Building a Query Engine for a Corpus of Open Data
Mauro Pelucchi, Giuseppe Psaila and Maurizio Toccu
University of Bergamo - DIGIP, Viale Marconi 5, I-24044 Dalmine (BG), Italy
Keywords:
Retrieval of Open Data, Blind Querying, Single Item Extraction.
Abstract:
Public Administrations openly publish many data sets concerning citizens and territories in order to increase
the amount of information made available for people, firms and public administrators. As an effect, Open Data
corpora has become so huge that it is impossible to deal with them by hand; as a consequence, it is necessary
to use tools that include innovative techniques able to query them.
In this paper, we present a technique to select open data sets containing specific pieces of information, and
retrieve them in a corpus published by a portal of open data. In particular, users can formulate structured
queries blindly submitted to our search engine prototype (i.e., being unaware of the actual structure of data
sets). Our approach reinterpret and mixes several known information retrieval approaches, giving at the same
time a database view of the problem. We implemented this technique within a prototype, that we tested on a
corpus containing more that over 2000 data sets. We noted that our technique provides focused results w.r.t.
the baseline experiments performed with Apache Solr.
1 INTRODUCTION
Public Administrations are publishing more and more
data sets concerning their activities and citizens.
Since these data sets are open, i.e. publicly avail-
able for any interested people and organizations, they
are called Open Data. The reader should not confuse
the concept of Open Data for Public Administration
(which is the context of this paper), with the concept
of Linked Open Data, i.e., the ability of describing the
content of web sites linked each other through RDF
(Resource Description Framework).
Public Administrations extended there web sites
with Open Data Portals; typically (see (H
¨
ochtl and
Lampoltshammer, 2016)) they exploit two frame-
works named Socrata and CKAN, that can be viewed
as CMS (Content Management Systems) to support
the publication of data sets within an Open Data Por-
tal.
Data sets published by Open Data Portals are
Structured data sets: in the simplest form, they are
CSV files with a fixed number of fields, with field
names reported in the first line. Alternatively, they are
often represented as vectors of JSON objects, without
nesting (they must be represented in a flat way), and
(less frequently) as structured XML documents (with-
out exploiting nesting capabilities of XML). In prac-
tice, data sets are always thought as tables: an item in
the data set (CSV row, JSON object, XML element)
represents a table row.
Nowadays, the number of data sets published by a
single Open Data Portal is very large and is continu-
ously increasing. Often, a corpus published by one
single publishing system can contain several thou-
sands of data sets.
People (typically analysts) that need to find the
right data sets among the mass of available data sets
in an Open Data Portal have to work hard to find what
they look for. They need a powerful search tool to
query the above corpus.
We observed that users, a priori, do not know
the actual structure of thousands of heterogeneous
data sets collected in the corpus. So, they can only
blindly query the corpus, making some hypothesis
about property or field names. Furthermore, we also
observed that often they are not even interested in an
entire data set, but only in some rows (items) among
those contained in the data set, i.e., only those sat-
isfying a given selection condition. For example, a
user/analyst might want to get information about high
schools located in a given city named “My City”,
being interested in their name, address and reputa-
tion. The query mechanism should be able to focus
the search on those data sets that actually contain the
items of interest, and retrieve them. However, at the
same time, it should be able to extend the search to
126
Pelucchi, M., Psaila, G. and Toccu, M.
Building a Quer y Engine for a Corpus of Open Data.
DOI: 10.5220/0006308801260136
In Proceedings of the 13th International Conference on Web Information Systems and Technologies (WEBIST 2017), pages 126-136
ISBN: 978-989-758-246-2
Copyright © 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

those data sets that could potentially contain items of
interest, by considering similar names in the corpus
catalog (i.e., the list of data sets in the corpus and their
fields). For example, a data set with field “city name”
or simply “city” in place of “cityname”.
Therefore, in this paper, we propose a technique
for querying a corpus of Open Data Sets on the basis
of queries that users blindly formulate on the corpus.
The technique mixes a query expansion mechanism,
which relies on string similarity matching, and a key-
word selection mechanism based on the concepts of
informativeness and representativeness.
The paper is organized as follows. Section 2 gives
some preliminary definitions. Section 3 formally de-
fines the problem we are dealing with. In Section 4,
the overall query process is presented. Then, Sec-
tion 5 presents the query assessment step and the key-
word selection step. Section 6 introduces the query
expansion step. Finally, Section 7 presents the steps
concerning open data set retrieval and filtering. The
experimental evaluation is shown in Section 8, while
Section 9 discusses some related work. Section 10
draws the conclusions.
2 PRELIMINARY DEFINITIONS
In order to facilitate the paper comprehension, we
give some preliminary definitions.
Definition 1: Data Set. An Open Data Set ods is
described by a tuple
ods: <ods id, dataset name, schema, metadata>
where ods id is a unique identifier, dataset name is
the name of the data set (not unique); schema is a
set of fields names, metadata is a set of pairs (la-
bel, value), which are additional meta-data associated
with the data set. 
Definition 2: Instance and Items. With Instance,
we denote the actual content of data sets. It is a set
of items, where an item is a row of a CSV file or data
table, as well as a flat object in a JSON vector. 
Many open data portals provide open data sets in
several formats. Commonly, they adopt CSV (comma
separated values), but it is becoming frequent the
adoption of JSON (JavaScript Object Notation); XML
is also used, but less frequently.
The adoption of JSON is motivated by the fact
that, now, it has become the de facto standard for data
sharing in the world of Web Services. In fact, it is easy
to read, much lighter than XML and JavaScript (the
programming language used for developing client-
side web applications) natively interprets JSON data.
However, in the open data world, it is used only
as an alternative representation for flat tabular data: a
CSV file can always be transformed into a vector of
JSON objects, where all objects in the vector has the
same structure, i.e., the same fields and fields have
only simple values (there are not nested objects).
Consequently, JSON capabilities to describe
nested complex structures, as well as collections of
heterogeneous objects, are not exploited in this con-
text.
To illustrate, consider Figure 1. We report a
fragment of the open data set named Local Author-
ity Maintained Trees published on the London Data
Store, that is the official open data portal of London
City. Notice how a flat CSV file (in the upper part of
the figure) can be easily transformed into a vector of
JSON objects (lower part of the figure). As the reader
can see, the price to pay is that the JSON version is
more verbose, since field names mast be repeated for
each object. However, in the optic of integrating the
items extracted from several open data sets, having
different structures, into a unique collection of results,
JSON is the natural representation format for our tool.
Definition 3: Global Meta-data. The list of fields
of items in the open data set ods is ods.schema. We
call the triple <data set name, schema, metadata>
the Global Meta-Data of the data set. 
Definition 4: Corpus and Catalog. With C = {ods
1
,
ods
2
, ...} we denote the corpus of Open Data sets.
The catalog of the corpus is the list of descriptions,
i.e., global meta-data, of open data sets in C (see Def-
inition 1). 
Definition 5: Query. Given a Data Set Name dn, a
set P of field names (properties) of interest P = { pn
1
,
pn
2
, ...} and a selection condition sc on field values,
a query q is a triple q: <dn, P, sc>. 
Definition 6: Query Term. With Query Term (term
for simplicity) we denote a data set name q.dn, a field
name appearing in q.P, in q.sc, a constant appearing
in q.sc. 
3 PROBLEM DEFINITION
Consider an open data corpus published by an Open
Data Portal, i.e., a system that publishes open data
sets.
Queries allow users to discover, in the corpus C,
those data sets, possibly having the specified name
q.dn or a similar one, that contain items (e.g., rows
or JSON objects) with the desired field names q.P (or
similar) and can be actually filtered on the basis of
Building a Query Engine for a Corpus of Open Data
127

CSV Version
ID,Borough,Species name,Common name,Display name,Ox,Oy
1,Barking ,ACER PSEUDOPLATANUS ’BRILLIANTISSIMUM’,,Maple,548320.8,189593.7
2,Barking ,TAXUS BACCATA FASTIGIATA,,Other,548297.9,189590.2
JSON Version
[
{
"ID": 1,
"Borough": "Barking ",
"Species name": "ACER PSEUDOPLATANUS ’BRILLIANTISSIMUM’",
"Common name": "",
"Display name": "Maple",
"Ox": 548320.8,
"Oy": 189593.7
},
{
"ID": 2,
"Borough": "Barking ",
"Species name": "TAXUS BACCATA FASTIGIATA",
"Common name": "",
"Display name": "Other",
"Ox": 548297.9,
"Oy": 189590.2
}
]
Figure 1: Fragment from the data set Local Authority Maintained Trees published in the London Data Store.
the selection condition q.sc, so that users can obtain
filtered items. By relying on the definition of a rel-
evance measure for data sets (in Section 7), we can
define the overall problem.
Problem 1: Given a Relevance Measure rm(ods)∈
[0,1] of a data set ods∈C, and a minimum threshold
th∈ [0, 1], a data set ods is relevant if rm(ods)> 0 and
rm(ods)≥th.
The result set RS = {o
1
, o
2
, . .. } of query q is the set
of items (rows or JSON objects) o
i
such that ods
j
is a
relevant data set, o
i
∈ Instance(ods
j
) and o
i
satisfies
the selection condition q.sc. 
Example 1: The sample query about schools in Sec-
tion 1 becomes:
q :<dn=Schools,
P= {Name, Address, Reputation},
sc= (City="My City" AND
Type="High School") >. 
4 PROCESSING STEPS
The contribution of this paper is a technique to solve
Problem 1.
The proposed technique is built around a query
mechanism based on the Vector Space Model (Man-
ning et al., 2008), encompassed in a multi-step pro-
cess devised to deal with the high heterogeneity of
open data set and the blindly query approach (the user
does not know the actual schema of data sets in the
corpus). Recall that, in our vision, the user performs
a blindly query, but asks for specific field names and
for items (rows or JSON objects) that satisfy a precise
selection condition. The system is fully aware of data
set schemata, but due to their heterogeneity it is not
possible to rely only on exact matching. Thus, it is
necessary to focus the search, identifying those terms
in the query that mostly represent the query essence
and the capability of the query to select items within
data set instances. At the same time, it is necessary to
expand the search space, considering terms which are
similar to those in the query, in order to address the
fact that users are unaware of actual schemata of data
sets.
Moving from the above considerations, we de-
fined the steps performed by our technique.
• Step 1: Query Assessment. Terms in the query
are searched within the catalog. If some term t
is missing, t is replaced with term t, the term in
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
128

the meta-data catalog with the highest similarity
score w.r.t. t. We obtain the assessed query q.
• Step 2: Keyword Selection. Keywords are selected
from terms in the assessed query q, in order to
find the most representative/informative terms for
finding potentially relevant data sets.
• Step 3: Neighbour Queries. For each selected
keyword, a pool of similar terms which are
present in the meta-data catalog are identified, in
order to derive, from the assessed query q, the
Neighbour Queries, i.e., queries which are similar
to q. Both q and the derived neighbour queries are
in the set Q of queries to process. For a neigh-
bour query, the set of relevant keywords is ob-
tained from the set of relevant keywords for q by
replacing original terms with similar terms.
• Step 4: VSM Data Set Retrieval. For each query
in Q, the selected keywords are used to retrieve
data sets based on the Vector Space Model (Man-
ning et al., 2008): in this way, the set of possibly
relevant data sets is obtained.
• Step 5: Schema Fitting. The full set of field names
in each query in Q is compared with the schema of
each selected data set, in order to compute the rel-
evance measure rm. Data sets whose schema bet-
ter fits the query will be more relevant than other
data sets.
• Step 6: Instance Filtering. Instances of relevant
data sets are processed in order to filter out and
keep only the items (rows or JSON objects) that
satisfy the selection condition.
5 QUERY ASSESSMENT AND
KEYWORD SELECTION
In Step 1, query q is assessed against the meta-data
catalog, in order to check for the presence of terms
in q: if all terms in q are present in the catalog, the
assessed query q is the same query q; otherwise, the
assessed query q is derived from q by substituting the
missing terms with the most similar and representa-
tive terms in the catalog. Since the concept of rep-
resentativeness is introduced hereafter in this section,
we will formally define the notion of most similar and
representative term in Section 5.2.
5.1 Keyword Selection
Consider the assessed query q. Terms appearing in q
are not equally important/informative for the query. If
the set of terms used to retrieve possibly relevant data
sets contains terms that are not important or much less
important than others, in Step 4 we can retrieve data
sets that may be not actually relevant for the query.
Thus, we adopted a keyword selection technique that
extracts, from the assessed query q, the set of (mean-
ingful) keywords K(q).
We were inspired by the technique introduced in
(Liu et al., 2006) for the context of classical search
engines. However, we introduced several variations
(structure of the graph and values of scores) to cope
with peculiarities of our problem.
The keyword selection technique proceeds as fol-
lows:
1. The Term Graph is built, where each node cor-
responds to a term in the assessed query q and
an edge connects two terms and weights their re-
lationship (see Figure 3 for a sample graph dis-
cussed after the keyword selection algorithm re-
ported in Figure 2).
2. An iterative algorithm:
(a) Chooses the most relevant term.
(b) Modifies the weights of not selected terms,
propagating a decreasing reduction of informa-
tiveness to neighbours of the selected term.
3. Only most informative terms are kept and inserted
into the keyword set K(q ).
Term Graph Construction. Each node in the Term
Graph corresponds to a term in the query. There are
three types of nodes: one Data Set Name Node, cor-
responding to the data set name q.dn; Field Nodes,
corresponding to field names in q.P or in q.sc; Value
Nodes, corresponding to constants appearing in com-
parison predicates in q.sc.
Nodes are connected each other by an edge if
terms are related to each other. In particular: a field
node is connected with the data set name node; two
field nodes are connected if they are the operands of
the same comparison predicate in q.sc; a field node
and a value node are connected if the field (described
by the field node) is compared with the constant value
(described by the value ode) by a comparison predi-
cate in q.sc.
Each node (term) has two scores: a representative-
ness score, denoted as r score, and an informativeness
score, denoted as i score.
Definition 7: Representativeness Score. The rep-
resentativeness score r score denotes the capability of
the term to find a relevant data set, i.e., how much it
represents a specific subset of mapped data sets. In
particular, we consider two components for represen-
tativeness: Intrinsic Representativeness and Extrinsic
Building a Query Engine for a Corpus of Open Data
129

Representativeness. The former is related to the pres-
ence of the term in the query; we denote it as irep(t)
(where t is a term). The latter is related to the pres-
ence of the term in global meta-data of data sets, thus
it measures the actual capability of the term to focus
the search; we denote the extrinsic representativeness
as erep(t). Both irep and erep are defined in the range
[0,1].
The r score of a term is obtained by combining
the intrinsic and the extrinsic representativeness, by
means of the average of them, i.e., r score= (irep(t)+
erep(t))/2. This way, both contribute and balance
each other. 
Intrinsic Representativeness. Depending on the role
played by a term in the query, we established different
values for the intrinsic representativeness.
- The intrinsic representativeness of the Data Set
Name is irep(t)= 1.0 if there exists a data set q.dn in
the corpus C; otherwise, irep(t)= 0.0.
- The intrinsic representativeness of a Field Name is
irep(t)= 0.5 if there exists at least a data set that has a
field named t; otherwise, irep(t)= 0.0.
- The intrinsic representativeness for string value
nodes and date value nodes, is irep(t)= 0.5, while it
is irep(t)= 0.2 for numerical value nodes.
Extrinsic Representativeness. The approach we
adopted to define the extrinsic representativeness, is
inspired by the idf (inverse document frequency) mea-
sure (Manning et al., 2008). Its classical definition is
idf =log(D / d(t)), where D is the total number of doc-
uments, while d(t) is the number of documents that
contain term t. The lower the number of documents
that contain t, the higher the value of idf, because term
t is more capable to focus the search on relevant doc-
uments.
Moving from this idea, we defined the extrinsic
representativeness in order to be in the range [0,1] and
to be suitable for our context.
Definition 8: Normalized Data Set Frequency.
Consider the corpus C of open data sets and its car-
dinality |C|. Consider a term t, the set Matched(t) of
data sets that contain term t in global meta-data (such
that Matched(t)⊆C) and its cardinality |Matched(t)|.
The Normalized Data Set Frequency of term t is
defined as
ndf(t)= log
2

|Matched(t)|
|C|
+ 1

that is defined in the range [0,1].
If term t is contained in the global meta-data of all
data sets, ndf(t)= 1 (the argument of the logarithm is
2. If term t is not contained in any global meta-data,
ndf(t)= 0, because the argument of the logarithm is 1.

Definition 9: Extrinsic Representativeness. Given
a term t and its normalized data set frequency ndf(t),
its Extrinsic Representativeness is defined as:
erep(t)= 1 − ndf(t) if ndf(t)> 0
erep(t)= 0 if ndf(t)= 0 
The rationale is the following: if a term is very
frequent, its extrinsic representativeness is close to 0,
because it is not able to focus the search. In contrast,
not frequent terms are able to focus the search, and
their extrinsic representativeness is close to 1.
Definition 10: Informativeness Score. The infor-
mativeness score i score measures the degree of in-
formation that could be obtained by selecting a term
as keyword. At the beginning, an initial value is set:
we chose i score= 1.0 for data set name nodes and
field nodes that appear in the selection condition q.sc.
We chose i score= 0.8 for value nodes and field nodes
appearing only in q.P. In fact, constants are naturally
less informative than data set names and field names;
furthermore, they are less likely to appear in meta-
data of data sets, thus the data set retrieval step based
on VSM would result ineffective. The same is for
field names not appearing in the selection condition:
they are less informative than those appearing in the
selection condition.
During the keyword selection process, the selec-
tion of a term affects the informativeness of neigh-
bours: in fact, selecting two neighbours may be re-
dundant. Consequently, the procedure will modify the
i score values of not selected terms. 
Keyword Selection Algorithm. The algorithm is in-
spired by the algorithm reported in (Liu et al., 2006),
but modified because, in our term graph, edges are
neither labeled nor weighted. The main procedure of
the algorithm, named KeywordSel, is reported in Fig-
ure 2.
The algorithm works on the input sets of nodes
and edges; notice that nodes have an additional field
named marked, denoting selected nodes when true.
The main cycle from lines 2 to 13, at each step se-
lects the node not yet selected with the highest value
for i score + r score, provided that the sum is greater
than 1. This is because below this threshold the capa-
bility of retrieving significant data sets becomes too
low. The for loop from line 4 to line 8 looks for the
highest i score + r score value (and the correspond-
ing node) in not marked nodes. Then, the if instruc-
tion at line 9 verifies that the selected node (if any)
has the sum i score + r score greater than 1: if so, the
node is marked as selected, the corresponding term is
inserted into the keyword set K and the procedure Up-
dateIScore is called (line 12) to propagate a reduction
of informativeness.
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
130

Procedure KeywordSel
Input: N set of nodes, E set of edges
a node is n =< t,i score,r score,marked = false >
Output: the set of selected keywords K
begin
1. K =
/
0; // set of selected keywords
do
2. max score= 0.0; //initial max score
3. max node=null; //no node selected
4. for each node n ∈ N with n.marked =false do
5. score= n.i score + n.i score;
6. if score>max score then
7. max score=score;
8. max node= n;
end if
end for each
9. if n 6=null and max score> 1.0 then
10. K = K ∪ {n.t}; //term selected as keywordd
11. n.marked =true; //node marked as selected
12. call procedure UpdateIScore(N, E, n);
end if
13. while max
score> 1.0;
end procedure
Figure 2: Procedure KeywordSel, that selects possibly rele-
vant keywords.
Procedure UpdateIScore performs the update of
the i score, by propagating from the selected node.
The propagation mechanism can be described in a re-
cursive way.
- Starting from the selected node s, for each node
z
0
connected with s, the reduction factor is rf(z
0
) =
s.r score /2/n, where n is the number of nodes con-
nected with s.
- Consider a node z that just received a reduction fac-
tor rf(z) and the set Out(z) of nodes connected to z
that have not received yet a reduction factor. For each
node z
0
∈Out(t), the reduction factor is rf(z
0
)=rf(z) / 2
/ |Out(z)|. The propagation continues until all nodes
have a reduction factor.
- For each node z, the i score is decreased by the re-
duction factor rf(z), i.e., z.i score = z.i score - rf(z)
(where the lower bound is z.i score = 0).
This way, the representativeness of the selected
node decreases the informativeness of neighbours; the
reduction becomes lower and lower as the distance
from the selected node increases.
Figure 3.a reports the initial term graph for the (as-
sessed) query in Example 1. Figure 3.b shows how
i score values change, showing the propagation from
the selected node (City) to the other nodes.
5.2 Deriving the Assessed Query
Having defined the concept of extrinsic representa-
tiveness for a term t (denoted as erep(t)), we can now
describe how query q is assessed into query q.
- The assessed query q is derived from query q by re-
placing each term t non appearing in the meta-data of
data sets with a substituting term t’.
- First of all, we consider the notion of string sim-
ilarity metric between two terms t and t’, denoted
as sim(t, t’)∈ [0,1]. We adopt the well known Jaro-
Winkler metric (see (Winkler, 1999), (Jaro, 1989)).
- For a pair of terms t and t’, we define the suitabil-
ity degree as suitability(t, t’)= (sim(t, t’)+erep(t’))/2,
i.e., a term t’ is more suitable to replace the missing
term t if t’ is as similar as possible to t and, at the
same time, it is as representative as possible of data
sets (extrinsic representativeness).
- A missing term t is replaced, in the assessed query q,
by the most suitable term t’ in the catalog, i.e., having
the greatest suitability(t, t’).
The rationale of this choice is simple: we want
to substitute missing terms with those that are more
likely than others to find relevant data sets, w.r.t. to
the query.
6 DERIVING NEIGHBOUR
QUERIES
The assessed query q could be not enough to find rel-
evant data sets for the user, because he/she writes the
query blindly, w.r.t. data set meta-data. As an exam-
ple, if the user wants to query a field named latitude,
he/she is usually interested also in data sets having
field lat (common short-cut for latitude). To make the
query process effective, we derive, from the assessed
query q, the neighbour queries nq, i.e., queries that
are close to q. Both q and nq
i
are collected into Q, the
set of queries to process.
Definition 11: Pool of Alternative Terms. Con-
sider a term t∈K(q). We denote as Alt(t) the set of
alternative terms for t.
A term t’∈Alt(t) if: 1) t’ is in the meta-data cat-
alog; 2) if its string similarity with t is grater than
or equal to a minimum similarity threshold th sim,
i.e., sim(t, t’)≥th sim; 3) given max alt the maximum
number of alternative terms, t’ is one of the max alt
top-most similar terms to t. 
In other words, Alt(t) contains the max alt terms
which are mostly similar to t.
Definition 12: Potential Neighbour Query. Con-
sider the assessed query q, with its keyword vector
Building a Query Engine for a Corpus of Open Data
131

a) b)
Figure 3: Initial state of the Term Graph (a) and after the first keyword selection (b). The triple in each node is (i score, irep,
erep).
K(q)=[t
1
, . .. , t
n
] and a vector of weights W(q), where
W(q)[i]= 1, for each term t
i
=K(q)[i]. A potential
neighbour query pnq and its keyword vector K(pnq)
are obtained as follows:
- |K(pnq)| = |K(q)|.
- One or more terms t
i
in K(q) are replaced with
one term chosen from the set of alternative terms
t
i
j
∈Alt(t
i
). If t
i
is not replaced, K(pnq)[i]=t
i
; other-
wise, K(pnq)[i]=t
i
j
. Correspondingly, the new poten-
tial neighbour query pnq is obtained by performing
the same replacements in the assessed query q.
- The vector of weights for the potential neighbour
query is as follows: if t
i
is not replaced, W(pnq)[i]=
1; otherwise, K(pnq)[i]=sim(t
i
, t
i
j
). 
Any query obtained by replacement of one or
more terms is a potential neighbour query. How-
ever, their number could be high and the query could
be possibly too far away from the assessed query;
therefore, we have to prune those potential neighbour
queries that may be too distant, based on the cosine
similarity metric applied to the weight vectors.
Definition 13: Neighbor Queries. Given the as-
sessed query q and a potential neighbour query p, p
is considered a neighbour query if CosineSim(W(q),
W(p))≥th neighbour, where th neighbour∈ [0,1] is a
minimum threshold. 
This way, the set Q of queries to process is pop-
ulated by the assessed query q and all neighbour
queries.
7 DATA SET RETRIEVAL
Each query nq ∈ Q has associated the vector K(nq) of
keywords selected in step 2 (for the assessed query)
or in step 3 (for neighbour queries). In step 4, the
Vector Space Model (Manning et al., 2008) is applied
to retrieve possibly relevant data sets.
The VSM approach relies on the concept of in-
verted index created on the bases of terms in the
global meta-data of each data set. Definitely, in our
context the inverted index IN implements a function
f (t) → {ods id
1
,ods id
2
,. .. }, that, given a term t, re-
turns the set of data set identifiers whose global meta-
data contain term t.
Consider a vector of keywords K(nq)=[k
1
, ...,
k
n
]. Then, given a data set identifier ods id, the
occurrences of keywords associated with ods id in
the inverted index IN are represented by the vector
m(ods id) = [p
1
, ..., p
n
], where p
j
= 1 if the data set
is associated with the keyword, p
j
= 0 otherwise.
The Keyword-based Relevance Measure krm(ods,
nq) of a data set ods for a query nq∈Q is the average
of p
j
values. The maximum value of krm(ods) is 1,
obtained when the data set is associated with all the
keywords, while values less than 1 but greater than 0
denote partial matching.
7.1 Schema Fitting
This step of our technique computes, for each data set
retrieved in step 4, the final relevance measure rm, by
composing the keyword-based relevance measure krm
of a data set with the fitting degree of its schema w.r.t.
the field names in the query. This step is necessary
to identify, among all retrieved data sets, those having
an instance that can be effectively processed by the
selection condition nq.sc.
We adopt an approach based on a weighted inclu-
sion measure, where the weight of each field name
depends on the role played in the query by the field
name.
Definition 14: Schema Fitting Degree. Consider a
data set ods and its schema ods.schema=<pn
1
, pn
2
,
.. .>. Then, consider a query nq∈Q and the set of
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
132

field names N = {n
1
,. .. ,n
k
} that appear in the selec-
tion condition nq.sc and/or in the list of field names
of interest nq.P. For each name n
i
∈ N, a weight
w
i
∈ [0,1] is associated (see Definition 15), depend-
ing on the role played by n
i
in the query. The Schema
Fitting Degree sfd(ods, nq) is defined as:
sfd(ods, nq)=
|N|
∑
i=1
w
i
× p(n
i
) /
|N|
∑
i=1
w
i
where function p(n
i
) = 1 if there exists a field name
pn
j
= n
i
in nq.schema, p(n
i
) = 0 otherwise. 
The following definition defines how names are
weighted.
Definition 15: Weight. The weight w
i
of a field name
n
i
∈ N is
w
i
= max(pw
i
,scw
i
)
where pw
i
is the weight due to the presence of n
i
in
the list nq.P of fields of interest in the query, while
scw
i
is the weight due to the presence of n
i
in the se-
lection condition nq.sc.
- pw
i
= 0.5 if field name n
i
∈ nq.P (may be relevant
to have this field in the resulting items); pw
i
= 0 oth-
erwise.
- scw
i
= 1 if the selection condition nq.sc contains one
single predicate with n
i
, or the predicates in the con-
dition are ANDed each other; scw
i
= 0.7 if predicates
are ORed each other. 
Finally, we define the relevance measure for a data
set.
Definition 16: Data Set Relevance Measure. Con-
sider a data set ods and a query nq∈Q. Given its
Keyword-based Relevance Measure krm(ods, nq) and
its schema Fitting Degree sfd(ods), the Relevance
Measure of data set ods w.r.t. query nq is
rm(ods, nq)=
(α − 1) × krm(ods, nq) + α × sfd(ods, nq)
where α = 0.6. The data set relevance measure
rm(ods) is
rm(ods)=max(rm(ods, nq
i
)),
for each nq
i
∈ Q. 
The rationale is that the schema fitting degree cor-
rects the keyword-based relevance measure. The rea-
son is simple: the keyword-based relevance measure
is obtained by querying the inverted index in a blind
mode, i.e., irrespective of the fact that a keyword is
a data set name, a field name or a word appeared in
meta-data. The keyword selection technique partially
corrects this blind approach, to discard those data sets
that could result accidentally relevant.
Finally, in order to perform item selection from down-
loaded data sets, how the field names in the query
match the actual schema of data sets becomes cru-
cial, because it is impossible to evaluate the selection
condition if fields in the condition are not present (all
or partially) in the schema of the data sets. This also
motivates the choice of α = 0.6.
Finally (see Section 3), the minimum threshold th
is used to discard less relevant data sets and keep only
those such that rm(ods)≥th.
8 EXPERIMENTAL EVALUATION
We implemented our technique in a prototype and
evaluated the effectiveness of the technique. Here, we
report a short evaluation, due to space limitations.
We performed experiments on a set of 2301 open
data sets published by the Africa Open Data portal
1
.
These open data sets are prepared in a non homoge-
neous way, as far as field names and conventions are
concerned. This fact makes the search more difficult
than with an homogeneous corpus (such as, e.g., New
York City open data).
We performed 6 different queries. They are re-
ported in Figure 4. Hereafter, we describe them.
In query q
1
, we look for statistical information
about malnutrition in Monbasa or Nairobi or Turkana
counties in Kenya. Apart from counties, we are inter-
ested in fields that describe underweight, stunting and
wasting. In query q
2
, we look for information about
civil unions since 2012, where the age of both spouses
is at least 35. The features of interest are year,
month, spouse1age and spouse2age. In query q
3
,
we look for information about availability of water,
a serious problem in Africa. In particular, we look
for distribution points that are functional (note the
condition, where we consider two possible values for
field functional-status (functional and yes). In
query q
4
we look for information about the number of
teachers in public schools, for each county. In query
q
5
, we look for data about per capita water consump-
tions at the date of December, 31 2013 (notice that the
date is written as "2013-12-31t00:00:00", to stress
the similarity search). Finally, query q
6
look for data
about orchid export in terms of weight (field kg) and
money.
To perform the evaluation, we downloaded all the
data sets and we explored and labelled every data set
and item as correct or wrong w.r.t. each query
2
.
We performed 9 tests, in order to evaluate the ef-
fect of varying some parameters. The maximum num-
ber of alternative terms maxc alt was set to 3 for all
tests, and the cosine similarity threshold for neigh-
bour queries th neighbour was set to 0.9995 for all
1
https://africaopendata.org/
2
We used Open Refine (http://openrefine.org) a powerful
tool to work with messy data
Building a Query Engine for a Corpus of Open Data
133

Table 1: Evaluation of Experimental Results (n.s.i. means no selected items).
Recall (%) for Data sets
Query Test 1 Test 2 test 3 test 4 test 5 test 6 test 7 test 8 test 9
q
1
13.33 13.33 13.33 13.33 13.33 13.33 80.00 80.00 80.00
q
2
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
3
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
4
66.67 66.67 66.67 66.67 66.67 66.67 66.67 66.67 66.67
q
5
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
6
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
Precision (%) for data sets
Query Test 1 Test 2 test 3 test 4 test 5 test 6 test 7 test 8 test 9
q
1
66.67 66.67 66.67 66.67 66.67 66.67 7.79 8.22 8.22
q
2
50.00 50.00 50.00 1.15 50.00 1.15 1.14 1.15 1.15
q
3
75.00 75.00 75.00 50.00 50.00 37.50 50.00 50.00 37.50
q
4
13.33 13.33 13.33 13.33 11.76 13.33 13.33 11.76 13.33
q
5
100.00 100.00 100.00 100.00 50.00 100.00 6.67 6.25 6.67
q
6
20.00 33.33 33.33 20.00 25.00 25.00 20.00 25.00 25.00
Recall (%) for Items
Query Test 1 Test 2 test 3 test 4 test 5 test 6 test 7 test 8 test 9
q
1
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
2
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
3
0.00 100.00 100.00 0.00 100.00 100.00 0.00 100.00 100.00
q
4
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
5
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
q
6
0.00 100.00 100.00 0.00 100.00 100.00 0.00 100.00 100.00
Precision (%) for Items
Query Test 1 Test 2 test 3 test 4 test 5 test 6 test 7 test 8 test 9
q
1
50.00 50.00 50.00 50.00 50.00 50.00 0.31 0.31 0.31
q
2
89.91 89.91 54.40 89.91 89.91 54.40 89.91 89.91 54.40
q
3
n.s.i. 51.43 51.43 n.s.i. 48.54 48.54 n.s.i. 48.54 48.54
q
4
10.39 14.39 9.72 10.39 9.72 9.72 10.39 9.72 9.72
q
5
100.00 100.00 100.00 100.00 100.00 100.00 24.73 32.85 49.46
q
6
n.s.i. 100.00 50.00 n.s.i. 100.00 50.00 n.s.i. 100.00 50.00
tests. Instead, we evaluated three different values for
th sim, i.e., the string similarity threshold for alter-
native terms, (0.9, 0.8 and 0.7) in combination with
three different values for th, i.e., the data set rele-
vance threshold (0.3, 0.25, 0.2). The combinations
are: for test1, th sim= 0.9 and th= 0.3; for test2,
th sim= 0.8 and th= 0.3; for test3, th sim= 0.7 and
th= 0.3; for test4, th sim= 0.9 and th= 0.25; for test5,
th sim= 0.8 and th= 0.25; for test6, th sim= 0.7 and
th= 0.25; for test7, th sim= 0.9 and th= 0.2; for test8,
th sim= 0.8 and th= 0.2; for test9, th sim= 0.7 and
th= 0.2.
In the first two sections of Table 1, we report re-
call and precision as far as the retrieval of data sets
is concerned, before filtering items. The reader can
observe that we often get very high values of recall,
even 100%, meaning that we are able to get all mean-
ingful data sets; the data set recall is sensitive to min-
imum relevance threshold variation only for query q
1
,
where with th= 0.2 the recall significantly improves.
Looking at the precision for data sets, results are more
various; for query q
2
, precision is very sensitive to
the variation of thresholds, due to the large amount of
false positive downloaded data sets; for query q
6
, the
variation of minimum thresholds generally improves
precision, meaning that we are able to better capture
the data sets of interest.
However, the final goal of our technique is to re-
trieve items that satisfy the selection condition. The
third and forth section in Table 1 reports the evalu-
ation of recall and precision for selected items. The
reader can notice that in this case things significantly
change: recall values are generally very high, mean-
ing that the technique is able to retrieve those (typ-
ically few) items of interest, in particular with low
settings for minimum thresholds. Looking at preci-
sion, we observe that sometimes (query q
4
) we get
low values (around 10%), meaning that a large num-
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
134

Table 2: Comparison of T Test 2 with the base the baseline obtained by means of Apache Solr.
Query
Baseline Test 2 - Data Sets Test 2 - Items
Recall(%) Precision(%) Recall (%) Precision(%) Recall(%) Precision(%)
q
1
100.00 22.09 13.33 66.67 100.00 50.00
q
2
100.00 1.96 100.00 50.00 100.00 89.91
q
3
100.00 25.00 100.00 75.00 100.00 51.43
q
4
100.00 33.33 66.67 13.33 100.00 14.39
q
5
100.00 100.00 100.00 100.00 100.00 100.00
q
6
100.00 10.00 100.00 33.33 100.00 100.00
q
1
:<dn=nutrition,
P= {county, underweight, stuting,
wasting},
sc= (county = "Monbasa"
OR county = "Nairobi"
OR county = "Turkana") >
q
2
:<dn=civilunions,
P= {year, month, spouse1age,
spouse2age},
sc= (year = 2012 AND (spouse1age >= 35
OR spouse2age >= 35) >
q
3
:<dn=wateravailability,
P= {district, location, position,
wateravailability} ,
sc= (functional-status = "functional"
OR functional-status = "yes" ) >
q
4
:<dn=teachers,
P= {county, schooltype,
noofteachers} ,
sc= (schooltype="public") >
q
5
:<dn=waterconsumption,
P= {city, date, description,
consumption per capita} ,
sc= (date = "2013-12-31t00:00:00") >
q
6
:<dn=national-export,
P= {commodity, kg, money} ,
sc= (commodity = "Orchids") >
Figure 4: Queries for the Experimental Evaluation.
ber of false positive items is selected; this is the effect
of the Jaro-Winkler string similarity measure adopted
to find similar fields.
The goodness of the approach is shown by the ex-
periments with query q
6
. The high precision for items
obtained in Test 2 (in spite of low precision for data
sets) is confirmed looking at downloaded items: many
of them have value "ORCHIDS/CYMBIDIUM" for field
commodity, instead of value "Orchid" written in the
query. Thus, our technique resulted particularly effec-
tive in the expansion phase with th sim.
Our technique obtained the best performance for Test
2, with th sim=0.8 and th=0.3. The reason is that
our approach is greedy during the VSM Data Set Re-
trieval; then, in the Schema Fitting step the search
space is narrowed and data set instances are down-
loaded. The filtering (and latter) phase is lazier: so
the average precision for items is above of 60%.
To complete the evaluation, we identified a suit-
able baseline. Among other tools, we decided to
adopt Apache Solr (Shahi, 2015), which is an open
source enterprise search engine, developed on top of
Apache Lucene.
We indexed the Africa Open Data corpus with
Apache Solr and executed the same 6 queries, eval-
uating both recall and precision for the retrieved data
sets (Apache Solr retrieves only data sets, while it is
not able to retrieve items).
Table 2 shows the results of our comparison: the
first couple of columns reports recall and precision for
data sets retrieved by Apache Solr, the second couple
of columns reports recall and precision for data sets
retrieved by our prototype in Test 2; the third couple
of columns reports recall and precision for items re-
trieved by our prototype in Test 2.
It is possible to see that recall is always 100% for
Apache Solr, while in two cases it is not 100% for
retrieved data sets in Test 2. In contrast, our proto-
type always overcomes Apache Solr as far as preci-
sion is concerned: this is mainly due to the generation
of neighbour queries and keyword selection.
Finally, when we consider retrieval of items (not
supported by Apache Solr), out prototype generally
behaves very well, with 100% of recall and generlaly
very high values for precision (except for query q
4
).
9 RELATED WORKS
The research area of this paper is certainly young,
and, at the best of our knowledge, we are in the pi-
oneering phase. However, many researchers are rea-
soning about Open Data management.
In (Khosro et al., 2014) Khosro et al., present the
state of art in Linked Open Data (LOD), with issues
and challenges. Moreover, the authors motivate this
topic by exploiting the projects analysed in the five
Building a Query Engine for a Corpus of Open Data
135

major computer science areas (Intelligence, Multime-
dia, Sensors, File System and Library), and present
the future trends and directions in LOD.
The Open Data world is related to the Linked
Data world. In fact, standardized proposals are typ-
ically used to describe published Linked Open Data.
The RDF (Data Set Description Framework) (Miller,
1998) is widely used for this purpose. Further-
more, the standard query language for RDF is called
SPARQL Standard Protocol and RDF Query Lan-
guage (Clark et al., 2008). W.r.t. our proposal, it is
very general and devoted to retrieve those data sets
with certain features. However, highly skilled people
in computer science are able to use it. In contrast, our
query technique is very easy for not skilled people and
closer to the concept of query in information retrieval.
Similar considerations are done in (Auer et al.,
2007), where the authors presents DBPedia: the RDF
approach is not suitable for non expert users that need
a flexible and simple query language.
In this paper we do not consider Linked Open
Dara: we query a corpus of Open Data Sets, in gen-
eral not related to each other, thus, not linked at all.
The idea of extending our approach to a pool of
federated Open Data Corpora is exciting. A pioneer
work on this topic is (Schwarte et al., 2011), but they
still rely on SPARQL as query language.
Finally, the heterogeneity of Open Data asks for
the capability of NoSQL databases. In (Kononenko
et al., 2014), the authors report their experience
with Elasticsearch (distributed full-text search en-
gine), highlighting strengths and weaknesses.
10 CONCLUSION
This paper presents a technique to retrieve items (rows
in CSV files or objects in JSON vectors) contained in
open data data sets from those published by an open
data portal. User blindly query the published corpus.
The technique both focuses the search on relevant
terms and expands the search by generating neigh-
bour queries, by means of a string matching degree.
The experimental evaluation shows that the technique
is promising and effective.
New and more extensive experiments will be per-
formed in the future, as well as the technique will
be refined and further improved, by adding seman-
tic information to drive the choice of similar terms.
In particular, as far as this point is concerned, we are
thinking to exploit dictionaries, such as WordNet, that
provide relationships between words. We think that
given a term in the query. we could discover its syn-
onyms and use them to rewrite the query (obtaining
new neighbour queries).
REFERENCES
Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., Cyganiak,
R., and Ives, Z. (2007). Dbpedia: A nucleus for a web
of open data. In The semantic web, pages 722–735.
Springer.
Clark, K. G., Feigenbaum, L., and Torres, E. (2008). Sparql
protocol for rdf. World Wide Web Consortium (W3C)
Recommendation, 86.
H
¨
ochtl, J. and Lampoltshammer, T. J. (2016). Adequate-
analytics and data enrichment to improve the quality
of open data. In Proceedings of the International Con-
ference for E-Democracy and Open Government Ce-
DEM16, pages 27–32.
Jaro, M. A. (1989). Advances in record-linkage methodol-
ogy as applied to matching the 1985 census of tampa,
florida. Journal of the American Statistical Associa-
tion, 84(406):414–420.
Khosro, S. C., Jabeen, F., Mashwani, S., and Alam, I.
(2014). Linked open data: Towards the realization of
semantic web - a review. Indian Journal of Science
and Technology, 7(6):745–764.
Kononenko, O., Baysal, O., Holmes, R., , and Godfrey,
M. (2014). Mining modern repositories with elastic-
search. In MSR. June 29-30 2014, Hyderabad, India.
Liu, J., Dong, X., and Halevy, A. Y. (2006). Answering
structured queries on unstructured data. In WebDB.
2006, Chicago, Illinois, USA, volume 6, pages 25–30.
Citeseer.
Manning, C. D., Raghavan, P., Sch
¨
utze, H., et al. (2008).
Introduction to information retrieval, volume 1. Cam-
bridge university press Cambridge.
Miller, E. (1998). An introduction to the resource descrip-
tion framework. Bulletin of the American Society for
Information Science and Technology, 25(1):15–19.
Schwarte, A., Haase, P., Hose, K., Schenkel, R., and
Schmidt, M. (2011). Fedx: a federation layer for
distributed query processing on linked open data. In
Extended Semantic Web Conference, pages 481–486.
Springer.
Shahi, D. (2015). Apache solr: An introduction. In Apache
Solr, pages 1–9. Springer.
Winkler, W. E. (1999). The state of record linkage and cur-
rent research problems. In Statistical Research Divi-
sion, US Census Bureau. Citeseer.
WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies
136