Intelligent Information Management, 2012, 4, 225-230
http://dx.doi.org/10.4236/iim.2012.45033 Published Online September 2012 (http://www.SciRP.org/journal/iim)
BINLI: An Ontology-Based Natural Language Interface
for Multidimensional Data Analysis
José Saias1, Paulo Quaresma1, Pedro Salgueiro2, Tiago Santos2
1Departamento de Informática-ECT, Universidade de Évora, Évora, Portugal
2Inductiva, Knowledge Technologies, Lda, Évora, Portugal
Received June 13, 2012; revised July 20, 2012; accepted July 30, 2012
ABSTRACT
Current technology facilitates access to the vast amount of information that is produced every day. Both individuals and
companies are active consumers of data from the Web and other sources, and these data guide decision making. Due to
the huge volume of data to be processed in a business context, managers rely on decision support systems to facilitate
data analysis. OLAP tools are Business Intelligence solu tion s for multidimensio n al analysis of d ata, allowing th e user to
control the perspective and the degree of detail in each dimension of the analysis. A conventional OLAP system is con-
figured to a set of analysis scenarios associated with multidimensional data cubes in the repository. To handle a more
spontaneous query, not supported in these provided scenarios, one must have specialized technical skills in data ana-
lytics. This makes it very difficult for averag e users to be autonomous in analyzing their data, as they will always need
the assistance of specialists. This article describes an ontology-based natural language interface whose goal is to sim-
plify and make more flexible and intuitive the interaction between u sers and OLAP solutions. Instead of programming
an MDX query, the u ser can freely write a question in his own human language. The system interprets this question by
combining the requested information elements, and generates an answer from the OLAP repository.
Keywords: NLP; BI; Ontology; Question Answering
1. Introduction
Current technology facilitates access to the vast amount
of information that is produced every day. A news article
about a company’s results can be read anywhere in the
world, from the very moment it is made available. The
amount of data potentially relevant for any topic, and the
lack of time on a process where the information gathered
on that topic is vital, lead to the adoption of automated
techniques for searching and filtering information.
Both individuals and companies are active consumers
of data from the Web and other sources, and these data
guide decision making. An investor may decide to buy
shares of a company based on the discovery of informa-
tion favorable to that company. In the activity of a man-
ager, to find that a group of customers shows a pattern
that requires interventio n, for instance, is a great achieve-
ment that depends on the access to relevant information
in the shortest time possible. Due to the huge volume of
data to be processed in a business context, managers rely
on decision support systems to facilitate data analysis.
Online Analytical Processing (OLAP) tools are Busi-
ness Intelligence (BI) solutions for multidimensional an
alysis of data, allowing the user to control the perspective
and the degree of detail in each dimension of th e analysis.
These systems are specialized tools for analysis and
visualization of large volumes of business data, usually
contained in a Data Warehouse (DW). A conventional
OLAP system is configured to a set of analysis scenarios
associated with multidimensional data cubes in the re-
pository, such as the total amount of sales per month. To
handle a more spon taneous query, not supported in these
provided scenarios, one must have specialized technical
skills in data analytics. Typically, the construction of such
a new analysis scenario would require the implementa-
tion of queries in an interrogation language, like MDX1.
This makes it very difficult for average users to be auto-
nomous in analyzing their data, as they will always need
the assistance of specialists. Instead of programming an
MDX query, the average user would feel more comfort-
able asking the system for the information he wishes to
see, using his own natural language. To view the amount
of sales per quarter, the user would simply write “What is
the total amount of sales per quarter”, as if he was ask-
ing it to a person.
This article describes an ontology-based Natural Lang-
uage Interface (NLI) called BINLI, whose goal is to sim-
plify and make more flexible and intuitive the in teraction
1Multidimensional Expressions (MDX) is the most widely supported
query language for reporting from multidimensional repositories.
C
opyright © 2012 SciRes. IIM

J. SAIAS ET AL.
226
between u s ers and OLA P tools.
In the following section we present some recent pub-
lications related to the work we are developing. The mo-
del we propose is described in Section 3. In Section 4 we
conclude by presenting some considerations about what
we have achieved, and enumerating some aspects to con-
sider for future work.
2. Related Work
A DW is “a subject-oriented, integrated, time-variant, and
nonvolatile collection of data in support of management’s
decision making process” [1]. DW is used as storage re-
sources for common OLAP systems. A multidimensio-
nal OLAP repository has cubes with many dimensions
relevant to a particular domain. The dimensions are at-
tributes associated with relevant perspectives for the ana-
lysis to be performed, such as a product category, or even
the date or location. A dimension can have multiple lev-
els organized into hierarchies, and each level may have
one or more members. The date dimension is one of these
cases, because it usually has a hierarchy of four levels:
Year, Semester, Month and Day [2]. Members are possi-
ble values for a level, for example 2010 or 2011 would
be memb ers of Year level.
A measurement is a value taken from the intersection
of all the dimensions, such as the value of a product P
sale on a given day D, at the store S. Quan titative data is
stored on a cube fact table. All the rows or measurements
in a fact table must be at the same grain (level of detail).
OLAP cubes are precalculated, in order to obtain a better
query performanc e [ 3].
The work in [4] is an OLAP system for analysis and
extraction of information on nursing records, starting
from the development of a repository of multidimensional
data. Authors have used several open source tools deve-
loped by Pentaho, a BI software company. That work is,
however, a typical implementation of an OLAP solution,
since it offers the user a set of fixed analysis scenarios,
within which allows common operations such as drill-
down, roll-up, slice and dice.
Conventional document retrieval systems are widely
used in order to find documents from a set of keywords
that describe the user’s information needs. But a collec-
tion of documents may not be the most interesting type
of response to situations that require rapid and specific
results. Question Answering (QA) systems allow users to
pose natural language questions [5], and instead of re-
turning full docu ments they provide concise answers.
Kuchmann-Beauger and Aufaure have recently pro-
posed a DW based QA system [6]. Their work is con-
cerned with the semantic analysis of a question, looking
for the data mod el objects that the quer y depends on, and
then producing a data visualization result. They identify
keywords or known terms in a natural language query,
and link those elements to data model objects, which
correspond to the OLAP repository dimensions, measures
or other schema objects. Using a set of business sales and
orders questions, the keyword direct match approach is
not always sufficient. When no answer has been found,
they rewrite the user question using a thesaurus-based
transformation, by applying synonym expansion. This
work also infers semantic closeness between terms based
on web search results, for unknown words in the question.
While many studies start with some unstructured natu-
ral language elements and look for a structured result
(such as a query), other works go in the opposite di rection.
Ioannidis work [7] aims at translating structured data
into natural language. The author concludes that text gen-
eration to produce a natural result is far from trivial, as
identifying t he right linguisti c const ruct s i s a complex task.
Conversation-based systems work with natural lan-
guage, such as QA systems, but have the characteristic of
generating a dialogue with the user, in order to resolve
ambiguity in the in terpretation of the question, or simply
to collect his feedback on a given result. An example of
conversation-based natural language interface to rela-
tional databases is described in [8]. Knowledge trees are
used to structure the do main knowledge and to dir ect the
conversational agent towards the goal of database query
generation as required by natural language input.
Frost and Fortier proposed a denotational semantics
model for natural language database queries [9], with ex-
plicit semantics for transitive verbs and negation. The
GINLIDB system is a generic interactive NLI for databa-
ses [10] meant to facilitate the interaction with datab ases
for common people who are not familiar with SQL syn-
tax. It has two major components: the linguistic handling
component and the SQL constructing component. That
system accepts English language requests, which are
interpreted and translated into SQL commands using a
semantic-grammar technique and a knowledge base with
the database schema.
In 2007, Li et al. proposed an XML database interac-
tive NLI [11] supporting aggregation, nesting, and value
joins. English sentences are translated into XQuery ex-
pressions, by mapping grammatical proximity of natural
language parsed tokens with the corresponding elements
in the XML data to be retrieved.
3. Proposed Approach
The main idea is that an ordinary person, with no back-
ground in interrogation languages or programming, can
interact with an OLAP system and perform queries in a
spontaneous way. These queries are written in natural
language, currently in Portuguese but the system is de-
signed to be multilingual. Queries may be related to any
subject or concept in the multidimensional repository,
about which the user wants some information, and they

J. SAIAS ET AL. 227
are not pre-computed. BINLI receives a question, freely
written by the user, interprets it, and generates an answer
from the multidimensional repository. We start by pre-
senting the modules of the system and then we describe
the methodology for processing the questions, presenting
some illustrating examples.
Figure 1 shows the main components of the BINLI
system and how they are interconnected. The user has a
Web interface to enter his questions and to receive the
system results. Question Analyzer is the most important
component of the BINLI system. It applies language spe-
cific tools according to the idiom used in the inserted
question. For now we consider Portuguese questions, but
other languages such as English, Spanish or French can
be supported as well, by integrating their respective dic-
tionaries and parsers.
In this part BINLI behaves as a QA system with a de-
fined domain: the DW content, the data in the OLAP
engine multidimensional repository. It must determine
what are the foci of interest for the user, which informa-
tion is sought on them, and still detect possible restric-
tions to consider, as occurs in typical QA question analy-
sis phase [5,12-14].
BINLI applies common procedures in NLP, i.e., pars-
ing and Named Entity Recognition (NER). The question
text is parsed for morpho-syntactic analysis, in order to
determine the main verb and the structure with the de-
pendency relationships between the query terms. Some
rules are also applied to handle special situations, de-
tected via superficial text pattern analysis. After NLP
techniques and the application of some rules, the system
seeks a mapping between core terms in user question and
OLAP schema objects. This is achieved by semantic si-
milarity computation and semantic reasoning [15,16]. In
the end of this phase, if a word or an expression admits
several meanings, all these semantic pa ths are considered,
resulting in a list of several possible interpretations for
the question. There is a weight associated with each mea-
ning of a query term, which is used as a sorting criterion
when considering v arious interpretations.
Figure 1. System Architecture.
The OLAP Query Generator is the module that will
generate an MDX query for each question interpretation,
according to the OLAP repository schema. If there are
multiple interpretations, they are sorted in descending
order of weight, and therefore the former is the more
plausible. The generated OLAP query code for the first
interpretation is then passed to the OLAP engine for
execution. For this purpo se, we chose the Pentaho Analy-
sis Services Community Edition, also known as Mon-
drian2, an OLAP server. Finally, the Answer Renderer
component shows the result in the form of tables and
charts using JPivot3 or similar tools. To find the elements
to display in response, the system searches for objects in
the repository schema that is relevant to the query. The
first approach is done by seeking expressions in the
query text that are a direct match with OLAP schema
objects. This correspondence is attempted only for query
terms that the former step of the analysis has marked as
possible links to the repository. If th e match is successful,
these are elements which will be sought on some proper-
ties or facts, in some repository table.
In addition to the direct match, the system also tries an
indirect concordance between candidate terms and schema
objects. Errors by exchange of two letters are common,
while writing a query. When a candidate term, such as
sotre, has no match in the scheme but the word store
would have a match, the system considers this possible
interpretation. We use the Levenshtein distance to deter-
mine if the similarity between the terms is acceptable.
The other approach for indirect relations is the test of
semantic compatibility between the terms. This is where
the support ontology plays an important role, allowing
the analysis of the semantic relations SynonymOf, Mero-
nymOf, HyperonymOf, InstanceOf and AkA. The ontogy
includes the terminology from repository multidimensio-
nal cubes. In addition to the base name associated with
each object in the DW schema, there is one or more al-
ternative designations for each supported language.
When applied to members in the OLAP schema, this
technique has a similar effect to the application of se-
mantic query expansion, in Information Retrieval, but
here the scope is broadened with the ontology content,
that may evolve.
When establishing the correspondence between query
terms and OLAP objects, the weight on a direct match
(100) is greater than the weight for indirect cases of
Levenshtein error correction (80) and the ontology based
semantic compatibility (90 for AkA relation; 85 for Syn-
onymOf relation; 60 for others). The weight of a query
interpretation is the average of the weights given to the
interpretation of its terms. This allows interpretation ran-
king.
2http://mondrian.pentaho
3http://jpivot.sourceforge.net

J. SAIAS ET AL.
228
Levenshtein distance allows automatic correction of that
expression into distrito.
Consider the following question, written in Portugue-
se:“Quais são as vendas de produtos por loja?” (or
“What are product sales per store?” in English). BINLI
Question Analyzer will perform a linguistic analysis to
the question text, discard the stop words, and it selects
for candidate terms, to connect to the repository scheme,
the words produtos (product), vendas (sales) and loja
(store).
The question refers to amounts for sales on mesas (ta-
ble products). There is no direct correspondence between
the term mesas and an object in the repository. However,
the ontology includes this business domain terminology,
and has a hint for the system: the term mesa is hypero-
nym of “mesa de centro” and “mesa de cozinha”. As msas
is the plural of mesa, the system performs an interpretation
of that question term involving some concepts in question
OLAP repository.
The module responsible for the generation of an
OLAP query is then activated and produces the MDX
code shown in Figure 2. In this simple example there is a
direct link between query terms and objects in the re-
pository schema. These OLAP schema object names ap-
pear underlined and highlighted with red color in Figure
2. They are an exact match of the relevant words we saw
in the question.
The text also includes the expression “last month”.
Such allusions are time constraints that th e system has to
solve. The user context includes the notion of location
and time. In this case, the system calculates which is the
month prior to the time of the question. Figure 4 shows
the result discriminating table sales values for each district.
The bars omitted in the second group means the absence
of values for those districts and for the desired month.
Figure 3 shows the result that the system returns to the
client through the Web interface. The Answer Renderer
fetches the execution result from the OLAP engine and
then generates the visual representation. The chart typ e is
set according to the number of dimensions to display.
SELECT
N
ON EMPTY{ DISTINCT( HIERARCHIZE( {{
[
Let’s take another example: “Qual é o valor das vend-
as de mesas no mês passado por distirto?” (in English:
“What is the value for table sales in last month per dis-
trict?”). In this case there is a spelling mistake in the
word distirto. The existence of a valid term at a very short
St ore ].[Name].Members}}))} ON COLUMNS,
N
ON EMPTY{ DISTINCT( HIERARCHIZE( CROSSJO IN({{
[Product]
Sales
.CHILDREN}},{{
[Measures].[TotalGross]}})))} ON ROWS
FROM []
Figure 2. MDX query for “What are product sales by store?”
Figure 3. Result for “What are product sales by store?”
Figure 4. Result for “sales in last month per the district”.

J. SAIAS ET AL. 229
4. Conclusions and Future Work
In this article we presented BINLI, an ontology-based
Natural Language Interface for multidimensional data
analysis. The language currently supported by this sys-
tem is Portuguese, but any of the most spoken western
languages can be supported in the near future. The model
has a generic architecture and is thought to be multilin-
gual, and may also operate in cross-lingual mode with
future improvements.
We developed a prototype system that is currently be-
ing evaluated by people in the management sector and
independent from the development process. Our experi-
ments show that the quality of the ontology is crucial to
achieve a correct interpretation. Th e current phase of tes-
ting aims to feed the ontology with semantic elements
that are gradually found, and that are usefull for the ques-
tion understandi ng process .
An OLAP cube can have many dimensions, each hav-
ing several hierarchies which describe the data under
diverse views. There may be cubes with homonymous
dimensions or hierarchies. This can be quite challenging
when matching the words from the NL query to the terms
found in the repository structure, producing many alter-
native interpretations. As in the work described in [6],
our system produces a list of possible interpretations for
a question. But instead of a thesaurus, the system de-
scribed in this article has an ontology as the support
knowledge base, which allows useful inference in proc-
esses of considerable semantic complexity.
The sorting criterion applied to the query interpreta-
tions needs to be revised. In addition to the interpretation
weight formula, we may consider other weight values in
individual term correspondence process depending on the
semantic relationship. In this initial phase of the project,
we focused on embracin g the OLAP schema and domain
terminology. We tried to minimize th e cases where there
was no response. To assess the impact of other interpre-
tation weight formulas, we need further testing using
queries chosen by BI experts and for which there are sev-
eral plausible interpretations.
Another feature we need to improve is the presentation
of alternative responses. When there are several possible
interpretations for a query, it should be easy for the user
to navigate through the results associated with each of
these interpretations. We try to automatically provid e the
most plausible result for the user’s need. Nevertheless, it
is interesting to see alternative interpretations.
In this line of improvement, the user can select the in-
terpretation that has to do with his interest when he asked
the question. The system will analyze patterns of prefer-
ence for that user and afterwards choose as first result a
more appropriate interpretation, for that question category.
In order to improve the interpretation capability, the
system can become interactive and ask the user for hints
about eventual query terms that might not be automati-
cally understood. The same can be done to reduce ambi-
guity. If a term T can be the name of a company but also
the name of a city, the system can prompt the user to cla-
rify the meaning of T before proceeding to calculate the
result, rather than choosing automatically one of those
interpretations.
Another important direction for future work is Text
OLAP [17]: the loading of data from unstructured sour-
ces into the OLAP repository. A significant part of busi-
ness data are unstructured or semi-structured documents,
to which we can apply information extraction and NLP
techniques. New data, obtained by the new approaches,
may lead to the discovery of relevant information for the
activity of the OLAP/BI user.
It is important to note that the system presented here is
not intended to fully replace the mouse-based interfaces,
with fixed and specific click or drag & drop operations.
Instead, the goal is to complement these conventional
interfaces with the introduction of a natural interface to
provide a simple and flexible solution for non-expert
users willing to use a BI system autonomously.
5. Acknowledgements
This research is partially supported by the QREN/PO
Alentejo program, under the project ALENT-07-0202-
FEDER-018599.
REFERENCES
[1] W. H. Inmon, “Building the Data Warehouse,” QED Te-
chnical Publishing Group, Wellesley, 1992.
[2] R. Bouman and J. van Dongen, “Pentaho Solutions: Busi-
ness Intelligence and Data Warehousing with Pentaho and
MySQL,” Wiley Publishing, Inc., Hoboken, 2009.
[3] R. Kimball and M. Ross, “The Data Warehouse Toolkit:
The Complete Guide to Dimensional Modeling,” 2nd
Edition, Wiley Computer Publishing, New York, 2002.
[4] J. Silva and J. Saias, “Olap em Âmbito Hospitalar: Trans-
formação de Dados de Enfermagem Para Análise Mul-
tidimensional,” In Actas das 2ªs Jornadas de Informática
da Universidade de Évora, JIUE’2011, Évora, 2011, pp.
77-85.
[5] C. Monz, “From Document Retrieval to Question An-
swering,” Ph.D. Thesis, University of Amsterdam, Am-
sterdam, 2003.
[6] N. Kuchmann-Beauger and M.-A. Aufaure, “A Natural
Language Interface for Data Warehouse Question An-
swering,” 16th International Conference on Applications
of Natural Language to Information Systems—NLDB
2011, Alicante, 28-30 June 2011, pp. 201-208.
[7] Y. E. Ioannidis, “From Databases to Natural Language:
The Unusual Direction,” Conference on Applications of
Natural Language to Information Systems, London, 24-27

J. SAIAS ET AL.
230
June 2008, pp. 12-16.
[8] M. Owda, Z. Bandar and K. A. Crockett, “Conversation-
Based Natural Language Interface to Relational Data-
bases,” Proceedings of the 2007 IEEE/WIC/ACM Inter-
national Conference on Web Intelligence and Interna-
tional Conference on Intelligent Agent Technology Work-
shops, Silicon Valley, 5-12 November 2007, pp. 363-367.
[9] R. A. Frost and R. J. Fortier, “An Efficient Denotational
Semantics for Natural Language Database Queries,” 12th
International Conference on Applications of Natural
Language to Information Systems, Paris, 27-29 June 2007,
pp. 12-24.
[10] F. A. El-Mouadib, Z. S. Zubi, A. A. Almagrous and I. El-
Feghi, “Interactive Natural Language Interface,” WSEAS
Transactions on Computers, Vol. 8, No. 4, 2009, pp. 661-
680.
[11] Y. Y. Li, H. H. Yang and H. V. Jagadish, “Nalix: A Ge-
neric Natural Language Search Environment for XML
Data,” ACM Transactions on Database Systems, Vol. 32,
No. 4, 2007.
[12] H. Q. Hu, “A Study on Question Answering System Us-
ing Integrated Retrieval Method,” Ph.D. Thesis, The
University of Tokushima, Tokushima, 2006.
[13] C. Amaral, A. Cassan, H. Figueira, A. Martins, A. Men-
des, P. Mendes, J. Pina and C. Pinto, “Priberam’s Ques-
tion Answering System in qa@clef 2008,” CLEF Work-
shop, 2008, pp. 337-344.
[14] J. Saias and P. Quaresma, “The Senso Question Answer-
ing System at qa@clef 2008,” Technical Report, CLEF
Workshop, 2008.
[15] R. Thollot, N. Kuchmann-Beauger and M.-A. Aufaure,
“Semantics and Usage Statistics for Multi-Dimensional
Query Expansion,” 17th International Conference of Da-
tabase Systems for Advanced Applications, Busan, 15-18
April 2012, pp. 250-260.
[16] J. Saias and P. Quaresma, “Semantic Networks and
Spreading Activation Process for QA Improvement on
Text Answers,” Proceedings of the 8th Brazilian Sympo-
sium in Information and Human Language Technology—
STIL2011, Cuiabá, 24-26 October 2011.
[17] B.-K. Park and I.-Y. Song, “Incorporating Text Olap in
Business Intelligence,” Business Intelligence Applications
and the Web: Models, Systems and Technologies, IGI
Global, 2012, pp. 77-101.