2.2. La atención médica de enfermedades agudas
2.2.3. El modelo prestador
2.2.3.1. Los profesionales médicos
We want to specify the tasks to the participants so that they would use the system to
answer questions similar to those that typical users might have. We reviewed the four
proposed use cases in Section 3.2, including choosing the right terms for SPARQL, finding
interesting facts, and detecting potential errors in instance co-reference or ontological
alignment; and from them we tried to see which use cases can be cast as a proper task for
The first use case is to use the system to choose the right terms for SPARQL. We
propose this use case as the most promising usage of the TC system. However, during
the study or in the real world, we shall not expect that the participants or average users
already understand the concepts of SPARQL before they start to use the system. A
practical way is to give a tutorial and tell the participants that we usually want to choose
the terms that are used in the most instances when no other clue is available for us to
decide which candidate term is better for queries. We illustrate the set of instances with
Venn diagram and use a made up example about the “movie” classes and the “directedBy”
properties. Then the task is stated as
Task 1. Given a pair of keywords for the class and the property as the query
purpose, can you find a matched class and a matched property from some
source (specified by their namespaces), so that using this pair of terms (the
best combination) would retrieve most of the instances in the dataset?
We also need to decide the pairs of keywords. There are a few considerations when
picking the keywords. Firstly the topic should not require any specific background knowl-
edge, i.e. it should not be in life science domains, nor about academic publications, given
that the participants could be a freshman without such knowledge. Also because we will
have multiple pairs, we want the topic of each pair to be different from those of all other
pairs. The most important reason is that we do not want a participant to acquire any
information in the task of one pair and have this information bias the subsequent task of
topic each request is related to. We list the final pairs and the best combinations we
have for the study as follows. Note that in this experiment we only consider the syntactic
matches, so that the tasks are more straightforward to the participants and the less open
questions make it easier to analyze the results. However in the future we should also
consider involving synonyms as answers.
• Task 1.1 Keyword for Class: “Company”, for Property: “location”
Best Combination: {dbpediaowl:Company, dbprop:location} (19208 instances)
• Task 1.2 Keyword for Class: “Scientist”, for Property: “award”
Best Combination: {dbpediaowl:Scientist, dbpediaowl:award} (3072 instances)
• Task 1.3 Keyword for Class: “Town”, for Property: “population”
Best Combination: {dbpediaowl:Town, dbpediaowl:populationTotal} (23104 in- stances)
• Task 1.4 Keyword for Class: “Actor”, for Property: “starring”
Best Combination: {yago:Actor109765278, dbpediaowl:starring-} (15225 in- stances)
After considering the second use case of “learning interesting facts”, we decided that it
is not specific enough to investigate with a user study. Although we believe the TC system
provides users an easy but flexible way to explore the dataset and learn something, it is
hard to quantify whether their findings are interesting and valuable, in fact, interesting and
the opportunity of interesting findings in other tasks.
The other two use cases are “detecting errors”, one involves co-reference errors and the
other ontological errors. We believe that non-expert users will find it difficult to analyze
errors in order to determine what the reason of the error is. However it might not be
hard for the users to find something that sounds improbable as long as they are presented
with a interface that they clearly understand what the data indicates. We give a made
up example of overlapping between instances of movies and book, and suggest that the
reason is because some instances of movies are considered the same as their original novel
books. This task is stated as
Task 2. Given a class as the context, browse all the overlapping class tags in
the system, can you find something that is incompatible, i.e. it is impossible
that something is both an instance of the context class and an instance of the
tag?
Again we follow the criteria we use for Task 1 to pick the context, and we also want
the total number of overlapping class tags to be shown in 10 - 15 pages. We finally decide
two contexts: Task 2.1 lgv:Stadium; Task 2.2 dbpediaowl:College. We shall discuss
more with these two sub tasks in Section 5.3.
In order to compare the TC and HL systems, we have each participant to equally use
both systems, which also means that each system will be equally used across participants.
Also we want each question to be equally answered using either system, so that we can
order of the questions shown to the participants but randomized the order of systems to
be used. To ensure the order does not change after refreshing the page, we used a pseudo
random order determined by each participant’s login user name.