A controlled natural language (CNL) is a language engineered to look and feel
like natural language and to be more suitable, for a speci c purpose, than
natural language itself. CNLs have been designed to improve communication between
humans, like PoliceSpeak [
According to a survey [
The described problem arises only when the user does not know the schema. In this situation, even if users are pro cient with SQL or other formal query languages, they cannot query the database without knowledge of the schema. When the schema is largely unknown, users lack the vocabulary necessary to formulate queries in the rst place. We must point out that CNL interfaces provide some help in such a situation, through so-called predictive editors. Thanks to a formal, unambiguous grammar and to a closed vocabulary, such an editor displays all syntactically legal completions of the user input at every single keystroke. This feature lets the user have a glance at the vocabulary of the language, which give a hint about the content of the database.
Nevertheless, even when the user knows the vocabulary supported, the user needs to know how the terms they can be combined meaningfully, and what constraints hold on the data, in order to formulate queries that the system can answer. For example, predictive editors that rely only on syntactical information do not stop users from composing a meaningless query like \Which songs are performed by two songs? ", but also a query like \Which albums have two artists? ", that is reasonable even if it fails when the schema associates artists with songs but not with albums. If that is the case, even if the latter query is syntactically correct, no data will ever match it. In this article we introduce the Quelo system, which addresses this very problem, by supporting the user in the task of formulating a precise query { which best captures their information needs { even when the user ignores completely the vocabulary and the constraints of the underlying information system holding the data.
Our idea develops a technique presented in earlier systems. \Conceptual
authoring" [
We developed a novel system, named Quelo and formally described in [
Given the schema A, our tool can avoid suggesting C and D as an extension of query B, because according to the constraints in A, X will always be a man in B. But Quelo goes even further: it drives the user in composing informative queries|that is to say|it tries to limit the insertion of redundant elements. A case in point: when Quelo retrieves the possible substitutions for \married" in query E, it will lter out \person", even if \person" is semantically compatible with \woman". The reason for this exclusion is that a \woman" is a \person" already, therefore \person woman" would not be more informative than \woman".
In a nutshell, Quelo works on a tree-shaped, conjunctive, description logics formula. It uses a DL reasoner to calculate which extensions would make the formula unsatis able with respect to a OWL DL ontology, in order to lter them out. In addition, it lters out some of the extensions that, if added, would produce an equivalent formula.
The system has a few, simple requirements. First, our \semantic grammar" is an OWL DL ontology schema. Such a schema is standard technology used to represent a formal conceptualisation of a knowledge domain, that is, a conceptual schema. If an OWL DL ontology Modelling the target knowledge base already exists, our tool needs no information besides the schema to provide its services. Second, the system requires a reasoner for OWL DL. Many such reasoners are developed commercially and available o -the-shelf under various licensing agreements.
The input query is a directed tree, where each edge is labelled with exactly one binary predicate and each node is labelled with one or more unary predicates. Such a tree encodes a rst-order logic query in which the root node represents a free variable; each other node represents a distinct, existentially quanti ed variable; the label R of an edge directed from node x to node y represents a formula R(x; y); each label C of a node x represents a formula C(x). The overall formula encoded by the tree is the conjunction of all the formulae represented by edge and node labels and therefore it is a conjunctive query: no negations, no disjunctions, no universal quanti cations may appear. Moreover, it is tree-shaped: no variables may appear twice as the second argument of a binary predicate. For instance, the query \I'm looking for a woman who loves a married man" corresponds to the following formula: 9y Woman(x) ^ loves(x; y) ^ Married(y) ^ Man(y) :
The above restrictions on the shape of the query allow Quelo to deliver its services
fast enough to provide them on-line, that is, at the same time the user types the
query. Given such a query in input, the services o ered by Quelo [
Each set of unary predicates included in the result is arranged in a hierarchy according to the constraints of the target ontology.
In order to evaluate and showcase the system, we implemented a web-based
natural language interface (NLI) [
To compose a query in the NLI, the user starts from an initial, system-de ned sentence ("I am looking for something."). At its margins (and, later on in the process, within it) there are gaps where the user can insert snippets of text chosen among those proposed by the system. Some parts of the sentence, determined by the system, can be deleted, or replaced with other similar snippets. The insertion, deletion and substitution of system-determined snippets are the only operations allowed. Through a sequence of such operations, the user edits the text. At the same time, hidden to the user, the system maintains a directed tree corresponding to the meaning of the query. Each operation on the text a ects the tree by inserting, deleting or substituting a label, a node or a portion of the tree. Each part of the text is connected with a speci c element of the tree, and each snippet that is inserted or removed carries in or out an element of the tree. In this way, the system builds the meaning of the text by observing the operations performed. In particular, the system does not take into consideration the text as a whole resulting from these operations in order to produce the tree. In this way, the system avoids the disadvantages associated with natural language understanding.
The snippets inserted and removed cannot be simple canned text, otherwise the result of their composition would hardly resemble English. For this reason, our interface uses NLG. Each of our snippets consists in a so-called NLG \template", that is, a portion of a parse tree conforming to a simple English grammar with feature structures. Each such snippet has three layers: a surface textual representation, which is displayed to the user when it is suggested by the system, a portion of parse tree, which is used to produce the text as whole after the insertion of the snippet in the query, and a portion of a labeled directed tree, which represents the contribution of the snippet to the meaning of the query.
So far, the system we described follows closely the conceptual authoring archetype: the query is presented to the user in a generated natural language tightly connected with the underlying semantics. The ontology schema can thus be e ectively exploited by the users in order to formulate a natural language query that best captures their information need, without any ambiguity. Our improvement on the basic conceptual authoring system is that the snippets available for addition and for replacement are ltered by Quelo. As explained in the second section Quelo operates by means of appropriate automated reasoning techniques over the ontology schema which describes the domain of the data in the information system. For each snippet, Quelo checks whether the portion of labeled directed tree of the snippet is a valid extension of the current query, e ectively restricting the user's choice to only those snippets which are relevant and meaningful in a given context.
The most powerful and innovative feature of the conceptual authoring interface of our system lies in the fact that users need not to be aware of the underlying organisation of the data, nor they need knowledge of the vocabulary, and still they tool will guide them into composing a query that is sound with respect to the underlying information system. Moreover, such knowledge can be gradually acquired by using the tool itself, gaining con dence with the vocabulary and with the constraints of the ontology. Users may also decide to just explore the knowledge without actually querying the information system, with the aim of discovering general information about the modelled domain.
To work properly, our NLI requires a map from elements of the ontology schema
to such \templates". To ease the adoption of the system, we devised a rule-based
technique to extract linguistic information from the ontologies themselves in order
to generate the NLG templates automatically. In order to achieve this result, we
harvested and analysed more than 12000 unique schema relations, and we carefully
selected the grammatical features supported by the generated natural language in
order to keep the templates simple to build while being exible enough to satisfy
the most common expression needs of existing ontologies. While limited to English,
the technique is e ective to the point that the NLG interface can be used with most
ontologies virtually without the need of human intervention. [
We developed a natural language interface for Quelo, but Quelo was not speci cally
designed for it. In fact, we are currently developing other interfaces as well. We
believe it could also be used within CNL-based predictive editors to limit the set
of suggestions. Intuitively, when the query composed so far can be unambiguously
transformed into a well-formed formula, it could be fed to Quelo in order to lter
out suggestions associated with semantic extensions that are syntactically correct
but not semantically appropriate. A complete description of this work together with
an extensive set of references can be found in [