluate:
#1
result changes, top 3 results change (ordered or
unordered), number of
results changes, number of queries getting zero results changes. And for
each of these we can manually inspect a random selection of affected
queries to decide whether the results are generally better or not.
We also run A/B tests, where we let a small sample of users get the
proposed change, while a similar number get the standard results. We do
statistical analyses on user engagement with results and various other
click metrics that let us compare the control and experimental
conditions.
For more on how we test search changes in
general, see Testing Search on
mediawiki.org.[2]
In both of these cases—RelForge testing and A/B testing in
production—overfitted models would perform poorly, and that would become
apparent.
For example, if most searches on "rent" do not pertain to "rent
seeking", then how will the machine learning
approach to search
results for "rent" guard against never presenting any results on "rent
seeking"?
Your wording has left me a bit confused, and I'm not sure whether your
concern is (a) that a query of "rent" should never return "rent
seeking",
and so the machine learning model should never present it, or (b) that we
should guard against building a model that *never* presents results on
"rent seeking" for a query of "rent". I'll briefly address each.
Case (a): "rent" should *never* return "rent seeking"
It's not clear to me that returning "rent seeking" for a query of
"rent"
is
necessarily a case of overfitting per se, but in
general the click models
that we use would take note that users who search for "rent", say, click
on
the musical 70% of the time and the
disambiguation page 29% of the time.
Those would be the "good" results and the model would prioritize moving
them to the top of the list.
*Never* presenting results on "rent seeking" would be an error. The word
is
present in the article, and in the title, so it
should be somewhere in
the
results. Moving it up or down the results list is
a question of ranking,
which is what the machine learning model is trying to figure out.
Case (b): "rent" should not be *prevented* from returning "rent
seeking"
Our click data shows that about 80% of clicks on search results are on
one
of the first two results, and more than 90% are
on the top 10. Our click
models for scoring the order of results reflect that. All of the value
then, from the machine learning model's point of view, comes from getting
the top 3 to 5 results in the best possible order. There's not a lot of
value in pushing down any particular result much farther than that. For a
single word query like "rent", title matches are the best. There are only
138 results for intitle:rent, vs over 44K for just rent—however, the
first
page of results for both is the same.
We are interested in use cases other than searchers who are looking for a
particular article or particular information, though that tends to
predominate. Editors might want to find all the articles with a
particular
word (e.g., a misspelling) and no result would be
excluded by the machine
learning model, just possibly ranked lower.
Hope that helps,
—Trey
[1]
https://en.wikipedia.org/wiki/Overfitting
[2]
https://www.mediawiki.org/wiki/Wikimedia_Discovery/Search/ Testing_Search
Trey Jones
Software Engineer, Discovery
Wikimedia Foundation
*(via Deb Tankersley's email address as Trey's original email got
moderated)*
On Tue, Jun 13, 2017 at 3:43 PM, James Salsman <jsalsman(a)gmail.com>
wrote:
> On Wed, Jun 14, 2017 at 5:25 AM, Deborah Tankersley
> <dtankersley(a)wikimedia.org> wrote:
> >
> > The Discovery team structure has now changed, but the new teams will
> still
> > work together to complete the goals as listed in the draft annual
> plan.[2]
> > A summary of their anticipated work, as we finalize these changes, is
> > below. We plan on doing a check-in at the end of the calendar year to
see
> > how our goals are progressing with the
new smaller and separated team
> > structure.
> >
> > Here is a list of the various projects under the Discovery umbrella,
> along
> > with the goals that they will be working on:
> >
> > Search Backend
> >
> > Improve search capabilities:
> >
> > Implement ‘learning to rank’ [3] and other advanced machine
learning
How will the Foundation's approach to machine learning of search
results ranking guard against overfitting?
For example, if most searches on "rent" do not pertain to "rent
seeking", then how will the machine learning approach to search
results for "rent" guard against never presenting any results on "rent
seeking"?
_______________________________________________
Wikimedia-l mailing list, guidelines at:
https://meta.wikimedia.org/
wiki/Mailing_lists/Guidelines and
https://meta.wikimedia.org/
wiki/Wikimedia-l
New messages to: Wikimedia-l(a)lists.wikimedia.org
Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l,
<mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe>
_______________________________________________
Wikimedia-l mailing list, guidelines at:
https://meta.wikimedia.org/wik
i/Mailing_lists/Guidelines and
https://meta.wikimedia.org/wiki/Wikimedia-l
New messages to: Wikimedia-l(a)lists.wikimedia.org
Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l,
<mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe>
_______________________________________________
Wikimedia-l mailing list, guidelines at:
https://meta.wikimedia.org/wik
i/Mailing_lists/Guidelines and
https://meta.wikimedia.org/wiki/Wikimedia-l
New messages to: Wikimedia-l(a)lists.wikimedia.org
Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l,
<mailto:wikimedia-l-request@lists.wikimedia.org?subject=unsubscribe>