Lecture06 Search tactics.ppt

advertisement
Search strategy
& tactics
Governed by
effectiveness
&
feedback
tefkos@rutgers.edu; http://comminfo.rutgers.edu/~tefko/
Tefko Saracevic
1
Central ideas
As a searcher you are striving
toward doing effective searches
• Searches are all about being effective
– finding what is needed, desired
• To measure effectiveness common measures of
precision and recall are used
• Various tactics are used to affect effectiveness
• Application of tactics for a desired results is a second
nature of professional searchers
Knowing searching = also knowing tactics to reach
toward desired effectiveness
Tefko Saracevic
2
ToC
1.
2.
3.
4.
Major concepts
Measuring effectiveness
Search tactics
Feedback types
Tefko Saracevic
3
can it be done?
1. Major concepts
Effectiveness, relevance
Tefko Saracevic
4
Some definitions
• Search statement (query):
– set of search terms with logical connectors and attributes file and system dependent
• Search strategy (big picture):
– overall approach to searching of a question
• selection of systems, files, search statements & tactics, sequence,
output formats; cost, time aspects
• Search tactics (action choices):
– choices & variations in search statements
• terms, connectors, attributes
Tefko Saracevic
5
Some definitions (cont.)
• Move :
– modifications of search strategies or tactics that are aimed
at improving the results
• e.g. from searching for digital AND libraries to digital(w)libraries
(Dialog) or “digital libraries” (Scopus)
• Cycle (particularly applicable to systems such as Dialog):
– set of commands from start (begin) to viewing (type)
results, or from a viewing to a viewing command
Tefko Saracevic
6
Search strategy – big picture
all kinds of actions from start to end
• The entire approach to a search – selection of
– files and sources to use
– approaches in proceeding to search & combining
• search terms
• operators to use
• fields to search
– formats for viewing results
– alternative actions if search yields
• too much
• too little
– problem-solving heuristics
Tefko Saracevic
7
Search tactics – specific actions
connected with a given search as it progresses
• A query - command line entered into a system in order
to retrieve relevant information
– terms, operators & attributes as allowed by a given system
– vocabulary & syntax used in conjunction with connectors
&/or limiters to search a system
• Again: depends on a system how it is done
– for example, a search statement might be:
• in Dialog: b 47; ss (garbanzo? or chickpeas) and (hum?us or humus)
• in Scopus you can enter: (garbanzo? or chickpeas) and (hum?us or
humus)
• how would you do that in
?
Tefko Saracevic
8
Search performance is expressed in terms of:
• Effectiveness :
– performance as to objectives
• to what degree did a search accomplish what desired?
• how well done in terms of relevance?
• Efficiency :
– performance as to costs
• at what cost and/or effort, time?
Both are KEY concepts & criteria for
selection of strategy, tactics & evaluation
• But here we concentrate on effectiveness
Tefko Saracevic
9
Effectiveness criteria
• Search tactics chosen & changed following some criteria
of accomplishment, such as:
–
–
–
–
–
none - no thought given – hard to imagine but happens sometimes
relevance (most often) – is it relevant to a need, task, problem?
magnitude (also very often)- is it to much retrieved to start with?
output attributes- is it trustworthy? Authorities? Understandable? …
topic – is it on the topic of the question?
• Tactics are chosen or altered interactively to match given
criteria - feedback plays an important role
Knowing what choice of tactics may produce
what results is key to professional searcher
Tefko Saracevic
10
Relevance:
key concept in IR & key criterion for assessing effectiveness
“Relevant: having significant and demonstrable
bearing on the matter at hand.”
“Relevance: the ability (as of an information retrieval
system) to retrieve material that satisfies the needs
of the user.” Merriam Webster (2005)
• Attribute/criterion reflecting effectiveness of
exchange of inf. between people (users) & IR
systems in communication contacts, based on
valuation by people
Tefko Saracevic
11
Variation in names
• Relevance or relevant can be expressed in different
terms
– utility, pertinence, appropriate, significant, useful, germane,
applicable, valuable …
• But the concept still remains relevance
That which we call relevance by any other word would
still be relevance
Relevance is relevance is relevance is relevance
thank you Shakespeare and Gertrude Stein
Tefko Saracevic
12
Some attributes of relevance
– in IR - user dependent – users are the final judges at the end
– multidimensional or faceted – users asses relevance on a
number of factors, not only topic, but also authority, novelty, source …
– dynamic – users may change relevance assessments & criteria as
search progresses or they learn or the problem is modified
– not only binary – users assess object not only as relevant/not
relevant, but also more on a continuum (scale) - partially relevant
included
– intuitively well understood – nobody has to explain to users
what it is
Tefko Saracevic
13
Consistency
Individual differences
– consistency of assessment varies over time & between
people to a (sometimes significant) degree – but
(in)consistency is similar as in other processes e.g. indexing, classifying
• Subject expertise affects consistency of relevance judgments.
Higher expertise results in higher consistency and stringency.
Lower expertise results in lower consistency and more inclusion.
– relevance is not fixed; individual differences may be large
• It is most helpful to discuss with a user what kind of
documents they may asses as relevant, what are
their criteria and then try to incorporate that in
searching & own assessments
Tefko Saracevic
14
Two major types of relevance
– leading to a dichotomy of systems- & users-view of relevance:
– Systems or algorithmic relevance
• the way a system or algorithm assesses relevance
• relation between a query & objects in the file of a system as retrieved
or failed to be retrieved by a given algorithm
– User relevance
• the way a user or user surrogate (searcher, specialist …) assesses
relevance
• relation between information need or problem at hand of a user and
objects retrieved or out there in general
– But Topical relevance can be considered by both system & user
• the degree to which topics or subjects of a query & topics or subjects
of objects (documents) in a file or retrieved match
• relation between topic in the query & topic covered by the retrieved
objects, or objects in the file(s) of the system, or even in existence
Tefko Saracevic
15
User relevance can be further
distinguished as to:
– Cognitive relevance or pertinence:
• possible changes in user’s cognitive state due to objects retrieved
• relation between state of knowledge & cognitive information need of a
user and the objects provided or in the file(s)
– Motivational or affective relevance
• matching & satisfying user’s intentions, purposes, rationales, emotions
• relation between intents, goals & motivations of a user & objects
retrieved by a system or in the file, or even in existence. Satisfaction
– Situational relevance or utility:
• value of given objects or information for user’s situation or changes in
situation
• relation between the task or problem-at-hand & the objects retrieved (or
in the files). Relates to usefulness in decision-making, reduction of
uncertainty ...
Tefko Saracevic
16
2. Measuring effectiveness
Precision and recall
Tefko Saracevic
17
How?
The basic way by which effectiveness is
established in IR searching is to compare
User
assessment
of relevance
System
assessment
of relevance
Where user assessment is the gold standard
Tefko Saracevic
18
Effectiveness measures
Two measures are used universally:
Precision:
– of the stuff retrieved & given to user how much (what %) was
relevant?
– more formally: probability that given that an object is
retrieved it is relevant, or the ratio of relevant items retrieved
to all items retrieved
Recall:
– of the stuff that is relevant in the file how much (what %) was
actually retrieved?
– more formally: probability that given that an object is relevant
it is retrieved, or the ratio of relevant items retrieved to all
relevant items in a file
Tefko Saracevic
19
Calculation
Items judged
RELEVANT
Items
RETRIEVED
a
Recall
Tefko Saracevic
=
b
No. of items retrieved &
judged relevant
Items
NOT RETRIEVED
Precision =
Items judged
NOT RELEVANT
c
No. of items retrieved
& judged not relevant
(junk)
d
No. of items not
retrieved & relevant
(missed relevant)
No. of items not
retrieved & not
relevant (missed junk)
a
a+b
High precision =
maximize a, minimize b
a
High recall =
maximize a, minimize c
a+c
20
Examples of calculation
• If a system retrieved 16 documents and only 4 were
assessed as relevant by a user then precision is 25%
• If a system had 40 documents in the file that were
relevant but managed to retrieve only 12 of them
then recall is 30%
• Precision is easy to establish, recall is not
• union of retrievals is used as a “trick” to establish relative recall
• you do a number of searches or use a number of tactics or
algorithms then you take together all that was retrieved (union)
and have that assessed; then you can calculate a relative recall of
each search, tactic or algorithm in respect to the union & see
which one provides better or worse relative recall
Tefko Saracevic
21
Interpretation: PRECISION
• Precision= percent of relevant stuff you have in your set
of answers retrieved
– or conversely percent of junk (false drops) in the answer set
– high precision = most stuff relevant
– low precision = a lot of junk
• Some users demand high precision
– do not want to wade through much stuff
– but it comes at a price: relevant stuff may be missed
• Tradeoff almost always: high precision = low recall
• we will get to tradeoff a bit later, but it is VERY important to consider
Tefko Saracevic
22
Interpretation: RECALL
• A file may have a lot of relevant stuff
• Recall = percent of that relevant stuff in the file that
you retrieved in your answer set
– conversely percent of stuff you missed
– high recall = you missed little
– low recall = you missed a lot
• Some users demand high recall (e.g. PhD students doing
a
dissertation; patent lawyers; researchers writing a proposal or article)
– want to make sure that important stuff is not missed
– but will have to pay a price of wading through a lot of junk
• Tradeoff almost always: high recall = low precision
Tefko Saracevic
23
3. Search tactics
Tradeoff between precision and recall
Using it for different tactics
Tefko Saracevic
24
Aim of search tactics
• Since there is no such thing as a perfect search:
– the aim is to search in a way that will insure a given, chosen or
desired, level of effectiveness
• That means that we have to agree on
– what do we mean by effectiveness in searching?
• general agreement is that retrieval of relevant answers is the major
criterion for effectiveness of searching
– what measures do we use to express achievement of
effectiveness in terms of relevance
• general agreement is that we use measures of precision and recall (or
derivatives)
Thus many search tactics are geared toward
achieving certain level of precision or recall
Tefko Saracevic
25
Precision-recall trade-off
• USUALLY: precision & recall are inversely related
– higher recall usually means lower precision & vice versa
Precision
100 %
0
Tefko Saracevic
Recall
100 %
26
Precision-recall trade-off
• It is like in life, usually:
• you get some, lose some; you can't have your cake and eat it too
• Usually, but not always (keep in mind these are probabilities)
– when you have high precision most stuff you got is relevant or
on the target but you also missed other stuff (could be a lot)
that is relevant – it was left behind
– when you have high recall you did not miss much but you also
got junk (could be a lot) - you have to wade through it
• There is price to pay either way
– but then a lot of users are perfectly satisfied with and are
aiming at high precision
• give me a few good things, that is all what I need
Tefko Saracevic
27
Effectiveness (Cleverdon’s) “laws”
High precision in retrieval is usually
associated with low recall.
High precision = low recall
High recall in retrieval is usually associated
with low precision.
High recall = low precision
You – with your user – have to decide if you are
aiming toward high recall or high precision &
you have to be aware of & explain tradeoffs
You use different tactics for high
recall from those for high precision
Tefko Saracevic
Originally
formulated by
Cyril
Cleverdon,
a UK librarian,
in 1960s, in
first IR tests
done in
Cranfield, UK,
thus “Cranfield
tests”
28
Search tactics
several ‘things’ or variables in a query can be
selected & changed to affect effectiveness
Variable
What variation possible?
1. LOGIC
choice of connectors among terms
AND, OR, NOT, W …
2. SCOPE
no. of terms linked - ANDs
A AND B vs A AND B AND C
3.EXHAUSTIVITY
for each concept no. of related terms - OR
connections
A OR B vs. A OR B OR C
4. TERM SPECIFICITY
for each concept level in hierarchy
broader vs narrower terms
5. SEARCHABLE FIELDS
choice for text terms & non-text attributes
e.g. titles only, limit as to years, various sources
6. FILE OR SYSTEM SPECIFIC
CAPABILITIES
using capabilities, options available
e.g. given fields, ranking, sorting, linking
Tefko Saracevic
29
Actions and consequences
BUT: each variation has consequence in output
if you do X then Y will happen
Tefko Saracevic
Action
Consequence
SCOPE
- adding more ANDs
Output size: down
Recall: down
Precision: up
EXHAUSTIVITY
- adding more ORs
Output size: up
Recall: up
Precision: down
USE OF NOTs
- adding more NOTs
Output size: down
Recall: down
Precision: up
USING BROAD TERMS
- low specificity
Output size: up
Recall: up
Precision: down
USING PHRASES
- high specificity
Output size: down
Recall: down
Precision: up
30
Tactics: what to do?
• These “laws” lead to precision & recall devices
– tactics to increase/decrease precision or recall
Tefko Saracevic
To increase precision
To increase recall
SCOPE
-add more ANDs
-- you restrict
NOTs
-add more NOTs
-you eliminate
EXHAUSTIVITY
-add more ORs
- you enlarge
BROAD TERMS
- you broaden terms
PHRASES
- you become more specific
USE MORE TRUNCATION
-you use grammatical variants of
terms
31
Precision, recall devices
With experience use of these devices
will become second nature
Tefko Saracevic
NARROWING
Higher precision
BOADENNING
Higher recall
More ANDs
Fewer ORs
More NOTs
Less free text
More controlled
Less synonyms
Narrower terms
More specific
Less truncation
More qualifiers
More limits
Building blocks
Fewer ANDs
More ORs
Fewer NOTs
More free text
Fewer controlled
More synonyms
Broader terms
Less specific
More truncation
Fewer qualifiers
Fewer limits
Citation growing
32
Other tactics
(but that gets us in the next unit on advanced searching)
• Citation growing:
–
–
–
–
find a relevant document
look for documents cited in
look for documents citing it
repeat on newly found relevant documents
• Building blocks
– find documents with term A
– review – add term B & so on
• Using different feedbacks
– a most important tool
Tefko Saracevic
33
4. Feedback
Types. Berry picking
Tefko Saracevic
34
Feedback in searching
• Extremely important!
• And constantly used, consciously or unconsciously
• Formally:
– the process in which part of the output of a system is
returned to its input in order to regulate its further output
• Simply put:
– you search, find something that may be relevant, then look
at it, then on the basis of AHA! change your next query or
tactic to get better or more stuff
Tefko Saracevic
35
Feedback in searching …
• Any feedback implies loops
– a completion of a process provides information for
modification, if any, for the next process
– information from output is used to change previous or
create new input
• In searching:
– some information taken from output of a search is used to
do something with next query (search statement)
• examine what you got to decide what to do next in searching
– a basic tactic in searching
• Several feedback types used in searching
– each used for different decisions
Tefko Saracevic
36
Feedback types
• Content relevance feedback
– judge relevance of items retrieved
– make decision what to do next
• switch files, change exhaustivity …
• Term relevance feedback
– find relevant documents
– examine what other terms used in those documents
– search using additional terms
• also called query modification & in some systems done automatically
• Magnitude feedback
– on the basis of size of output make tactical decisions
• often the size so big that documents are not examined but next search
done to limit size
Tefko Saracevic
37
Feedback types (cont.)
• Tactical review feedback
– after a number of queries (search statements) in the same
search review tactics as to getting desired outputs
• review terms, logic, limits …
– change tactics accordingly
• Strategic review feedback
– after a while (or after consultation with user) review the
“big” picture on what searched and how
• sources, terms, relevant documents, need satisfaction, changes in
question, query …
– do next searches accordingly
– used in reiterative searching
Tefko Saracevic
38
But then
• There are other ways to look at searching where
many of these things are combined
• Here is one way that looks at the whole search as a
process of complex wandering
• Shifting exploration & feedback are implied
Tefko Saracevic
39
Bates’ Berry-picking model of searching
“…moving through many actions
towards a general goal of
satisfactory completion of research
related to information need.”
– query is shifting (continually)
• as search progresses queries are changing
• different tactics are used
– searcher (user) may move through a
variety of sources
Elaborated by
Marcia Bates
UCLA
• new files, resources may be used
– think of your last serious search
• isn’t this what you were doing?
Tefko Saracevic
40
Berry-picking …
– closer to the real behavior of information searchers
– new information may provide new ideas, new
directions
• feedback is used in various ways
– question is not satisfied by a single set of answers, but
by a series of selections & bits of information found
along the way
• results may vary & may have to be provided in appropriate
ways & means
– you go as if berry-picking in a field
Tefko Saracevic
41
A berry-picking evolving search
(from the article)
“1) The nature of the query is an evolving one, rather than single and
unchanging, and 2) the nature of the search process is such that it follows a
berrypicking pattern, instead of leading to a single best retrieved set.”
Tefko Saracevic
42
Conclusions
search tactics have to be mastered
• Users are not concerned about searching but about
finding
• Effective searching is a prerequisite for finding the
right stuff
• Search tactics are critical for effective searching
• They need to be understood and followed
if you do X you can expect Y
in order to get to Y you have to know what X to do
Tefko Saracevic
43
Thank you
M. C. Escher
Tefko Saracevic
44
Download